• KLA’s economics improve as semiconductor complexity rises, rare for industrials

• Services + data moat + packaging create a structurally stronger model

• Expected returns look flat unless the bull case is more likely than the market assumes

In the spring of 1997, a team of engineers at KLA-Tencor gathered in a conference room in Milpitas, California, to confront an uncomfortable reality. The semiconductor industry was migrating from the 350-nanometer to the 250-nanometer process node, and defect rates were spiraling out of control. Chipmakers who had previously treated optical inspection as a nice-to-have, a quality check bolted onto the end of the manufacturing line, were suddenly discovering that invisible particles, measured in fractions of a human hair’s width, could destroy entire wafer lots worth millions of dollars.

What happened next defined KLA’s strategic identity for three decades. Rather than diversifying into the larger, more glamorous equipment markets, deposition, etch, lithography, CEO Ken Levy doubled down on inspection. The bet was simple but counterintuitive: as transistors shrank, the laws of physics would increasingly fight back. The probability of killer defects would rise exponentially, making inspection not just important but existential for every chipmaker on earth. Every competitor chasing bigger markets would eventually have to come to KLA for the quality assurance layer.

By 2000, KLA owned more than half the wafer inspection market. It has never relinquished that position.

Twenty-nine years later, the bet has compounded beyond anything Levy could have imagined. KLA Corporation is a $222 billion company trading near all-time highs. Its process control tools are embedded in every advanced fab on earth, TSMC, Samsung, Intel, SK Hynix, Micron. Its 55% market share in inspection is more than four times its nearest competitor. And the same structural dynamic that made KLA essential in 1997, shrinking geometries creating more defect modes, is not decelerating. It is accelerating.

The 2-nanometer node, high-bandwidth memory stacks, and advanced chiplet packaging are the 1997 moment compounded across three dimensions simultaneously. The question this essay tries to answer is whether KLA is just a beneficiary of a strong capital expenditure cycle, or whether it is the rare industrial company whose economics structurally improve as its industry gets harder. And if the answer is yes, whether the stock price already knows it.

What KLA Actually Sells

To understand why KLA’s position is unusual, you have to define what the company actually does, and it is not what most people think.

A semiconductor foundry like TSMC does not buy a KLA machine to make chips. It buys ASML machines to print patterns on wafers. It buys Applied Materials and Lam Research machines to deposit and etch thin films. These are the tools that build the chip. The specific job a customer hires KLA to do is different: find the invisible flaw immediately, before it propagates through dozens of subsequent process steps, ensuring that the multi-billion-dollar fab actually prints money instead of expensive sand.

The value chain is clarifying. ASML makes the brush. Applied and Lam make the paint and the canvas. KLA is the magnifying glass that decides whether the painting is worth selling.

This matters because of the asymmetry embedded in the economics. A KLA tool costs millions of dollars. A failed 2-nanometer wafer lot with advanced packaging can destroy tens of millions. The return on investment for inspection is extreme and rising, not because KLA’s tools are getting more expensive relative to alternatives, but because the value of the wafer being inspected is getting more expensive relative to everything else. This is why customers do not want to buy KLA tools. They are forced to by physics. And it is why KLA maintains 63% gross margins and 44% operating margins regardless of the macroeconomic backdrop. The value is not in the atoms of the tool. It is in the asymmetry of the customer outcome: a small reduction in yield loss can be worth vastly more than the cost of the inspection system that prevented it.

Most semiconductor equipment companies sell capacity, the ability to produce more wafers. KLA sells the ability to make that capacity economically useful. This distinction is critical because it explains why KLA’s share of total fab spending rises at every node transition, a dynamic that has no parallel elsewhere in the equipment stack.

What Changed in the World

The 1997 dynamic never went away. But three things happened simultaneously in the last several years that made it far more powerful.

Logic got harder, in three dimensions

For decades, chipmakers improved performance by shrinking transistors on a flat plane. That era is over. The 2-nanometer gate-all-around transition is the most significant architectural shift in semiconductor manufacturing since Intel introduced FinFET transistors in 2011. Gate-all-around transistors replace the three-sided fin structure with a fully wrapped nanosheet, introducing four to five entirely new process modules and extending total manufacturing steps by roughly 20%. Each new step requires inspection and metrology that did not exist at prior nodes. The result: gate-all-around adds approximately 50% more inspection layers and 30% more metrology layers compared to FinFET.

Simultaneously, the number of extreme ultraviolet lithography layers per chip is escalating rapidly, approximately 14 layers at the 5-nanometer node, 20 at 3-nanometer, 25 to 26 at 2-nanometer, and 30 or more expected at 1.4-nanometer. Each additional EUV layer linearly increases demand for reticle inspection and patterned wafer inspection, both areas where KLA holds dominant share.

And backside power delivery, now in production at Intel’s 18A node and ramping at TSMC’s A16, creates an entirely new metrology category that did not previously exist: measuring through the full wafer thickness for front-to-back alignment, where pattern-dependent deformation creates non-linear overlay errors. This is net-new demand that appears in no historical model.

The cumulative effect: each advanced node transition adds 100 to 200 basis points of process control spending as a share of total wafer fab equipment investment. This is permanent and cumulative. Chipmakers cannot reduce inspection steps without accepting lower yields, and lower yields at 2-nanometer geometries are economically catastrophic.

Memory stopped being commodity memory

High-bandwidth memory is not simply more DRAM. It is a fundamentally different manufacturing challenge that transforms back-end memory assembly into a front-end yield problem.

The arithmetic is punishing. A 16-high HBM4E stack, where each individual die has a 95% yield and each bonding step has a 99% yield, produces an effective stack yield of approximately 41%. Improving individual die yield from 95% to 97% raises effective stack yield from 41% to 53%, a 12-percentage-point improvement worth hundreds of millions of dollars in output for a single memory fab. That yield sensitivity directly justifies massive investment in inspection at every stacking layer.

HBM adds roughly 100 basis points of process control intensity compared to standard DRAM. Each HBM wafer undergoes all standard front-end inspection plus multiple additional steps: through-silicon via formation inspection, post-thinning backside inspection, pre-bonding surface inspection, bonding alignment metrology, post-bonding defect inspection, and known-good-die verification. A single one-micron particle trapped between layers during hybrid bonding can cause delamination or void formation, destroying the entire stack.

The roadmap extends through HBM5 and HBM5E into the 2030s. And HBM4 introduces a revolutionary architectural change: the base die shifts from DRAM process to logic process technology, adding logic-grade process control requirements to what was previously a pure memory flow. This is not a one-time capacity buildout. It is a structural shift in how memory is manufactured.

Packaging became a front-end yield problem

For decades, semiconductor packaging was a commoditized, low-margin assembly process, the industrial equivalent of putting a chip in a box. Today, advanced packaging architectures like TSMC’s CoWoS and SoIC shrink interconnect pitches from 90 micrometers down to sub-15 micrometers. Hybrid bonding pushes defect sensitivity from greater than 5 micrometers to sub-1 micrometer, a five-to-ten-times increase that brings front-end-grade process control requirements to what was previously back-end assembly.

KLA’s advanced packaging revenue tells the story: approximately $300 million in 2023, $500 million in 2024, $950 million in 2025, and projected to exceed $1.1 billion in 2026. Market share in advanced packaging has quadrupled in four years. This is not KLA entering a new market with a new product. It is KLA adapting its existing front-end inspection capabilities to back-end packaging processes that now require front-end-like precision.

NVIDIA’s next-generation Rubin architecture will require approximately 2.7 times more interposer area per package than the current Hopper generation, with proportionally more inspection surface. Panel-level packaging, targeting mass production by 2028, will move from 300-millimeter round wafers to large rectangular panels, requiring entirely new inspection tools and creating a replacement cycle on top of growth that no one is currently modeling.

The connecting thread across all three shifts is this: artificial intelligence matters to KLA not through more wafers, but through more complexity per wafer. Larger die sizes, denser packaging, and vertical stacking all increase the probability that a single defect destroys an enormously expensive device. KLA’s tax base is not silicon volume. It is silicon complexity. And complexity, unlike volume, is a one-way ratchet.

The Hidden Quality Shift

The consensus view of KLA is straightforward: it is a best-in-class semiconductor equipment company benefiting from the AI cycle, trading at a deserved premium for its dominant market position. This is accurate. It is also incomplete.

The deeper story is that KLA’s business quality is improving beneath the surface in ways the market has not fully incorporated into its valuation framework. Three shifts are underway simultaneously.

Services are not a footnote

KLA’s services business generated $2.9 billion in calendar year 2025. It has grown for 52 consecutive quarters, more than 13 years without interruption. Over 80% of this revenue comes from multi-year contracts with renewal rates exceeding 95%. Management targets doubling services to approximately $6 billion by 2030, implying a 13 to 15% compound annual growth rate.

This is, functionally, a subscription business embedded inside a capital equipment company. The market values it at KLA’s blended equipment multiple. If it were a standalone company with these growth rates, this retention, and this margin profile, it would trade at a software-like revenue multiple.

The strategic significance became visible during fiscal year 2024, when KLA’s product revenue declined 11% in a mild WFE downturn. Services still grew 10%. This asymmetry, services grow in downturns because the installed base does not shrink, makes KLA’s overall business fundamentally more durable than the “semiconductor equipment” label suggests. By 2030, services will approach 25% of total revenue at higher margins than systems. The troughs of future downturns will be shallower and more profitable than any prior cycle.

Advanced packaging is not an adjacency

The shift from monolithic chip designs to chiplet architectures means packaging process control demand grows faster than packaging volume, because each successive generation of packaging technology adds inspection steps independent of how many packages are produced. NVIDIA’s migration from CoWoS-S to CoWoS-L, Broadcom’s adoption of CoWoS-L for advanced TPU models, and the emergence of outsourced assembly providers like ASE and Amkor standing up CoWoS lines for the first time, each of these creates greenfield KLA demand that is largely absent from consensus models.

By 2030, advanced packaging could plausibly represent 15 to 20% of KLA’s total revenue, up from approximately 7 to 8% today. That is not a product extension. It is a second core business.

The learning system that strengthens at scale

KLA has approximately 57,000 tools installed across every major fab on earth, with a median operating life of 24 years. These tools have accumulated three decades of proprietary defect libraries from TSMC, Samsung, Intel, SK Hynix, Micron, and dozens of smaller manufacturers. Each new generation of tools improves detection accuracy through machine learning algorithms trained on this data, and deploying those algorithms across the entire installed base carries zero marginal cost.

This creates something close to a compounding industrial learning system. More complex chips produce more novel defect types. More defect types generate more proprietary data. More data improves algorithmic accuracy. Better accuracy increases switching costs, qualifying a new inspection vendor takes 12 to 18 months and directly risks production yield, making no fab manager willing to gamble billions in yield loss to save a few million on tool capital expenditure. Higher switching costs cement market share. Larger market share expands the installed base. And a larger installed base generates more service revenue, more customer touchpoints, and more data, restarting the loop.

A competitor can theoretically build a comparable optical inspection tool. They cannot replicate 30 years of production defect data from every advanced fab on earth. This is not a barrier to entry. It is a barrier to existence.

The market sees a better semiconductor equipment company. The deeper reality is a better business model.

The Price of the Physics Tax

Everything described above is real. The structural dynamics are genuine. The compounding mechanism is durable. And nearly all of it is already in the stock price.

At 49 times trailing earnings and approximately 36 times forward earnings, KLA trades at 75% above its 10-year average price-to-earnings ratio. The stock has doubled in the past year. Every independent probability-weighted three-year price target I have encountered, from institutional research, from independent analysts, from my own scenario work, converges on a range of $1,470 to $1,693. The stock trades at $1,727.

The positioning data confirms the crowding. KLA is the number-one most overweight semiconductor stock among active fund managers, according to Bank of America’s February survey. “Long global semiconductors” is tied as the single most crowded trade on Wall Street, cited by 35% of survey respondents. Short interest has collapsed 27% from January to February, falling to just 2.2% of float. Bears have capitulated. There are no contrarians left to be proven wrong.

Insider activity is uniformly in one direction. CEO Rick Wallace sold approximately $22.9 million in shares over the past six months. CFO Bren Higgins sold approximately $4.8 million. EVP Brian Lorig sold approximately $7.4 million. No insider purchases have been detected. While insider selling at all-time highs is common and often programmatic, the complete absence of buying is notable for a management team that publicly presented a target of $84 in earnings per share by 2030 at their March investor day.

The January earnings report demonstrated exactly how this dynamic plays out in practice. KLA beat revenue estimates by 1.4%, beat earnings per share estimates, and guided the next quarter above consensus. The stock fell 15% over the following sessions. At these multiples, execution gets punished. Only acceleration gets rewarded.

The business case and the stock case have diverged. The business is getting structurally stronger. The stock has already capitalized that strength.

Three Scenarios

The range of outcomes over the next three years depends almost entirely on two variables: the trajectory of wafer fab equipment spending and the terminal multiple the market assigns to KLA’s earnings stream. The business fundamentals, share, intensity, services, are relatively stable across scenarios. What changes is the macro and the market’s willingness to pay.

Bull Case (25% probability)

The AI capital expenditure supercycle sustains through 2028. Wafer fab equipment spending reaches $175 billion or more. Process control intensity accelerates toward 9% ahead of schedule as 2-nanometer yields prove more difficult than expected, forcing near-100% inline inspection at critical layers. Advanced packaging revenue exceeds $2 billion. Services track toward the $6 billion target.

Under these assumptions, KLA reaches approximately $22 billion in revenue by calendar year 2028, with operating margins expanding to 46 to 47% and non-GAAP earnings per share of roughly $68. If the market sustains a 35 times forward multiple, consistent with a monopoly whose cyclicality is demonstrably declining, the stock reaches approximately $2,380. That is a 38% return from current levels.

This requires management’s 2030 targets to be on track and the market to sustain a premium multiple through a full cycle. This combination has no precedent in semicap history.

Revenue CAGR: 17 to 18%. Operating margin: 46 to 47%. Terminal multiple: 35 times.

Base Case (55% probability)

AI spending continues but moderates. Wafer fab equipment grows at high-single-digit rates to $150 to $160 billion. Process control intensity reaches approximately 8.5% by 2028. China stabilizes at 20 to 25% of revenue. Advanced packaging reaches $1.5 to $1.8 billion. Services reach approximately $4.5 billion, tracking toward but not yet at the 2030 target.

Revenue reaches approximately $18.5 billion by 2028. Operating margins remain stable at 44 to 45%. Earnings per share reach approximately $55. At a 30 times forward multiple, a premium to history but below the current level, the stock trades at approximately $1,650. That is a 4% decline from today, roughly offset by cumulative dividends.

The business executes well. The stock goes approximately nowhere because the current multiple already embeds this outcome.

Revenue CAGR: 12 to 13%. Operating margin: 44 to 45%. Terminal multiple: 30 times.

Bear Case (20% probability)

The AI capital expenditure cycle experiences a digestion period in 2027 to 2028, consistent with Gartner’s repeated flagging of this risk window. Wafer fab equipment stagnates at $130 to $135 billion. Further export controls, potentially including the proposed MATCH Act’s retroactive prohibition on servicing existing tools in Chinese fabs, sever a portion of KLA’s highest-margin recurring revenue. Process control intensity stalls at approximately 8% as foundries delay advanced node migrations to conserve capital.

Revenue stagnates near $15.5 billion. Operating margins compress to 41% as fixed R&D costs deleverage. Earnings per share reach approximately $42. The multiple compresses to 25 times forward, still above the 20-year average of 24 times, reflecting the structural improvements in the business. The stock trades at approximately $1,050. That is a 39% decline.

Revenue CAGR: 5 to 6%. Operating margin: 41%. Terminal multiple: 25 times.

The Thesis Scoreboard

Seven metrics will determine which scenario is unfolding, listed in order of importance.

Process control intensity as a percentage of wafer fab equipment spending is the single most important number. If it approaches 8.5% by the end of calendar year 2026, the structural case strengthens materially. If it stalls near 7.5%, the thesis weakens.

Advanced packaging revenue trajectory must sustain growth above 15% annually. Any deceleration below 10% signals the second engine is faltering.

Services revenue growth rate must remain above 12% to track toward the $6 billion 2030 target. If growth drops below 8% for two consecutive quarters, particularly due to lost Chinese installed-base access, the compounding thesis weakens.

Gross margin trajectory above 62% signals pricing power is intact and tariff and component cost headwinds are being passed through. Below 61% signals absorption and margin ceiling risk.

ASML’s quarterly order intake is the most reliable leading indicator of wafer fab equipment direction six to twelve months forward.

Hyperscaler capital expenditure guidance from Microsoft, Google, Meta, and Amazon cascades to TSMC and then to KLA within two to three quarters. Any synchronized downward revision is an early warning.

Insider buying would be the most powerful signal against the current all-sell pattern. Any C-suite purchase, of any size, from a management team guiding to $84 earnings per share, would carry outsized informational value.

The Inspector’s Dilemma

In 1997, Ken Levy bet that physics would make inspection indispensable. He was right, spectacularly, permanently, comprehensively right. KLA’s market position today is stronger than at any point in its history. Its addressable market is growing faster than the industry it serves. Its financial model produces margins and returns on capital that rival the best businesses in technology. And the same dynamic that made KLA essential 29 years ago, shrinking geometries creating more defect modes, is accelerating into the most complex era of semiconductor manufacturing ever attempted.

But here is the irony that Levy’s successors must contend with. The very qualities that make KLA an extraordinary business, dominant share, rising intensity, compounding services, a deepening data moat, have become so widely recognized that the stock market has priced them into a 49 times earnings multiple. At the 10-year average multiple, KLA would trade at roughly $970. The $750 gap between $970 and $1,727 is the premium the market pays for the expectation that the next decade will look like the last.

KLA’s inspectors find the defects that would otherwise destroy billions of dollars in silicon. Investors should apply the same rigor to the stock. The business is getting better as the industry gets harder. The company may be one of the rare industrial franchises whose economics improve with every increase in manufacturing complexity, a toll on avoidable failure that chipmakers will pay in perpetuity because the cost of not paying is worse.

The question is whether that truth is still an insight, or already common knowledge embedded in a very demanding stock price. In semiconductor manufacturing, defects that escape inspection always matter eventually. In equity markets, valuation defects have the same property. They are invisible for long stretches. And then, quite suddenly, they are not.

Subscribe now

• ABF (Ajinomoto Build-up Film), is not “just packaging material”; it is the insulating layer that makes high-end organic substrates possible, and AI is driving far more ABF per chip through larger packages, more layers, and harsher yield economics.

• The industry is structurally split: Ajinomoto sits upstream in a capital-light near-monopoly, while substrate makers operate a capital-heavy, qualification-bound oligopoly where only a few players can really serve AI-grade demand.

• The market still mostly sees Ajinomoto as a food company, but the economics increasingly look like semiconductor infrastructure, and that mismatch is the heart of the thesis.

In 1996, a young researcher at Ajinomoto, the Japanese company that invented MSG, dominates global soy sauce markets, and sells seasoning packets in 130 countries, was handed an unusual assignment. Intel needed a better insulating film for its next-generation CPU substrates. The materials it was using were inconsistent, difficult to process uniformly, and insufficient for the routing density the Pentium III would require. What Intel needed was a thermosetting film that could be laminated uniformly onto circuit panels, survive laser drilling at microscopic scales, support copper plating in semi-additive processes, and do all of this reliably across millions of packages.

The researcher was Shigeo Nakamura. He had joined Ajinomoto’s Central Research Laboratory four years earlier, drawn to fermentation chemistry and materials science, working in what the company then described as an “emerging business segment.” Electronic materials was not Ajinomoto’s core business. It was not obvious it ever would be.

“In 1996, we started development of Ajinomoto Build-up Film,” Nakamura wrote in his CEO message for 2025. “Along the way, we faced many crises of survival, as it was not a core business, but I devoted myself to research and development due to the tenacity of my superiors and a pride-based sense of urgency making me unwilling to let the research end so unceremoniously.”

The film shipped commercially in 1999 for the Pentium III. When it was adopted, Nakamura assumed it might stay in production for ten years. It has now been over twenty, and ABF is still going strong.

Nobody outside the packaging industry noticed at the time. No analyst note captured the significance. A food company had quietly become the upstream supplier to the entire high-performance computing industry, and it had done so not through acquisition or grand strategy, but through one researcher who refused to let an obscure materials project die.

On February 3, 2025, Ajinomoto named Nakamura as President and CEO of the entire Ajinomoto Group, noting that he had “led the launch and growth of the Electronic Material business centering on Ajinomoto Build-up Film.” He became, as his official biography records, the Ajinomoto Group’s first CEO with a technical background.

That promotion is the tell. When the inventor of your most strategically important product becomes group CEO, not of the electronics subsidiary, not of the materials division, but of the whole company, the organization is communicating something about where its center of gravity is moving. But to understand why it matters specifically now, in 2026, you have to start not with Ajinomoto but with what AI has done to the substrate underneath every chip it powers, and why that shift is forcing a rerating of an entire supply chain that most investors still price as a commodity cyclical.

The Package Became the Problem

The standard frame for the semiconductor supply chain puts the interesting action on the die. Transistors scale. TSMC builds them smaller. Performance improves. The chip ships. Packaging is plumbing, necessary, unglamorous, and strategically peripheral.

That frame was largely correct through the PC era. It is wrong now and understanding why is the key to understanding everything that follows.

For fifty years, Moore’s Law delivered performance by shrinking transistors. That scaling has slowed, not stopped, but slowed enough that the industry has had to find performance elsewhere. The answer, increasingly, is the package itself. AI systems gain performance today not primarily by making the die smaller, but by integrating more compute dies together on one package, stacking high-bandwidth memory directly beside them, and connecting everything through increasingly sophisticated interconnect structures. The package, in other words, has moved from plumbing to compute infrastructure. It is now part of the scaling curve.

This shift has a specific material consequence. Every routing layer inside a high-performance substrate, every insulating layer that separates copper traces, allows laser-drilled microvias to connect one layer to the next, and maintains electrical integrity at multi-gigahertz frequencies, is built from ABF. The film is laminated, cured, drilled, and plated, then the cycle repeats for every additional layer the package requires. A substrate for a PC CPU might go through that cycle six times. A substrate for a current-generation AI accelerator goes through it eighteen times or more.

That difference is not incremental. It represents a completely different business.

What AI Changes About Package Economics

Unimicron chairman Tzyy-Jang Tseng has noted that substrate areas for CoWoS technology have grown more than ten times. That statement is worth sitting with. The substrate under an AI accelerator is not a bigger version of a PC substrate. It is a categorically different product, more layers, larger body, tighter tolerances, harsher yield economics, that happens to use the same insulating film and similar equipment. A company that makes six-layer PC substrates is not in the same business as a company making twenty-layer AI substrates. They share an industry name. They do not share an economics.

This is where the market has been making its central error. The ABF substrate industry is not a single market that AI has made better for everyone. It is two markets, advanced AI substrates with pricing power and structural scarcity, and commodity PC substrates with overcapacity and margin pressure, that share branding and some equipment but have diverged completely in economics. Understanding which tier each company actually operates in is the only thing that matters for understanding who benefits from the structural shift and who doesn’t.

The 2024 semiconductor downturn was the clearest diagnostic event in recent sector history, and most analysis drew the wrong conclusion from it. When Ibiden’s revenue fell eighteen percent and Nan Ya PCB’s ABF segment dropped more than twenty percent, the conventional read was that AI substrate demand was softer than expected. The better read is almost the opposite: the companies with genuine AI substrate exposure held; the companies with AI narratives, still largely PC businesses with AI stories bolted on, were exposed by PC overcapacity hitting their core business. Kinsus, which had concentrated its advanced lines on AI clients, achieved a thirty-one times profit recovery in 2025. Samsung Electro-Mechanics posted record annual revenue of KRW 11.3 trillion. AI-powered applications reshaped the electronics manufacturing landscape in 2024, with Unimicron and Kinsus posting double-digit revenue growth, while Nan Ya PCB struggled. The bifurcation was not a temporary anomaly. It is the permanent structure of the industry.

Monopoly Chemistry, Oligopoly Manufacturing

Before you can understand the substrate makers, you have to separate two things the industry routinely conflates: ABF film and ABF substrates. They have different competitive structures, different capital requirements, and completely different economic outcomes. Conflating them is how most coverage of this sector goes wrong from the first paragraph.

ABF film is Ajinomoto’s product. It is a thermosetting composite, epoxy resin, hardener, silica filler, delivered in rolls to substrate manufacturers worldwide. Ajinomoto’s market share in high-performance CPU and GPU substrate grades: above ninety-five percent. Sekisui Chemical, its nearest competitor, holds a few percent in lower-end applications and has not credibly challenged Ajinomoto’s advanced grade position in over two decades of trying.

“We expect sales of electronics materials, mainly ABF, to grow at an annual rate of more than 10% through 2030,” Nakamura told Nikkei Asia. “We will invest the same amount or more by 2030 as demand increases.” The investment required to achieve this: a ¥25 billion, approximately $166 million, plan to grow ABF production capacity by fifty percent by 2030.

ABF substrates are a completely different business. Seven manufacturers convert Ajinomoto’s film into the finished packages that sit under every AI chip. Their capital requirements: $800 million to $1.2 billion per new fab. Their qualification timeline: thirty to thirty-six months before a single revenue-generating substrate ships. Their gross margins: decent when utilization is high and product mix skews AI, brutal when it doesn’t.

The asymmetry in this table is the industry in miniature. $166 million buys a fifty percent expansion in the film that every downstream business depends on. $800 million to $1.2 billion buys a single substrate fab that still needs three years to qualify before shipping revenue. The film monopolist expands at one-fifth the cost of a single downstream factory, and every one of those factories depends on its film allocation to run.

This is the structure: capital-light monopoly chemistry feeding capital-heavy oligopoly manufacturing. Ajinomoto is not a supplier in the conventional sense. It controls the standard layer on which the rest of the industry runs, and it does so with pricing power, minimal competitive pressure, and capital requirements that barely register against the businesses downstream. As of today, no U.S. entity manufactures ABF or has secured licensing rights, making the United States entirely import-dependent on a material essential for advanced semiconductors. That is not a supply chain vulnerability that announced itself loudly. It accumulated quietly, one qualified substrate generation at a time, while everyone was watching the chip companies.

Qualification, Not Capacity, Is the Real Moat

The lazy frame for the ABF substrate supply situation is: everyone is expanding, so the shortage should ease. More than $15.5 billion in substrate capacity was announced industry-wide in 2021 and 2022. Ibiden committed ¥500 billion, roughly $3.3 billion, over three years. Unimicron raised its 2026 annual capex to a record NT$34 billion. AT&S has invested more than €2.5 billion across three continents.

The better frame: announced capacity and real supply are completely different things, separated by thirty-six months and a yield learning curve that no amount of money can compress.

Bloomberg reported that Ibiden CEO Koji Kawashima said sales of the company’s AI-use substrates are “robust with customers buying up all that Ibiden has,” with demand likely to last at least through the following year. “Our customers have concerns,” Kawashima told reporters. “We’re already being asked about our next investment and the next capacity expansion.”

That language, customers asking about the investment after next, reveals the fundamental structure of competition in this industry. The qualification cycle is so long that customers who want substrate supply in 2028 are negotiating factory commitments in 2025. The production timeline and the demand timeline are separated by years, and the sophisticated customer knows it. This is precisely why Ibiden carries ¥92 billion in customer advance payments, not as a balance sheet metric but as a demand conviction signal. When NVIDIA pre-pays for substrate capacity years before the substrate ships, it has calculated that supply disruption is worse than paying above-market prices today.

Toyo Securities analyst Hideki Yasuda stated the competitive reality plainly after assessing the field: “Nvidia’s AI chips need sophisticated substrates, and Ibiden is the only one that can mass produce them at a good production yield.”

Not the largest. Not the most capital-invested. The only one that can mass produce at good yield. Yield is the moat, and yield is accumulated over decades of production runs, refined through failure analysis, embedded in proprietary chemical recipes, plating parameters, and laser settings that are not disclosed and cannot be transferred. A competitor with identical equipment and unlimited capital still needs years to match the yield curve of an incumbent with thirty years of process history on the specific substrate type in question.

The customer co-investment structure makes this explicit. Intel, AMD, and NVIDIA have collectively subsidized approximately fifty percent of expansion capex at their top substrate suppliers. This is not generosity. It is a revealed preference about irreplaceability. When a company with NVIDIA’s pricing power decides the correct move is to pre-fund a substrate manufacturer’s factory, they have revealed that a financially weak or supply-constrained substrate maker is worse for them than overpaying for guaranteed allocation today.

This also explains why more announced capex does not automatically tighten the supply picture for everyone. AT&S has invested over €2.5 billion in substrate capacity. That investment is real. But AT&S’s ability to capture AI substrate economics is constrained by AMD customer concentration above seventy-seven percent of Americas revenue, €1.3 billion in net debt, a sub-twenty percent equity ratio, and three top executives replaced within twelve months during a critical ramp period. Capital deployed is not the same as qualification achieved. The distance between those two things is where most analysis of this sector goes wrong.

A Market That Does Not Lift All Boats

The seven substrate makers are not in the same business, and AI has made them less equal over time, not more. The bifurcation is not a temporary divergence that supply normalization will correct. It is the permanent structure of the industry, driven by yield economics and qualification depth that compound with every generation.

Where Each Player Actually Stands

Ibiden and Unimicron form the genuine top tier, and the gap between them and the rest is wider than market share numbers suggest. Industry observers describe the two as operating in a “quiet duopoly” in one of the most strategically critical segments of the semiconductor supply chain, as the only suppliers capable of manufacturing the most advanced ABF substrates that AI accelerators require. Every other substrate maker is competing for secondary position, commodity volume, or a specific niche that may or may not grow.

Kinsus deserves more attention than its seven percent market share implies. It is the purest AI-leverage play in the group, not because it is the most technically capable, but because it has concentrated its most advanced production lines almost entirely on AI clients. The thirty-one times profit recovery in 2025 is the clearest financial expression of what happens when a substrate maker successfully makes the move from PC exposure to AI exposure. The risk is the mirror image of the opportunity: that level of concentration creates fragility if any major customer shifts allocation.

Samsung Electro-Mechanics is the highest-variance name. It is the newest entrant in high-performance server substrates, mass production started only in October 2022, yet it has secured supply relationships with AMD, NVIDIA’s NVSwitch product line, and multiple hyperscalers within eighteen months. Samsung Electro-Mechanics has accelerated its glass substrate timeline, moving equipment procurement and installation to September and launching a pilot production line a full quarter ahead of schedule, with expectations of producing glass substrates for high-end applications in 2026. If glass cores reach meaningful volume after 2027 and Samsung is the only substrate maker with production-ready glass capability at scale, the thirty years of organic substrate yield expertise that Ibiden has accumulated becomes less relevant than its absence of glass substrate expertise. That is a binary outcome, either Samsung leapfrogs the field or it has split its resources between two substrate technologies while incumbents concentrated entirely on one.

AT&S is where the geopolitical diversification narrative and the financial reality are most in tension. The company is genuinely the only non-Asian major substrate maker. That is strategically valuable for Western chip companies with supply chain concerns. But strategic value and investment value are not the same thing when the balance sheet carrying that value has €1.3 billion in net debt, revenue targets that have been revised materially downward, and customer concentration in a single client navigating a difficult market. The bull case requires AMD to win in AI accelerators at scale. That remains an open question.

What Could Actually Break This Structure

Three things could disrupt the current dynamics. Only one is imminent.

The most-discussed threat is glass core substrates. Glass offers real advantages: near-zero coefficient of thermal expansion matching silicon, elimination of warpage in packages exceeding 100mm, and theoretical interconnect density improvements of up to ten times. Intel launched the first commercial glass core product in early 2026. Samsung Electro-Mechanics is aiming for mass production as early as 2026, potentially challenging Intel’s first-mover advantage, while TSMC is integrating glass into its CoWoS ecosystem through a new variant called CoPoS.

The sobering counterpoint: glass substrates currently cost two to three times organic equivalents, manufacturing yields remain at seventy-five to eighty-five percent versus organic substrates’ ninety to ninety-five percent, and brittleness requires specialized handling automation that the industry is still developing. Mass production readiness is unproven. The most likely near-term outcome is a hybrid architecture, glass core with ABF build-up layers, which benefits Ajinomoto: glass cores don’t compete with ABF build-up layers, they add a new substrate component while ABF remains the build-up dielectric. The company most vulnerable to full glass displacement is Ibiden, whose moat is entirely built on organic substrate yield expertise. The company best positioned to benefit is Samsung Electro-Mechanics, for precisely the opposite reason.

The geopolitical risk is less discussed but more immediately material. Eighty percent of Japanese container traffic transits the Taiwan Strait, while China is rapidly investing in alternatives to reduce its dependence on Ajinomoto’s product. Over forty-five percent of substrate production sits in Taiwan. The CHIPS Act investments, Ibiden’s Arizona fab, AT&S’s Texas facility, are real, but meaningful US capacity is 2028 at the earliest. Until then, the supply chain’s geographic concentration is essentially unchanged.

The demand risk that matters most in the near term is a hyperscaler capex pause. AI infrastructure spending is discretionary in a way that smartphone components are not. The advance payment structure is the best available leading indicator of demand conviction, and it is currently rising. But it has risen for only two years, through a period of extraordinary AI investment enthusiasm. The AI infrastructure buildout has not yet been tested through a full economic cycle.

The Category Error Nobody Has Named Yet

Here are the numbers the market has not yet fully processed.

Ajinomoto’s total annual revenue is approximately ten billion dollars. ABF film revenue is approximately $570 million, five to six percent of total group revenue. The functional materials segment containing ABF generates an estimated twenty percent of total group profit. Nakamura has set a target for electronics materials sales to grow at more than ten percent annually through 2030, with $166 million of capex buying a fifty percent capacity increase, at a time when a single substrate fab costs five to eight times that amount.

Five to six percent of revenue. Twenty percent of profit. One-fifth the capex cost of a single downstream customer’s factory. No credible challenger. Growing at double digits. Trading at food company multiples.

This is the same category error the market made with AWS inside Amazon’s retail reporting for years. The evidence was in the financials. The frame, Amazon is a retailer, was so established that the strategic significance of the cloud business was invisible to most investors until it became undeniable. Ajinomoto is a food company that controls the upstream material standard for advanced semiconductor packaging. Its most economically significant operation is a specialty chemicals monopoly generating near-monopoly margins on the insulating film inside every AI chip on earth. It is covered primarily by consumer staples analysts who model it on seasonings, frozen foods, and pharmaceutical ingredients.

The gap between those two descriptions is where the mispricing lives. And it is not unique to Ajinomoto. The broader substrate sector is being rerated from cyclical industrial to structural infrastructure, but not equally, and not all seven companies participate in that rerating. The market is beginning to recognize that Ibiden and Unimicron are not commodity manufacturers riding an AI demand wave. They are qualified infrastructure providers for a compute stack that has no near-term alternative. The qualification timeline, the advance payments, the co-design relationships, these are not just operational details. They are the mechanism by which what looks like a cyclical business is actually a structural one.

Unimicron’s single-quarter profit in Q4 2025 surpassed the combined total of the previous three quarters, that inflection is what a structural rerating looks like when it begins to show up in the financials.

Back to 1996

In 1999, when ABF first shipped commercially, Nakamura assumed it might stay in production for a decade. Twenty-six years later, every AI accelerator that powers the largest models in the world, every NVIDIA GPU, every AMD MI300X, every custom silicon chip that Google and Amazon and Microsoft run their most demanding workloads on, sits on a substrate built from his film.

“We expect sales of electronics materials, mainly ABF, to grow at an annual rate of more than 10% through 2030,” Nakamura said after taking the CEO role. “We will invest the same amount or more by 2030 as demand increases.”

He did not say: we will invest dramatically more. He said: the same amount or more. The monopolist at the top of the stack does not need to spend like the manufacturers below it. It needs only to expand its capacity modestly and let the geometry of AI packaging, more layers, larger bodies, ten times the film per chip, do the rest.

The question the market is slowly beginning to ask is whether the ABF substrate companies are commodity cyclicals or structural infrastructure. The answer, for the top tier, is the latter, and the evidence is in the advance payments, the qualification depth, and the co-design relationships that make switching suppliers a multi-year project rather than a procurement decision. Ibiden and Unimicron are not riding the AI wave. They are part of the physical architecture that the wave requires.

The harder question, the one the market has not yet asked, is what Ajinomoto actually is. The financials say semiconductor infrastructure. The valuation says food company. That gap closes eventually.

Thesis confirmation: Customer advance payments at Ibiden and Unimicron continue rising through 2026–2027 as the Gama Plant ramps, confirming AI demand conviction beyond the current cycle. Semiconductor analysts initiate standalone coverage of Ajinomoto’s functional materials business. Ajinomoto introduces separate segment reporting for electronics materials, signaling management is preparing to surface a valuation that the current conglomerate structure obscures.

Thesis break: Sekisui Chemical or a Chinese challenger successfully qualifies for high-performance ABF grades at a tier-one substrate maker by 2027, the first credible crack in a monopoly that has held for twenty-six years. Or hyperscaler AI capex guidance turns negative in consecutive quarters, unwinding the advance payment structures that currently underwrite ¥500 billion in expansion commitments across the industry.

In 1996, Nakamura was told to build something that probably wouldn’t last. He built it anyway. What he built, though neither he nor Intel fully understood it at the time, was the upstream standard for the physical infrastructure of modern compute. The food company running that standard won’t stay hidden forever.

Subscribe now

• AI has shifted the bottleneck from chip design to packaging. As GPUs get larger, hotter, and more complex, the substrate underneath them has become one of the hardest and most valuable parts of the stack.

• Ibiden’s edge is yield, not glamour. Its accumulated manufacturing know-how, customer prepayments, and position in the highest-complexity substrate segment suggest it may be more than just another cyclical semiconductor supplier.

• The core investment question is who captures the economics. Ibiden is clearly strategically important, but the stock will depend on whether rising complexity translates into lasting returns rather than just higher capital intensity.

In 1912, a company was founded on the banks of the Ibi River in Gifu Prefecture, Japan, to do something that sounds unremarkable until you think about it for a moment: turn moving water into electricity. The river ran fast, Japan needed power, and Ibiden was the kind of enterprise that exists because a natural force needs to be converted into something usable. For the next eighty years, the company stayed close to that logic. It generated electricity. It made ceramics. It built housing materials. Always the same pattern: take something difficult and make it work for someone else.

In the 1990s, Intel needed a supplier for the small circuit boards underneath its processors. Ibiden took the contract. For twenty-five years, it made substrates for Intel, and Intel accounted for 70–80% of Ibiden’s substrate revenue. It was a comfortable arrangement. Intel set the specifications. Ibiden built to them. The substrate, the multilayered organic board that connects a chip to the outside world, was necessary but unglamorous. Nobody thought much about it, because the chip above it was where the value sat.

Then, within the span of about eighteen months, two things happened. Intel stumbled. And AI made the substrate the hardest part of the chip to build.

The World Where Substrates Did Not Matter

To understand why the situation changed, you first must understand why it stayed the same for so long.

For most of the history of modern computing, progress meant shrinking transistors. Smaller transistors meant faster chips. Faster chips meant better computers. The package underneath, the substrate, was just the floor. A PC processor substrate had four to eight layers of electrical routing, measured 30–45 millimeters across, and cost under a dollar. Making one was not trivial, but it was not the constraint on anything. Five or six companies could do it at adequate yields.

Ibiden built a solid business in this world. The moat was not glamorous, it was a yield-learning curve. Ibiden did not have a patent wall or a proprietary standard. What it had was thirty years of accumulated knowledge about how to manufacture multilayer organic substrates at high yield. That knowledge translated into lower cost per good unit, which translated into Intel coming back every generation. Other companies could make substrates. They could not learn as fast at the same scale.

The market understood this and valued Ibiden accordingly. Cyclical Japanese industrial. Goes up when the PC cycle is strong, goes down when it weakens. When Intel lost share to AMD and struggled with its foundry strategy through 2023 and 2024, Ibiden’s numbers deteriorated. Revenue was flat at ¥369 billion in FY2024. Operating margins: 12.9%. The sell-side consensus: wait for the server cycle to recover, check back later.

That assessment was not irrational. It was also about to become obsolete.

When the Floor Became the Constraint

What changed was not that AI created more demand for chips. What changed was where the engineering difficulty went.

Transistor shrinks started slowing down. The industry’s response was not to give up on performance; it was to find a different way to get it. Instead of making one chip smaller, engineers started connecting multiple chips together inside a single package: compute chiplets, memory stacks, interposer layers, power delivery networks. The performance that used to come from transistor scaling now comes from how those pieces are assembled. And all of them sit on the substrate.

An Nvidia H100 GPU substrate has sixteen to twenty routing layers, exceeds 100 millimeters in body size, and must deliver 700 watts of power through dedicated planes while routing thousands of high-speed signals. The next generation, the B200, pushes to 1,000 watts, eight stacks of high-bandwidth memory, and a dual-die design that was physically impossible on the previous packaging architecture. Each generation demands a larger, more complex, more difficult substrate.

Ajinomoto, the Japanese food company whose subsidiary produces the dielectric insulating film used in virtually all high-performance substrates, has published the math: a high-performance computing substrate uses over ten times more ABF film than a PC substrate. Not ten percent more. Ten times. Both the body size and the layer count rise together, and the manufacturing difficulty compounds with each additional layer.

This is the first thing the market can get wrong about Ibiden. If you think “substrates” is one market, the company looks cyclical, dependent on total chip volumes, subject to the usual semiconductor swings. If you recognize that the advanced segment has separated from the commodity segment, AI substrates sold out with prices up 15–40%, PC substrates in oversupply with flat or declining prices, the business starts to look like something else.

The proof is in the segment data. Ibiden’s electronics operating margin expanded from 13.6% in FY2024 to 19.2% through the first nine months of FY2025, despite absorbing ¥44.5 billion in new depreciation from capacity buildouts. Margins rising while fixed costs rise. That only happens when the product mix is shifting toward something substantially more valuable.

Manufacturing Difficulty as a Compounding Advantage

This is where the strategic logic gets interesting, and where I think most analysis of Ibiden stops too early.

An ABF substrate is built one layer at a time. Each layer requires vacuum lamination of the insulating film, thermal curing, laser drilling of microvias, chemical cleaning, electroless copper seeding, photoresist patterning, electrolytic plating, resist stripping, and flash etching, all at sub-ten-micron tolerances. A sixteen-layer substrate requires this sequence six to eight times per side. Over sixty discrete process steps. Cumulative yield for a fourteen-layer substrate is estimated at 75–80%. For a twenty-layer substrate at five-micron line and space, what the latest AI accelerators require, yields are substantially lower, and the exact numbers are closely guarded.

The economics of yield at AI scale are punishing. A modern manufacturing panel holds only about sixteen AI-grade packages at 120-millimeter body size. One defective package causes a 6.25% yield loss. Compare that to a PC substrate panel, which holds roughly one hundred packages, where one defect costs about 1%. At AI complexity, yield is not a manufacturing detail. It is the entire business.

Now consider what happens over time. Each generation of AI chip demands a more complex substrate. More layers. Larger body. Finer features. Embedded components for power delivery. The number of suppliers that can manufacture these substrates at commercially viable yields shrinks with each step up in difficulty. Customers concentrate orders at the survivors. The survivors accumulate more yield-learning data from higher volumes. Their yields improve. They become better positioned to win the next generation.

This is the loop: harder packages narrow the qualified field. A narrower field concentrates volume. Concentrated volume accelerates learning. Faster learning makes the leader better equipped for the next leap in difficulty.

Ibiden’s management confirmed the mechanism in the most direct way possible. When asked about the Ono Plant, the company’s newest and most advanced facility, management said the advantage was “not so much in productivity but in yield. We are already seeing results.”

The parallel to TSMC’s position in leading-edge logic fabrication is obvious, and it clarifies the limits of the comparison. TSMC has no credible competitor at the most advanced node. Ibiden has four or five credible competitors, Unimicron, AT&S, Shinko Electric, Samsung Electro-Mechanics, Kinsus, at one generation behind. They can make substrates. Some of them are very good. But at the highest complexity, twenty-plus layers, embedded components, body sizes exceeding 100 millimeters, the field narrows sharply, and Ibiden claims 70–80% share. That is not a monopoly. But it is a frontier position in a concentrated manufacturing layer, and it appears to strengthen rather than weaken as complexity increases.

The Food Company That Controls the Supply Chain

If you want to understand why the advanced substrate market is as concentrated as it is, you must look one level up, to the most unlikely bottleneck in the entire AI hardware ecosystem.

Every high-performance substrate in the world, every one made by Ibiden, Unimicron, AT&S, Shinko, Samsung Electro-Mechanics, or anyone else, requires an insulating dielectric film called ABF, Ajinomoto Build-up Film. It is produced almost exclusively by Ajinomoto Fine-Techno, a subsidiary of the Japanese company best known for MSG seasoning and frozen dumplings. Ajinomoto holds over 95% of the global market for this material. It operates two manufacturing plants, both in Japan.

The capital asymmetry between Ajinomoto and its downstream customers is one of the most striking features of the entire semiconductor supply chain. Ajinomoto’s planned capacity expansion, roughly ¥50 billion through 2030, will increase output by about 50%. That investment is less than what Ibiden will spend on a single new substrate fab. The cheapest monopoly in the chain gates the most capital-intensive oligopoly beneath it.

Most analysts’ frame Ajinomoto’s dominance as a supply risk for substrate makers. I think this framing is backwards, at least for the incumbents.

Every substrate maker depends on the same material. But Ibiden has been Ajinomoto’s largest customer since the late 1990s. When Ajinomoto allocates scarce film during shortages, which it did during the 2020–2022 crisis, when lead times stretched to five months and prices rose 30–50%, Ibiden gets priority. A company trying to enter the advanced substrate market does not just need to build a billion-dollar fab and survive thirty-six months of customer qualification. It also needs to secure ABF film allocation from a monopolist that prioritizes its existing high-volume customers.

The Ajinomoto relationship does not just protect Ibiden from competition. It actively raises the barrier to entry for everyone else.

The genuine constraint is different, and management has been direct about it. Ibiden has secured enough ABF film for current guidance, but “not yet secured materials to accommodate potential upside.” Even if customer demand and factory capacity are available, film supply may cap how fast the company can grow. That is a real limitation. But it constrains every competitor equally, and Ibiden is best positioned within it.

If I were looking for a single leading indicator that almost nobody tracks, it would be Ajinomoto’s capacity expansion decisions and allocation priorities. Not Nvidia’s quarterly shipments. Not hyperscaler capex guidance. The food company’s factory plans.

Who Gets Paid for This?

Here is where the story becomes an investment question, and where I think the tension is genuinely unresolved.

In February 2026, Ibiden announced a ¥500 billion capital expenditure plan for AI substrate capacity. The critical detail came in the Q&A:

“Our basic policy is to conclude contracts with customers and then receive an amount equivalent to the investment as an advance payment.”

Ibiden’s balance sheet already carries ¥92.1 billion in customer advance payments. Nvidia, Intel, and hyperscalers building custom AI chips are paying for the factories. Not investing in Ibiden’s equity. Paying for the physical capacity. Intel, AMD, and Nvidia have collectively co-funded roughly 50% of the expansion capex at their top four substrate suppliers.

Customers behave this way when they believe they have no practical alternative. When the penalty for not having supply, a delayed GPU launch, a missed data center deployment, a slipped AI training cluster, exceeds the cost of funding someone else’s factory. That is what the advance payments tell you about power dynamics. Five years ago, Intel dictated specifications and Ibiden built to them. Now customers are competing to reserve Ibiden’s future capacity.

But here is the part that makes this more complicated than a simple dominance story. Despite holding 70–80% share in a sold-out market, Ibiden told analysts in October 2025 that the company was “not optimistic about product prices.” Competitors are expanding. AT&S is building in Malaysia and Austria. Samsung Electro-Mechanics is investing over a trillion won per year. Unimicron has raised annual capex to a record.

Management’s candor about pricing reveals the real question. The market does not need to be convinced that the work is hard. It does not need to be convinced that AI packaging is getting more complex or that Ibiden has a strong position. What the market needs to decide is whether Ibiden captures the economics of that difficulty, or whether it ends up doing increasingly important, increasingly expensive work while the returns flow elsewhere in the value chain.

The advance payments suggest customers think Ibiden matters. Management’s pricing caution suggests Ibiden may not be able to fully charge for how much it matters. The resolution probably comes through mix rather than rate: each generation’s substrate is larger, more complex, and higher-priced than the last, even if the pricing on any given product stays disciplined. That is a durable growth model. But it is not the same as pricing power, and the distinction matters for how much the stock is worth.

Two Companies in One Earnings Report

One reason this investment question is so hard to answer from reported results is that Ibiden still presents itself as one company when it is economically closer to two.

The AI and ASIC substrate business is growing into a tighter, higher-value market. Demand exceeds capacity. Customers fund expansion. Margins are expanding through mix.

The legacy PC, general-purpose server, and ceramics businesses are competitive, price-pressured, and in the case of ceramics, silicon carbide diesel particulate filters for European cars, structurally declining.

The blended FY2025 operating margin guide of 14.5% understates the first business and overstates the second. The Q3 FY2025 quarter “missed” expectations not because AI demand weakened, management said explicitly it did not, but because PC, server, and networking orders disappointed. The AI business delivered. The old business did not. The blended result looked mediocre. The trajectory underneath was unchanged.

By FY2028, ceramics will likely be 12–15% of revenue, a slowly shrinking business that the electronics segment is outgrowing. The risk is not that ceramics drags down the thesis. It is that blended reporting keeps making good quarters look mediocre and mediocre quarters look bad, whipsawing a stock that the market has not yet learned to read properly.

What the Stock Needs to Prove

On consensus estimates, Ibiden trades at roughly 50 times FY2026 earnings, 40 times FY2027, and 30 times FY2028. The market is already pricing significant improvement. The question is not whether the company is better than it was in 2023, obviously it is. The question is whether the earnings path from here justifies the entry point.

The variant view, and this is my view, is that the market is using the wrong time horizon and partly the wrong classification. If Ibiden’s reported return on capital understates the economic return to shareholders because customer prepayments fund a portion of the asset base, and if the Gama Plant represents a step-function increase in earnings power that is not yet in any forecast (management confirmed this explicitly in October 2025), then the forward multiples are less demanding than they appear. JPM raised its FY2028–2029 estimates in March 2026 specifically to capture Gama and new 2.5D packaging products. Consensus has not fully followed.

The market wants to treat this as a cyclical semiconductor company with AI exposure. The better frame may be a scarce, customer-reserved producer of something harder to replace than the market realizes. The gap between those two classifications is where the stock either works or doesn’t.

Three Futures

Bear, ¥5,200. Revenue compounds at 9–10% to ¥540–560 billion by FY2029. Margins settle at 14–15%. EPS: ¥200–210. At 25 times. Pricing weakens, Gama underwhelms, and Ibiden proves good but not great, difficulty keeps increasing, but the economic rent sits elsewhere.

Base, ¥9,800. Revenue compounds at 14–15% to ¥650–680 billion. Margins reach 18–19%. EPS: ¥340–350. At 28 times. Ono delivers, Gama contributes, ASIC diversification hits 15%. A very good business that earns its premium without transforming its classification.

Bull, ¥14,500. Revenue compounds at 18–19% to ¥780–820 billion. Margins hit 22–23%. EPS: ¥450–470. At 31–32 times. Gama becomes economically meaningful, the market reclassifies the business, and the old frame gives way.

What Would Prove This Right

The thesis does not need another quarter of noisy results. It needs evidence that the economics are catching up to the strategic importance.

Watch quarterly. Electronics operating margin, sustaining above 18%, or compressing under depreciation? Customer advance payments, growing toward ¥130 billion or more, or stagnating? ASIC revenue as a share of electronics, crossing 15%?

Watch semi-annually. Return on invested capital, inflecting above 10%? That single metric is the most important test. It separates “well-positioned cyclical” from “business the market has misclassified.” Gama Plant disclosures, any named customer, confirmed timeline, or financial projection? Competitor yields at twenty-plus layers, is the gap narrowing or holding?

Watch annually. Free cash flow, positive by FY2028? Glass substrate milestones, commercial production for AI applications before 2029? Ajinomoto allocation decisions, is film supply keeping pace, or becoming the binding constraint nobody planned for?

The Constraint Beneath

Ibiden is either the most expensive cyclical stock in Japan or a structural compounder the market has not yet properly valued. Those are not close to the same thing.

I think the evidence tilts toward something the market has not fully absorbed. The customers closest to the problem, the chip designers whose products cannot exist without what Ibiden builds, are behaving as if this is not temporary scarcity. They are paying for factories. They are reserving future capacity. They are co-designing substrates eighteen months before tape-out. That is not how you treat a replaceable supplier.

But management is guiding as if the advantage might not last. Pricing caution despite a sold-out market. Conservative revenue targets that exclude the most strategically important factory. Language calibrated to keep expectations in check while the balance sheet fills with customer cash.

One of them, the customers or the management, is wrong about how durable this position is. The customer advance payments will tell you which. If they keep rising, the market has underestimated what Ibiden is becoming. If they stall, the old classification was right, and the stock is expensive.

The Ibi River still runs through Gifu. The hydroelectric plants still generate a modest current, 147,340 megawatt-hours last year. The company that started by converting a river’s force into usable power is now converting manufacturing difficulty into the physical foundation of AI. Whether difficulty compounds faster than capital consumes it is the question that will determine what Ibiden is worth.

The river has not stopped flowing in 113 years. But rivers do not care about earnings multiples. The advance payments do.

Subscribe now

• We were right about the reservoir, but early on the pipes. Constellation’s nuclear fleet, PTC support, and rising capacity prices have all validated the original thesis that reliable clean power is becoming more valuable, but co-location turned out to be a slower and more regulated monetization path than we expected.

• The real update is strategic, not quarterly. Calpine changes Constellation from a seller of clean power into a broader power orchestration platform: nuclear, gas, batteries, demand response, and commercial relationships can now be packaged into customer-specific solutions.

• The stock is being priced for delay, not destination. The market sold the name because there was no new hyperscaler deal and 2026 guidance only met the long-term framework, but the bigger story is that Constellation’s earnings flywheel and contracting inventory are stronger than before.

“I recognize that the last time we spoke, I indicated that we expected to be done with an important transaction by this call, but we’re not ready to announce anything today.”

That was CEO Joe Dominguez on March 31, 2026, the same day Constellation Energy introduced a 20% base earnings growth commitment through 2029, authorized $5 billion in share repurchases, and laid out the most structurally sound long-term financial framework in the power sector. The stock fell 9.3%.

The market wanted a hyperscaler deal. It got a platform thesis. In the language of our previous analysis, the reservoir is fuller than we expected. The plumbing turned out to need permits nobody anticipated. The question that matters, the only question, really, is which of those two facts determines what the company is worth.

What We Got Right, What We Got Wrong

In August 2025, we published “The Reservoir Company,” arguing that Constellation wasn’t in the business of selling electricity. It was selling pressure, reliable, schedulable, carbon-free power available every hour of every day. Nuclear plants were the reservoir, long-term contracts the pipes, the production tax credit the levee that prevents the water level from falling, and PJM capacity revenues the toll paid for readiness.

That framework proved more right than we expected. Nuclear capacity factor held at 94.7%, four points above the industry average, for the second consecutive year. The PTC floor continued to accrete with inflation, reaching roughly $44.75 per megawatt-hour. PJM capacity prices in the Eastern Mid-Atlantic zone surged from $52 to $270 per megawatt-day, a fivefold increase that validates the “toll” mechanism we described. Long-term contracted clean output for 2030 delivery tripled, from 12 million to 48 million megawatt-hours. The thesis that reliable pressure would command an increasing premium is now consensus.

Where we were wrong was about the pipes. We framed behind-the-meter co-location, physically wiring a data center to a nuclear plant, as the primary monetization path. It turned out that FERC’s rejection of the Talen interconnection agreement, followed by PJM’s unworkable compliance proposal and the executive order that forced hyperscalers to renegotiate contract terms, created a regulatory thicket that made direct co-location impractical at scale. We also didn’t anticipate that the Calpine acquisition would close with DOJ-mandated divestitures of highly efficient gas plants, creating near-term earnings noise that obscured the deal’s strategic value.

Our thesis was directionally correct but the timeline was too aggressive. The value is still there, arguably more of it now. But the path to realization runs through regulatory resolution and management execution, not just asset scarcity.

The Grid as an API

The strategic development that matters most since our last writing is not any single deal or number. It’s the transformation of what Constellation actually sells.

Consider an analogy from enterprise computing. In the early days, companies that needed secure, high-speed connectivity would build physical cross-connects, direct wires between their servers and their partners. It was fast, it was private, and it completely bypassed public network costs. It also couldn’t scale. Every new connection required bespoke infrastructure and manual routing.

The industry solved this by abstracting the network. Instead of physical cross-connects, companies used the public internet as a platform, paying modest transit fees in exchange for infinite speed, flexibility, and reach. The value shifted from owning the wire to owning the application that ran on it.

The U.S. power grid is undergoing exactly this transition, and Constellation is the company best positioned to exploit it.

Faced with a regulatory blockade on behind-the-meter co-location, Constellation is pivoting to front-of-the-meter deals that use the existing PJM transmission grid as the interconnection layer. The mechanics are straightforward: Constellation injects clean, firm power into the grid from its 180-million-megawatt-hour nuclear fleet. The hyperscaler builds its data center wherever zoning is friendly and permitting is fast, anywhere in the PJM footprint, and draws that power down. Hourly matching software verifies the clean energy delivery.

This is not a retreat from the data center thesis. It’s an architectural upgrade. Constellation can now execute deals anywhere in the grid without waiting for bespoke FERC rulings on each site. The hyperscaler pays modest transmission fees, a cost they’ll absorb willingly because their bottleneck is time, not electricity prices. In the AI arms race, waiting three years for co-location approval is far more expensive than paying grid charges.

And this is why Calpine changes the company’s architecture, not just its income statement. Before Calpine, Constellation sold clean power. After Calpine, it sells orchestration, the ability to package nuclear baseload, gas flexibility, battery storage, demand response, and structured contracts into a single customer solution that solves for clean, firm, fast, and affordable simultaneously. Dominguez described this on the call: some customers want nuclear paired with batteries and demand response; others will pay for backstop capacity from PJM and buy energy and attributes directly; still others will use AI-enabled load shifting to curtail during peak hours. The point is that no two deals look the same, and Constellation is the only company that can structure all of them.

The historical parallel is Amazon Web Services circa 2008. Early cloud computing was a single product, virtual machines. AWS became dominant when it realized the real business was not selling compute, but selling a platform of services that made compute, storage, networking, and deployment work together. Once Amazon made that shift, no competitor could replicate the full stack.

Constellation is making an analogous move. The 190 million megawatt-hours of commercial distribution, the relationships with 80% of the Fortune 100, the 55-gigawatt fleet spanning nuclear, gas, geothermal, and renewables, the investment-grade balance sheet, these are not isolated assets. They are a system. And systems are harder to replicate than plants.

The Metric That Tells the Real Story

The most important thing Constellation reported was not 2025 adjusted operating earnings of $9.39 per share, or 2026 guidance of $11 to $12. It was the formalization of a new earnings architecture.

Management now wants the company understood through two lenses: base earnings and enhanced earnings. Base EPS includes long-term contracted generation, uncontracted nuclear priced at the PTC floor with 2% inflation, minimum expected gas fleet margins, and commercial margins based on a ten-year weighted average. Enhanced earnings are everything above that, power and capacity prices exceeding the PTC floor, wider spark spreads, stronger commercial optimization.

This distinction matters because it reveals what management is actually arguing. Companies introduce new metrics when they want the market to see a different business. The old Constellation was a strong merchant-contract hybrid. The new Constellation wants to be valued as a contracted platform with a durable earnings floor and a large inventory of unmonetized optionality. Management is making a re-rating argument without quite calling it one.

The argument has a lot of truth in it. The 20% base EPS CAGR through 2029 deliberately excludes every discretionary lever, the $5 billion buyback, all unannounced contracts, any inflation above 2%, and any gas fleet utilization improvement. CFO Shane Smith confirmed that free cash flow in 2028 and 2029 simply earns interest income in the projections. Management built a framework designed to be beaten, using the same playbook that has produced four consecutive years of guidance outperformance.

But here is where intellectual honesty matters. Enhanced earnings are still roughly 40% of 2026 EPS. That means the company is not yet the pure infrastructure annuity the new framework implies. It is moving in that direction, enhanced earnings should decline to 30–35% of the total over time as more contracts move into the base, but the migration is not complete. The opportunity for investors lies in whether that migration happens faster than the market expects. The risk lies in whether it stalls.

The Number That Nobody Noticed

Buried in the statistics section of the Q4 earnings release is a data point that deserves more attention than it received.

In Q4 2025, Constellation’s nuclear fleet experienced 30 non-refueling outage days, compared to 3 in Q4 2024. Full-year non-refueling outage days rose from 36 to 57. The quarterly capacity factor slipped to 93.1%.

For any other company, this would be a footnote. For Constellation, it is the most important operational metric in the business. The entire thesis depends on the nuclear fleet running at best-in-class levels. Management has argued for years that their four-percentage-point capacity factor advantage over the industry translates into roughly 8 million additional megawatt-hours annually, effectively an extra reactor, generated for free by operational excellence.

One lumpy quarter doesn’t break a thesis. But the trend, 36 outage days in 2024, 57 in 2025, warrants monitoring, especially as management simultaneously integrates a $26.6 billion acquisition, restarts a shuttered nuclear plant, navigates FERC proceedings, and negotiates hyperscaler contracts. The “best operator in the world” narrative gets tested when every plate is spinning at once.

What the Reservoir Is Worth

Narrative without numbers is just storytelling. Here is what we think Constellation is worth under three realities, using a three-year horizon.

Bull case, $480 to $520 (30% probability)

Revenue compounds at roughly 12% as full Calpine contribution layers onto two or three new premium PPAs and an on-schedule Crane restart. Adjusted operating EPS reaches approximately $21 by 2028 as EBITDA margins expand past 32%, driven by base earnings dominance, rising gas utilization, and PTC inflation adjustments at 3%+. The market re-rates the stock to 23–25 times forward earnings, recognizing Constellation as an infrastructure compounder rather than a utility. The FERC approval of capacity injection rights transfer for Crane pulls forward the restart timeline. The $5 billion buyback executes at prices below $320, mechanically adding 450–500 basis points to EPS growth beyond the stated CAGR. This scenario requires regulatory resolution, hyperscaler deal execution, and sustained fleet reliability above 94%.

Base case, $370 to $410 (50% probability)

Revenue compounds at 8–9% on Calpine consolidation, modest new contracting, and organic growth. EPS reaches $17–18 by 2028 as margins expand to 29–31% and enhanced earnings gradually migrate into the base. The market values the stock at 21–23 times forward earnings, a premium to utilities but not yet full infrastructure-compounder pricing. One new major PPA is signed by mid-2027. Crane begins operations by late 2028. PJM capacity prices moderate but stay well above historical averages. This is roughly where Bloomberg consensus sits today, and it implies a 12–15% annualized return from current levels including dividends.

Bear case, $200 to $240 (20% probability)

No new major PPAs materialize through 2027. Enhanced earnings normalize faster than new contracts replace them. Gas fleet utilization disappoints. Crane is formally delayed past 2029. Nuclear fleet capacity factor slips below 93% on a sustained basis as management attention is divided across too many priorities. The market re-rates to 14–16 times forward earnings, traditional IPP territory, on EPS of roughly $14–15 in 2028. The PTC floor and buyback authorization provide mechanical support, preventing a collapse below $200, but the premium multiple evaporates.

Probability-weighted three-year target: approximately $380. At $279 today, that implies roughly 36% total return over three years including dividends, attractive, but the entry point matters. Below $260, the math becomes compelling across all scenarios. Below $240, the PTC floor and buyback create a hard bid.

The Scoreboard

The temptation with a company this complex is to track everything. The discipline is to track only what actually determines whether the thesis is working.

First, contracted megawatt-hours for 2030 delivery. Currently 48 million, up from 12 million a year ago. Above 70 million by year-end 2026 means the contracting engine is converting scarcity into base earnings. Below 55 million means the regulatory friction is winning.

Second, the enhanced-to-base earnings mix. Currently 40% enhanced. Below 35% by 2027 means the infrastructure compounder thesis is real and the multiple should expand. Still above 40% means the market is right to withhold the premium.

Third, nuclear fleet capacity factor. Sustained above 94% means the crown jewel is intact despite integration demands. Two consecutive quarters below 92% means management has a resource allocation problem that directly impairs the earnings trajectory.

Fourth, the FERC ruling on the Eddystone-to-Crane capacity injection rights transfer. This is the single highest-impact near-term catalyst that almost nobody is discussing. Approval collapses the four-year gap between Crane’s physical readiness in 2027 and PJM’s 2031 full deliverability timeline. Denial preserves the most visible bear case on the stock.

Fifth, buyback execution. More than $2 billion retired by year-end 2026 at prices below $300 means management is backing conviction with capital. Token execution at higher prices means signaling, not commitment.

That is the scoreboard. Not the next quarter’s earnings per share. Not the next rumor about a hyperscaler deal. The scoreboard.

The Bet

Last August, we argued that owning the reservoir beats chasing the weather. That thesis is stronger today, not weaker. The reservoir is larger, Calpine added the pumping stations. The levees are higher, PTC inflation adjustments are accreting mechanically. The tolls are more lucrative, capacity prices are five times their prior-year level. And the demand for pressure has never been greater, as artificial intelligence workloads collide with a grid that wasn’t built to serve them.

What changed is that the market learned something we should have emphasized more: reservoirs are only valuable if the pipes reach the customers. Constellation is building those pipes, through front-of-meter deals, regulatory filings, AI-enabled flexibility partnerships, and the broadest commercial platform in the power sector. But building pipes in a regulated industry takes longer than filling reservoirs in a scarce one.

The variant perception is this: the market sees a nuclear utility that missed its guidance and failed to announce a deal. The correct framing is a power orchestration platform whose monetization set just widened, whose base earnings are growing at 20% with every optional lever excluded, and whose stock is being sold by momentum investors who bought the AI narrative and left when it didn’t arrive on their schedule. The market is pricing the delay. It should be pricing the destination.

We’d accumulate below $280 and add aggressively below $260. The reservoir is full. The pipes are coming. Patience is the trade.

Subscribe now

• The debate has shifted. The old question was whether Zhen Ding had real AI exposure; the new one is when AI server, optical, and substrate become financially undeniable rather than merely strategically interesting.

• “One ZDT” is the real thesis. The moat is not “PCBs,” but a manufacturing system that can absorb rising interconnect complexity across substrates, server boards, optical, and mobile, with Thailand turning geography into part of the product.

• The proof points are mix, margins, and cash flow. If AI/server/optical/substrate mix keeps climbing and gross margins stay above the depreciation wave, the stock can re-rate; if not, it stays trapped in the market’s old mobile-PCB classification.

“Zhen Ding Technology’s sales growth is set to accelerate in 2026 as it enters a new expansion cycle... Gross margin should expand despite rising depreciation as high-end AI volume and improved utilization outpace expenses.”,

• Bloomberg Intelligence, March 2026

Our January view was that Zhen Ding’s edge was not “PCBs” but manufacturing scale: the ability to turn increasingly exotic interconnect technologies into reliable, high-volume production. The key triggers were ABF yield parity, Thailand readiness, and the quarter when AI server plus substrate revenue crossed a credibility threshold of roughly 25% of mix. I still think that frame was right. What I would update now is not the direction of the thesis, but the burden of proof.

The company has done enough to prove the AI story is real. It has not yet done enough to prove the AI story is large enough to dominate the financials.

That distinction matters because FY2025 was precisely the kind of year that reveals a transition in progress. Revenue reached a record NT$182.5 billion, gross margin improved to 19.8%, operating margin improved to 7.6%, and Q4 gross margin hit 22.6%, expanding 220 basis points year-over-year in a quarter where revenue declined. AI-related applications accounted for nearly 70% of revenue. IC substrate revenue grew 31.3%. Management said 2026 would begin a “new high-growth cycle” and committed more than NT$50 billion to ten new facilities. But parent net income fell 26%, EPS fell to NT$6.91, and free cash flow turned negative. That is not the signature of a broken business. It is the signature of a business whose strategic center is improving faster than its reported earnings.

The Burden of Proof Has Shifted

The old debate was whether Zhen Ding had meaningful AI exposure or whether this was mostly another mobile-cycle stock with a nice deck. That debate is over. Management made a very specific case: iHDI and HLC products for GPU and ASIC customers are entering mass production; optical demand is driven by 800G and 1.6T boards with customers already reserving 2027 capacity; ABF capability has reached 158×158mm at 28 layers; and the company is spending at a rate that only makes sense if customer commitments are binding.

The updated debate is simpler and harsher: when does this become financially undeniable?

That is the real significance of the 25% credibility threshold. It was never a magic number. It was a way of saying the market would not reclassify the stock until AI server plus substrate revenue became too big to dismiss as interesting-but-immaterial. FY2025 reinforced that judgment. The strategic businesses are rising, combined server/optical/IC substrate reached 11.6% of revenue, up from 9.9% in 2024 and 6.2% in 2022. They are not yet the consolidated story. Mobile remained 61.3% of 2025 revenue. The company is still, financially speaking, what the market thinks it is.

By 3Q FY2026, the burden must shift from narrative to conversion. Not all at once, and not perfectly. But enough that the stock is no longer valued primarily as a mobile supplier with an AI call option.

The Metric That Will Settle It

Source: Company filings. Quarterly mix estimated from annual disclosures. FY24 and FY25 mix figures are reported.

The metric that best reveals Zhen Ding’s trajectory is the decoupling of gross margin from revenue growth. In 4Q25, revenue declined yet margins expanded 220 basis points. That only happens when you’re selling different, higher-value products. The question for 3Q26 is whether that decoupling survived the seasonal trough of 1H26 and the accelerating depreciation from the NT$50 billion capex program.

If 3Q26 gross margin is above 24%, the depreciation wave has been absorbed by revenue growth and the margin expansion is structural. If it’s between 21-23%, the transition is working but the old business still dominates the margin profile. If it’s below 20%, the depreciation is winning and the bears are right to question the capex timing.

The competitor comparison sharpens this. In Q4 2025, Zhen Ding’s gross margin of 22.6% was nearly 700 basis points higher than Unimicron’s 15.8%, yet the market treated Unimicron as the “substrate pure-play” and Zhen Ding as the “mobile company.” The gap is classificational, not operational. By 3Q26, I want to see whether that gap has held, narrowed, or widened. If Zhen Ding is still running structurally higher margins than the company the market gives the AI premium to, the mispricing becomes harder to ignore.

“One ZDT” Is a Theory of Industrial Power

The wrong way to hear “One ZDT” is as generic bundling. The right way is as a claim about what becomes valuable when computing fragments into more layers: flex, HDI, HLC, server boards, optical interconnect, BT, ABF, foldables, AI glasses, and eventually technologies that blur the line between substrate and board entirely. Zhen Ding’s product is not the individual board. Its product is the manufacturing system that can absorb that complexity.

This is the same logic that made TSMC indispensable, not because it designed the best chips, but because it could manufacture anyone’s design at scale with acceptable yields. Zhen Ding is making the same structural argument one layer down in the stack. In a world where compute is increasingly commoditized but the physical interconnect is increasingly constrained, the company that can reliably produce 28-layer substrates, 1.6T optical boards, and server backplanes across multiple geographies becomes a utility. Utilities that are essential and capacity-constrained do not trade at cyclical hardware multiples.

Thailand is the proof that this argument is landing with customers. Building greenfield manufacturing in a new country is not adding capacity. It is building a geopolitical product attribute, something customers pay for, not just a cost of doing business. By 3Q26, we should know whether Thailand is at meaningful utilization and whether non-China-sourced boards command a pricing premium. If they do, geography has become a moat. That’s new for PCBs.

The mobile business, meanwhile, deserves a reframe. The memory squeeze was supposed to be the headwind that sank the legacy base. But if foldables and AI-edge devices are driving higher PCB content per unit, enough to offset the volume decline, then mobile isn’t a “legacy drag.” It’s a high-precision manufacturing franchise whose technical requirements are converging with server-grade complexity. The same yield-curve expertise developed for ABF substrates gets ported back into mobile FPC, widening the gap against second-tier players. The cash cow isn’t dying. It’s evolving.

The Financial Structure Problem

Even if every operational thesis is confirmed, there is a gap between what the business is becoming and what equity holders actually receive. This is the part of the Zhen Ding story that bulls tend to skip and that I want to be explicit about.

The cash flow trajectory is the most important resolution by 3Q26. In March, the biggest analytical disagreement wasn’t about revenue or margins, it was about cash. GF Securities modeled FY2026 operating cash flow at NT$55.9 billion. Bloomberg consensus modeled roughly NT$28.7 billion. A NT$27 billion divergence, about 11% of the entire market cap. GF saw a company funding its own transformation. Consensus saw a company levering up to NT$27.5 billion net debt. By 3Q26, we have three quarters of actual data. The answer to this question determines whether the thesis is “high-return growth funded by operations” or “leveraged bet that requires perfect demand to avoid a deleveraging crisis.”

The minority interest problem is exactly as structural as I warned in March. In FY2025, minority partners in Chinese subsidiaries took NT$3.82 billion, 36% of net income. For every dollar of operating improvement, shareholders receive roughly 64 cents. This isn’t changing. It won’t change. The JV structures provide political cover and local government relationships essential for operating in China. The leakage scales with success: as the Chinese operations become more profitable from AI and substrate production, the absolute dollar amount grows even if the percentage holds. I now believe this should be modeled as a permanent 20% valuation discount to operating-level metrics.

The depreciation wave is the mechanical test. D&A was NT$17.9 billion in FY2025 and is rising toward NT$24+ billion annualized. If revenue ramps fast enough, depreciation is invisible, absorbed into a growing base. If revenue growth disappoints while the depreciation schedule is locked in, the margin expansion reverses mechanically. The depreciation doesn’t care about your strategy. It shows up every quarter regardless. By 3Q26, the early evidence should show which path the company is on.

The point is not to be bearish. It is to be honest about the distance between “the business is transitioning” and “shareholders are getting paid.”

Variant Perception

Consensus still tends to see a smartphone-heavy manufacturer with a better story than its financials. Bloomberg Intelligence makes essentially this argument: growth accelerates, but high-growth segments remain limited in contribution and margins stay below peers because of depreciation.

The better framing is that Zhen Ding is not becoming an AI pure-play. It is becoming an AI-era manufacturing utility for the interconnect layer of computing. If you expect substrate pure-play margins, you will be disappointed. If you recognize that the moat is the ability to industrialize difficult physical systems across multiple categories, and to finance that with a still-large mobile base, then the company starts to look less like a cyclical component maker and more like infrastructure with a lagging valuation framework.

The classificational arbitrage remains the most concrete version of this argument. If Zhen Ding still runs higher gross margins than Unimicron while trading at a lower multiple, and the only explanation is Bloomberg database tags, then the mispricing is real and the catalyst for closing it is knowable: mix crossing 20-25%.

Three Years, Three Prices

Bull requires AI capex sustaining through 2028, ABF yields reaching Ibiden-adjacent levels, Thailand commanding a geopolitical premium, and server/substrate mix exceeding 30%. Zhen Ding becomes the TSMC of interconnects.

Base assumes the transition is real but incomplete. Mix reaches mid-20s%, margins improve but stay below Ibiden, mobile stabilizes through premiumization, and FCF turns positive in FY2027. The stock re-rates from mobile-PCB multiples to “industrial compounder with AI upside.” This is the most likely outcome.

Bear requires AI capex to decelerate before new capacity is utilized, or T-glass supply constraints from Nitto Boseki (whose Fukushima expansion doesn’t arrive until early 2027) to leave factories idle. Depreciation crushes margins, net debt climbs past NT$40 billion, and the company spends the late 2020s deleveraging. This is trackable: watch Nitto Boseki’s timeline, NVIDIA’s platform guidance, and hyperscaler quarterly capex disclosures.

The single variable that governs all three scenarios: sustained hyperscaler AI infrastructure spending through 2028.

What Confirms It, What Breaks It

Four items. Not eight. Stated plainly.

Gross margin above 24% for two consecutive quarters. Confirms the depreciation wave has been absorbed by revenue growth. We got one quarter at 22.6% in 4Q25. We need to see it sustained above 24% as the NT$50bn capex comes online. Structural confirmation.

ABF operating margin above 10% by first half 2027. Confirms the substrate business is a real profit center, not just a strategic initiative that breaks even on a good day. The start-up tax was NT$1.50 per share. The return on that investment needs to be visible in segment profitability.

T-glass supply secured through 2028. The single most concrete supply-side risk. If Nitto Boseki’s Fukushima expansion arrives on schedule and Zhen Ding has allocation, the bear case loses its sharpest edge. If it slips or allocation is insufficient, the factories sit idle regardless of demand. Track Nitto Boseki timelines and Taiwan Glass/Fulltech Fiber Glass qualification progress.

Server/optical/substrate mix crossing 25%. The mechanical threshold for algorithmic reclassification. Not analyst opinion, database categorization. At 11.6% in FY2025, the company is on the right trajectory but needs another 12-18 months. The pace of that climb through 3Q26 will tell us whether the 2028 bull case is realistic or aspirational.

Our January view was that Zhen Ding’s second act would become obvious when the market stopped seeing a PCB company and started seeing an interconnect platform for the AI era. My updated view is only slightly different, but more demanding: that second act is real, but by 3Q FY2026 the argument must show up in financial conversion, not just strategic coherence. The margins must hold. The cash flow must inflect. The mix must keep climbing.

The story no longer needs belief. It needs proof.

Subscribe now

• The market focused on Ping An’s RMB 8 billion buyback, but it was too small to signal real capital allocation conviction.

• The more important development was CSM holding at roughly RMB 726 billion, the first time the backbook has stopped shrinking since 2022.

• That shifts the thesis: Ping An no longer needs just a management catalyst, because the underlying operating model may have crossed into self-sustaining earnings growth.

Everyone was waiting for the buyback. The real story was hiding in the backbook.

Ping An Insurance Group reported first-half 2026 new business value growth of 17.8%, a deceleration from the 29.3% pace of full-year 2025. The company announced an RMB 8 billion share repurchase program and disclosed that its contractual service margin stabilized at RMB 726 billion, the first period without decline in three years.

— Bloomberg News, August 2026

My first thought was: eight billion. That’s it.

For eighteen months I’ve argued that the single most important catalyst for Ping An’s stock was a buyback large enough to signal management conviction, RMB 15 billion or more, enough to tell the market that the people with the best information think the stock is cheap. Instead, management announced RMB 8 billion: not nothing, but not conviction either. A gesture dressed up as a decision.

The stock rose 3% on the day, gave it back within a week, and settled roughly where it started. The market’s verdict was clear: token. And I was ready to write that piece, the one about how Ping An keeps passing the test on operations and failing it on capital allocation.

Then I looked at the CSM line.

The contractual service margin, the backbook’s reservoir of locked-in future profits, came in at RMB 726 billion. That’s essentially flat versus the RMB 725 billion at year-end 2025. This sounds unremarkable. It is, in fact, the most important number Ping An has reported since the agent restructuring began, because it means the backbook has stopped shrinking for the first time since 2022.

The reason this matters more than the buyback is that CSM is the structural ceiling on Ping An’s most important profit line. The insurance service result, RMB 84 billion in FY2025, the single largest component of Life & Health operating profit, is mechanically driven by CSM releases. If the CSM keeps declining, that profit line declines with it, regardless of how fast NBV grows. If CSM stabilizes, the ceiling lifts. And if CSM starts growing again, Ping An’s earnings power compounds in a way that no buyback can replicate.

Everyone was watching the buyback. The real story was the backbook.

What We Said, and What Changed

In January 2026, I wrote “The Trust Bundler,” arguing that Ping An’s collapsing agent force wasn’t a crisis but a strategy, the company was replacing expensive, low-trust distribution with a healthcare service ecosystem that created stickier, higher-value customer relationships. In March, after FY2025 results, I updated that thesis in “The Convertible Bond Test”: the operating model was proven (NBV +29.3%, persistency at 97.4%), but two new risks had emerged that I’d underweighted, the aggressive doubling of equity allocation to 19% of the portfolio, and management’s refusal to signal capital return conviction.

My framework was clear: the stock needed a catalyst, and that catalyst was capital allocation, not operations. I was half right. The operating improvement was necessary but insufficient for the stock. But I was wrong about what would make the operating improvement sufficient. I assumed it was a management decision, a buyback. What actually happened was more interesting: the operating improvement itself crossed a structural threshold. The service-attached products, which command 1.5x to 23x the premium of standard policies, finally reached enough scale to replenish the CSM backbook at roughly the rate it was being released.

Source: Company filings, author estimates for 1H2026.

The 17.8% NBV growth is a deceleration from 29.3%, and bears will seize on it. They’re missing the point. The NBV rate was always going to normalize, you can’t grow bancassurance 138% forever, and the agency channel’s productivity gains from culling weak agents are largely complete. What matters now is whether the quality of new business is good enough to sustain the backbook. For the first time, the answer appears to be yes.

Why CSM Stabilization Changes the Math

I need to explain why this matters mechanically, because it’s the kind of accounting detail that gets ignored in favor of flashier metrics.

Ping An’s Life & Health operating profit comes from three sources: CSM release (the biggest), risk adjustment release, and investment service result. In FY2025, CSM release was RMB 68.2 billion, more than half of L&H operating profit. But the CSM balance had been declining for three straight years (RMB 819B → 768B → 731B → 725B), which meant the release base was shrinking. Even if the release rate held steady at ~8.6%, the absolute dollar amount would decline because the pool was getting smaller.

CSM stabilization reverses that dynamic. If the balance holds at ~RMB 725B and the release rate stays at 8.6%, annual CSM release is ~RMB 62B, a stable, predictable profit stream. If new business CSM margin recovers (8.8% in 1H26 vs. 8.3% in FY25, driven by higher-margin service-attached products), the balance could start growing again in 2027. That would mean the insurance service result, which declined 2.3% in FY2025, inflects to growth. And that inflection would be structural, not cyclical.

This is the connection to our original thesis that I didn’t fully articulate before. The Britannica-to-library metaphor was about distribution: replacing agents with services. But the CSM stabilization shows that the service model isn’t just changing how Ping An sells insurance, it’s changing the economics of what it sells. Service-attached products have higher persistency (97.6%), which means fewer lapses eating into the CSM. They have higher FYP (1.5x to 23x uplift), which means more CSM generated per policy. And they attract customers who buy more contracts (2.94 per customer, with 5-year customers holding 1.7x more), which means the cross-sell feeds back into backbook stability.

The library metaphor extends: the library isn’t just attracting more visitors. The visitors are staying longer, borrowing more books, and renewing their memberships at 97% rates. At some point, membership revenue exceeds the cost of new acquisitions. That point, I think, is now.

The Token and the Ceiling

So what about the buyback?

RMB 8 billion on a market cap of roughly RMB 1 trillion is 0.8% of shares outstanding. It’s better than nothing and worse than meaningful. For context, the threshold I set in March was RMB 15 billion, the amount that would signal genuine conviction rather than board-level box-checking. Eight billion signals that the capital return debate has started inside the company, but hasn’t been won.

The most important thing CFO Fu Xin said during the briefing was this: “We will evaluate the pace and scale of the repurchase program based on market conditions and regulatory considerations.” That’s boilerplate, but the word “pace” matters. It tells you management is framing this as an ongoing program, not a one-time event. The question is whether RMB 8 billion is the first tranche of something larger, or the entire commitment.

I lean toward the former, for one reason: the convertible bond proceeds are finally being deployed. Management disclosed that approximately 60% of the proceeds have been allocated, partly to the buyback, partly to a healthcare infrastructure investment in the Yangtze River Delta region. This isn’t the transformative hospital acquisition that would change the narrative overnight, but it’s directionally correct. Cash that was sitting idle for two years is moving. That’s new.

The stock’s muted reaction makes sense through this lens. The market wanted RMB 20 billion and got eight. But the market also got CSM stabilization and partial convertible deployment, two things that weren’t in anyone’s model. The net effect is that the base case got more probable and the bear case got less probable, without the bull case being triggered. That’s worth a modest re-rating over time, not a single-day pop.

Three Years From Here

The value of scenario analysis is clarifying what you’re betting on, not pretending you can predict China’s equity market. Here’s how I think about Ping An from mid-2026 through mid-2029:

What changed since March: I’ve shifted roughly 5 percentage points from bear to base. The CSM stabilization removes the structural headwind I was most worried about. The partial convertible deployment, while underwhelming, eliminates the worst interpretation (that management was hoarding cash against undisclosed losses). The equity portfolio survived a volatile first half without forced de-risking.

The variant perception hasn’t changed in kind, but it’s sharpened in degree. Consensus still sees a cheap insurer with conglomerate complexity. I think the better frame is a service utility under construction, now showing its first evidence of self-sustaining economics in the backbook. The gap between these views is still the stock, but the evidence pile supporting the service-utility frame just got meaningfully heavier.

Probability-weighted, the expected three-year target is roughly HKD 85-90, or about 55-65% total return including dividends, approximately 16-18% annualized. That’s attractive for an income-generating position with improving fundamentals.

What I’m Watching

Three signals, in order of what they’d make me do:

Buyback scaling above RMB 15B cumulative, confirms that the RMB 8B was a first tranche, not a gesture. This is the clearest path from base to bull. If the buyback stalls at 8B through year-end 2026, the “management will never fully commit” narrative reasserts itself.

CSM trajectory at year-end 2026, if the balance holds above RMB 725B or grows, the structural bear case is dead. If it resumes declining toward RMB 700B, the 1H stabilization was a one-off, and the operating profit ceiling is still falling.

Net investment yield versus the 4.0% EV assumption, this remains the silent risk nobody has incentive to flag. If net yield stays below 3.5% for two consecutive reporting periods, the embedded value anchor needs to be rebuilt on a lower base, and every “cheap at 0.7x EV” argument loses its foundation.

The Library, Revisited

In January I wrote that Ping An was building a library to replace the encyclopedia salesman. In March I wrote that the library was open but management wouldn’t let shareholders use it. After this quarter, I’d put it differently.

The library is working. People are coming in, staying longer, and renewing their memberships at rates that would make any subscription business envious. The lending desk is finally staffed, if modestly. And for the first time, the rate of new book acquisitions is keeping pace with the rate of borrowing, the collection is no longer shrinking.

The stock still trades like a building with a “For Sale” sign on it. I think it’s becoming a franchise. The single thing that would make me more confident is a second buyback tranche by March 2027. The single thing that would make me less confident is the CSM resuming its decline. Until one of those happens, I own the position, collect the dividend, and wait for the market to walk through the door.

Subscribe now

· Rolls-Royce’s FY2025 results suggest this is no longer a simple recovery story: Civil Aerospace margins hit 20.5%, free cash flow reached £3.27bn, and management raised mid-term targets while announcing a multi-year buyback.

· The key new insight is in LTSA economics: management showed the cash value of long-term service agreements has more than quadrupled since 2022, with roughly 75% of the incremental value still unrealised by 2028.

· The article’s variant view is that the market still treats Rolls-Royce like a cyclical industrial, when it increasingly looks like a mission-critical installed-base compounder whose pricing power, durability gains, fleet growth, and cost efficiency reinforce each other over time.

What We Said, What Happened, What Changed

In January, we published The Physics of Pricing Power, arguing that Rolls-Royce had completed one of the most remarkable transformations in industrial history. The thesis was straightforward: Rolls-Royce had spent a decade buying exclusivity on the A350 and A330neo, locking airlines into TotalCare contracts they could not exit, and then, under Tufan Erginbilgic, had finally exercised the pricing power it had spent twenty years refusing to use. The turnaround wasn’t cost-cutting. It was monetising switching costs.

We believed every word of that thesis. We also said the stock, at 1,273p, left no room for error. Our probability-weighted return was negative. The opportunity, we wrote, “lies not in the thesis but in patience.”

On 26 February, Rolls-Royce published the best results in its 120-year history. Operating profit reached £3.46 billion, five times the 2022 level. Free cash flow hit £3.27 billion. Civil Aerospace margins reached 20.5 percent. Every mid-term target was raised by 25 to 40 percent. A £7 to £9 billion buyback, the first multi-year programme in the company’s history, was announced. The stock hit an all-time high of 1,420p.

Then a war started in Iran, Middle East airspace closed, and the stock fell 19 percent in three weeks.

We got the business right and the stock setup wrong, or rather, we got it wrong for approximately four hours before geopolitics corrected the price for us. The entry point we said to wait for arrived, but not because the business disappointed. It arrived because a conflict 4,000 miles away temporarily grounded some flights.

That is interesting. But it is not the most interesting thing that happened. The most interesting thing was a chart that Erginbilgic showed during the earnings presentation, a chart that was not in the press release, not in the financial tables, and that most analysts have still not fully incorporated into their models. That chart is the reason our view has changed.

The Annuity Reveals Itself

The chart showed that the cash value of Rolls-Royce’s LTSA contracts, the long-term service agreements under which airlines pay per engine flying hour, has more than quadrupled since 2022. More than two-thirds of that increase came from management actions: contract renegotiations, time-on-wing engineering improvements, and EFH rate optimisation. Volume recovery contributed less than a third.

By 2028, only 25 percent of the incremental cash associated with those improved contracts will have been realised.

Read that again. The installed base of 6,000-plus widebody engines, flying under contracts that airlines cannot exit, on aircraft where no alternative engine exists, has been re-priced so aggressively that the embedded future cash flows have quadrupled, and three-quarters of that value has not arrived yet. Moreover, only half of LTSA contracts will be on new terms by 2028, meaning the re-pricing runway extends well into the 2030s.

Our January piece described Rolls-Royce’s economics correctly. We called it a toll booth. What we missed was that the toll booth was still being expanded. We treated the pricing power as a static advantage, the company had monopoly positions and was finally charging for them. The FY2025 disclosure revealed it is a compounding advantage, a flywheel with four interlocking elements that reinforce each other at scale.

The first element is time-on-wing. Rolls-Royce has now raised its durability target to more than 100 percent improvement across in-production Trent engines, with more than half already delivered. The Trent XWB-84 life extension programme completes in 2026. The practical consequence is that engines now fly roughly twice as long between shop visits. During those extended intervals, the airline pays per flying hour under the LTSA. Cash accumulates longer before the next cost event. The cost event itself is cheaper because visit frequency falls. Management quantified the result: shop visits peak in 2026 at 1,480 to 1,550, then decline to 1,300 to 1,400 by 2028. Over the same period, operating profit grows by roughly £1.5 billion.

Fewer shop visits. Higher profitability. That inversion is not a recovery. It is the flywheel made visible in the guidance.

The second element is contract re-pricing. Each LTSA that renews or renegotiates comes in at higher margins than its predecessor. Management disclosed that LTSA contract margins have risen 19 percentage points since 2022, while the income statement margin, which lags because revenue is recognised over time, has risen 28 points. Since only half of contracts will be on new terms by 2028, each year that passes lifts the average margin of the entire book as more contracts flip to the new economics.

The third element is fleet growth. The installed widebody engine base is expanding at 6 to 7 percent annually. Each new engine arrives with a new LTSA, contracted at the higher margin. The Trent 700, the largest mature fleet, still has more than 40 percent of its total lifetime flying hours ahead, with only 6 percent retired in three years. The A330 freighter conversion market, where Rolls-Royce captures roughly 90 percent of conversions, extends the revenue tail of a thirty-year-old engine programme further still.

The fourth element is cost efficiency. The total cash cost to gross margin ratio has fallen from 0.80x in 2022 to 0.36x in 2025, best-in-class in the industry. This is the quiet compounder: the fixed cost base spreads across more engines, more flying hours, and more aftermarket revenue with each passing year.

The engine is the physical infrastructure. The LTSA is the subscription. Time-on-wing is the churn reduction programme. Together they produce economics that resemble a contracted annuity more than a manufacturing business, except the contract duration is measured in decades, and the annuity appreciates rather than depreciates with each passing year.

There is a tension worth naming. Management says the contribution from contractual improvements, the onerous releases and catch-ups that added £392 million in FY2025, will be reduced by 2028. If margins compress as that tailwind fades, the flywheel thesis weakens from “permanent re-basing” to “one-time repricing with partial reversion.” If margins hold above 20 percent without that support, the thesis is confirmed. The proof window is 2026 to 2027.

The Second Business and the Category Question

There is a quieter development that most coverage treats as a footnote but that may matter as much as the LTSA disclosure over a three-year horizon.

Rolls-Royce is no longer one business. Civil Aerospace generated 52 percent of group operating profit in FY2025 and absorbs 90 percent of investor attention. But Power Systems, backup power generators for data centres, engines for NATO land and naval platforms, battery energy storage, delivered £852 million of operating profit, up 60 percent, on 17.4 percent margins. Defence contributed £689 million on a £17.4 billion order backlog representing more than three years of revenue. Together, these two divisions now account for 45 percent of group earnings, and both are growing.

The temptation is to frame Power Systems as an “AI data centre play.” Data centre revenue did grow 35 percent, and management targets 20 percent annual growth in power generation to 2028. But wrapping diesel and gas generators in hyperscaler language is partly narrative engineering. Data centres are 5 to 7 percent of group revenue. That does not transform the category alone.

The more important insight is structural. Rolls-Royce’s pricing power formula, captive customers operating mission-critical infrastructure with high switching costs and long replacement cycles, is not unique to civil aviation. Data centres cannot easily swap backup power providers mid-build. NATO armies cannot switch Leopard 2 engine suppliers mid-procurement. The commercial optimisation playbook that produced 20.5 percent margins in Civil Aerospace is now producing 17.4 percent margins in Power Systems, up from 8 percent in 2022. Same mechanism, different end market.

If Power Systems sustains 18 to 20 percent margins and Defence maintains its £17 billion-plus backlog, the equity story broadens from “civil aerospace aftermarket” to “mission-critical installed-base compounder.” That category shift, if the market accepts it, supports a structurally higher multiple than the aerospace sector alone would warrant.

The Variant Perception and the Iran Dislocation

The consensus view of Rolls-Royce is now favourable. Thirteen of eighteen analysts rate the stock a Buy. The average price target is approximately 1,400p. The transformation is respected.

But the consensus still frames Rolls-Royce as a cyclical industrial trading at a full multiple. The trailing P/E of 17x looks optically cheap because it includes non-recurring statutory gains; the forward P/E of 33x looks expensive because the market values each year of earnings as if it might be the peak. The sell-side worries about margin normalisation as contractual catch-ups fade. It discounts each quarter of EFH data as if the current flying hour rate is the permanent one.

Our variant view is different. The correct valuation lens is forward price-to-free-cash-flow, which stands at 28x, and the PEG ratio, which is 0.8. An FCF margin of 18.6 percent, best-in-class among global aerospace peers, applied to a business growing earnings at 17 percent with net cash on the balance sheet and a massive buyback compressing the denominator does not screen as expensive on any quality-compounder framework. It screens as cheap.

The Iran war made the disconnect vivid. A temporary grounding of Middle East flights, affecting 15 to 20 percent of widebody traffic for a period of months, triggered a 19 percent stock decline. The market priced each lost flying hour as if it were permanent. It ignored the fact that COVID collapsed engine flying hours by 57 percent and the LTSA contracts survived intact. It ignored the 75 percent of contracted future cash that sits beyond 2028. It ignored that Defence and Power Systems, 45 percent of earnings, are either neutral or positively exposed to geopolitical instability.

The business got materially better. The price got materially cheaper. That combination has a name: dislocation.

Three Years Out

The three scenarios below are three different answers to one strategic question: is Rolls-Royce a turnaround ending, or an annuity beginning to compound?

The probability-weighted value is approximately 1,340p, implying 16 percent upside and an annualised return of 9 to 10 percent including dividends. The buyback mechanically retires 600 to 700 million shares over the period, lifting earnings per share by 3 to 4 percent annually regardless of which scenario unfolds.

What to Watch, and What We Would Do

Five things determine which scenario wins.

First, Civil Aerospace operating margin in H1 2026. Above 22 percent confirms the flywheel is self-sustaining without contractual catch-ups. Below 18 percent means the catch-up benefit was the margin story, not structural re-pricing. This is the single most important data point.

Second, FY2026 engine flying hours relative to 2019. Above 115 percent means the Iran conflict’s impact was contained. Below 108 percent implies a demand impairment that delays the mid-term earnings trajectory.

Third, Power Systems book-to-bill. Above 1.1x confirms structural demand from data centres and defence. Below 0.9x suggests the growth was a cyclical pull-forward, and the 18 to 20 percent margin target is at risk.

Fourth, shop visit trajectory. Management says visits peak in 2026 then fall. If they remain stubbornly above 1,500 into 2027, the time-on-wing engineering is lagging the timeline, and the flywheel’s cost leg weakens.

Fifth, Airbus engine selection signals. The existential long-term risk is not recession, Iran, or supply chains. It is Airbus opening a future widebody programme to competitive engine selection, breaking the exclusivity that underpins the entire annuity. Any signal of dual-sourcing, at Farnborough, in Airbus capital markets days, or in airline lobbying, demands immediate reassessment of terminal value.

In January we argued that Rolls-Royce had become a toll booth. The FY2025 results proved something we did not fully appreciate: the toll booth compounds. Each year, the installed base is worth more than the year before, through longer engine life, higher contract margins, fleet growth, and cost efficiency that reinforces itself at scale. Management disclosed this for the first time in a form that makes the economics measurable rather than theoretical.

At 1,157p, you are paying 28 times forward free cash flow for the highest FCF margin in aerospace, a compounding installed base with three-quarters of its value still ahead, a £2.5 billion annual buyback shrinking the denominator, and the cleanest balance sheet in the company’s 120-year history. The January piece said the opportunity lay in patience. The patience paid off. The flywheel is visible. The price has come to you.

Subscribe now

• Meituan’s 4Q25 numbers were ugly on the surface, but the underlying network held: users stayed, engagement rose, and Core Local Commerce improved sequentially even under extreme competitive pressure.

• The biggest update to the thesis is that the moat survived, but the old margin structure probably did not; the key question is now steady-state earnings power in a three-player oligopoly.

• This is no longer just a recovery story. It is a reclassification story: if Meituan is really the operating system for local demand, the market may still be valuing it too narrowly as a food delivery business.

From Caixin Global, February 14, 2026:

“Meituan expects a 2025 net loss of up to 24.3 billion yuan, reversing a 35.8 billion yuan profit in 2024, mainly due to fierce subsidy wars with Alibaba and JD.com.”

Then, on March 26, Meituan delivered Q4 2025 results that were simultaneously ugly and encouraging, a combination only possible when expectations have been beaten into rubble by twelve months of corporate bloodletting. Revenue grew 4.1% to RMB 92.1 billion. The net loss was RMB 15.1 billion. And the Core Local Commerce segment, the business that actually matters, posted a loss of RMB 10.0 billion, down from Q3’s record RMB 14.1 billion.

My first thought was not about the loss. It was about what survived inside it.

In September 2025, we argued that the market fundamentally misunderstood the subsidy war. We framed Meituan as the “digital bellhop”, a company whose value lay not in delivering food but in making urban complexity navigable. We predicted that Alibaba could not sustain RMB 10 billion quarterly burn rates while simultaneously funding an AI infrastructure arms race, and that Meituan’s RMB 170 billion cash pile and superior unit economics would let it simply outlast the siege. We said math was permanent; subsidies were temporary.

That was the thesis. Here is the scorecard.

What we said, what happened

Our September 2025 view rested on three pillars: the moat would hold, competitors would exhaust themselves, and Meituan’s monetization rails, membership, advertising, satellite kitchens, instant retail, would create earnings power beyond delivery commissions.

The moat held. That much is unambiguous. In a year when two of the richest companies in China threw over RMB 100 billion at Meituan’s users, platform GTV and transaction volume still grew double digits. Daily active users, monthly transacting users, transaction frequency, and ARPU all hit all-time highs. The membership program expanded, with high-value members growing steadily and purchasing across more categories. These are not the metrics of a platform losing relevance. They are the metrics of a platform under pricing stress.

The exhaustion thesis is playing out, but more slowly than we anticipated. Alibaba’s free cash flow turned negative for the first time in its public history. JD.com scaled back delivery spending in Q3 2025. The State Administration for Market Regulation intervened, summoning all three platforms in May and July, launching a formal antitrust investigation in January 2026, and reposting state media commentary titled “The Takeout War Should End” the day before Meituan’s earnings. The regulatory trajectory is real. But the war mutated rather than ended cleanly, and that is where our thesis needs updating.

Where we were too optimistic was on the shape of the post-war market. We implied that margins would largely snap back once the math forced competitors to retreat. The Q4 data tells a different story. Core LC revenue fell 1.1% year-over-year despite healthy order and GTV growth, because incentives compressed the revenue Meituan could recognize per transaction. Selling and marketing expenses consumed 34.4% of Q4 revenue, nearly double the pre-war 18-19%. Full-year S&M spending exploded 61% to RMB 102.9 billion, eating nearly all of the gross profit.

The updated thesis in one sentence: the moat survived, but the old margin structure probably did not. Meituan remains the strongest local commerce platform in China. It will not return to being the monopolist it was in 2024.

The users stayed. The money didn’t.

The market’s debate about Meituan is still framed as “when does the subsidy war end?” I think that is the wrong question. The market is waiting for the war to end. It should be asking what the peace treaty looks like.

The better question, the one that actually determines what the stock is worth, is: what is the steady-state earning power of Core Local Commerce in a three-player oligopoly?

I have stared at this table for a long time, and three things stand out.

First, the margin trajectory. Core LC operating margin improved from -21.0% to -15.5% sequentially, a 29% reduction in losses and the single largest quarter-over-quarter improvement since the war began. This was driven by something specific: management said they “pulled back resources from low-AOV, low-quality orders” while S&M expenses declined from RMB 33.8 billion to RMB 31.7 billion. Meituan is voluntarily ceding the lowest-value segment of the market. That is not a company retreating. It is a company optimizing.

Second, the revenue anomaly. Core LC revenue was down 1.1% despite management saying order volume and GTV grew healthily. That arithmetic only works if the value of each transaction declined, which is exactly what happened. Commission revenue held better (+3.0%) because it is volume-driven; delivery services revenue fell 9.9% because it is AOV-driven. The subsidy war did not shrink Meituan’s network. It shrunk what Meituan could charge for running it.

Third, and this is the detail the market has not processed, management highlighted that delivery services revenue increased sequentially from Q3 to Q4. That means the AOV trough has passed. When delivery revenue is growing quarter-over-quarter while the company is simultaneously cutting subsidies, the per-order economic inflection has likely occurred. CEO Wang Xing confirmed this on the call: “We are on track to see a more meaningful sequential improvement in our food delivery per-order loss in Q1 versus Q4 last year.”

Now, the competitive update. The food delivery front against Alibaba and JD, the war everyone obsessed over, appears to be winding down. SAMR’s escalating intervention provides political cover for all players to de-escalate, and Wang Xing leaned into it with unmistakable intent: “We are firmly against involution.” In China, “involution” is not a business term, it is a political critique of destructive zero-sum competition. By adopting the government’s own vocabulary, Wang Xing is weaponizing Beijing’s policy to force a truce that Meituan could not achieve unilaterally.

But as the food delivery front cools, a different war is heating up. CFO Shaohui Chen, responding to a question about Douyin’s escalating in-store subsidies, said something remarkable in its directness: “Yes, we see the competitors’ recent ramp up in investment. This may negatively impact our short-term profitability.” That is the CFO of a RMB 550 billion company telling you, on a recorded earnings call, that his highest-margin business is under attack and it will hurt.

The operating system thesis

Here is where I step back from the quarterly noise and say what I think is actually happening.

The market classifies Meituan as a food delivery company with temporarily bad margins. I think this is a classification error. I think Meituan is the operating system for local commercial life in urban China, the default interface through which consumers discover, evaluate, transact with, and receive services from the physical world around them. Delivery is the backbone, but it is not the product. The product is the relationship: the membership that ties you to the platform, the reviews that make you trust its recommendations, the AI assistant that knows what you need, and the courier network that makes everything available in 30 minutes.

The Q4 evidence supports this reading in three ways.

Membership. Management said high-value members are growing steadily, with transaction frequency and spending “rising notably” and purchases expanding “across a broader range of categories.” This is the Costco playbook, lock in your best customers with a subscription, make them economically irrational to switch. Fifty-three million paying members ordering 2.3x more than non-members create a demand floor that is impervious to competitor subsidies. These users are not leaving when the coupons disappear.

AI. Wang Xing devoted an extraordinary portion of the call to Meituan’s AI strategy, the in-house LongCat model, the Xiao Tuan assistant embedded in the Meituan app, and his vision of evolving user interaction from keyword search to natural-language requests. His argument is specific and, I think, correct: a generic AI can answer questions about restaurants, but it cannot guarantee a table at 7pm, confirm the kitchen has capacity, dispatch a rider, and track delivery in real time. That requires Meituan’s proprietary data, the API of the physical world: merchant POI, real-time operational status, authenticated reviews, rider GPS coordinates. The AI strategy is not about chatbot magic. It is about interface control. If users ask Xiao Tuan “where should I eat tonight?” instead of scrolling Douyin, Meituan keeps control of the discovery layer.

Capital allocation. The Dingdong acquisition ($717M for mainland China grocery operations) reveals management’s strategic map. Quick commerce is the one category that is structurally Douyin-proof, you do not impulse-buy Tylenol from a video feed; you buy it on-demand when you need it. By acquiring Dingdong’s East China warehouse network, Meituan is neutralizing JD’s advantage in the segment most resistant to the algorithmic content threat.

But Keeta’s international expansion is where skepticism is warranted. New Initiatives operating loss exploded from RMB 1.3 billion in Q3 to RMB 4.6 billion in Q4, almost entirely from Keeta’s four-country expansion. The Saudi data point is encouraging, when Meituan cut subsidies significantly, order volume held, suggesting real service quality moat. But Brazil, where Meituan faces an 80% market share incumbent, received conspicuously cautious language. The domestic core has earned the right to reinvest. It has not earned the right to wander.

Whether the operating system thesis holds depends on five observable signals. Three lean bullish today. Two remain genuinely uncertain, and those two are the swing factors.

Variant perception

Consensus sees a battered delivery company recovering from a bad year. A repair story.

I see a reclassification story. The market values Meituan at roughly 1x EV/Revenue, versus DoorDash at 6x, Grab at 4x. Part of that gap is the China discount. But part is a classification error. If Meituan is “just” food delivery, the discount is deserved. If Meituan is the default operating system for local demand in China, with delivery as backbone, membership as retention engine, advertising as monetization layer, AI as interface, and fulfillment as infrastructure, then the stock is materially undervalued even after the margin reset.

The market is waiting for the war to end. It should be asking what the peace treaty looks like.

Three-year price framework

It is a framework for thinking about what the stock is worth under different assumptions about the business Meituan is becoming.

The structural insight from this table is not the specific numbers. It is the point underneath them: lower margins on a much larger base can still produce attractive absolute earnings power. But only if Meituan retains the interface and shows discipline on capital allocation. Q4 satisfied the first condition. The second remains an open question.

Five numbers to watch

My overall conviction: moderately bullish, but with less certainty than six months ago. The moat thesis is stronger, the stress test proved the network is durable. The margin thesis is weaker, the old economics are not coming back, and the Douyin front is a genuine new risk. At HK$86, the risk/reward still favors the patient investor. But you need to accept that you are buying a different company than the one that traded at HK$190, and that the path from here runs through 12-18 months of messy consolidated numbers before the operating system re-monetizes underneath.

The bellhop lost his monopoly on tips. But he learned every service in the building, signed you up for a membership, and built an AI that knows what you need before you ask. The question is no longer whether he survives. It is what he earns. At this price, I think the answer is more than the market expects, and less than the old monopoly bulls remember.

Subscribe now

• The original operating thesis held. Margins expanded, ad outgrowth persisted, and Business Services accelerated, proof that Tencent’s AI dividend is real and already compounding in the core business.

• What broke was the capital allocation assumption. Tencent is no longer avoiding the infrastructure cycle; it is choosing to reinvest heavily, betting that WeChat can become China’s execution layer for AI agents.

• The market may still be asking the wrong question. This is no longer mainly about Yuanbao versus rival chatbots, it is about whether Tencent owns the rails that agents will need for identity, service access, payments, and user communication.

Three months ago we wrote that Tencent’s edge was being profitable while rivals burned capital. The thesis was clean: AI value accrues to those who own the user relationship, not those who build the biggest datacenter. We called it the AI dividend, the 700 basis point spread between gross profit and revenue growth that proved you could extract superior AI returns without entering the infrastructure arms race.

Then Martin Lau said this on the Q4 FY25 earnings call:

“Our spending on our two biggest new AI products, HunYuan and Yuanbao, was CNY 7 billion in Q4 of 2025 and CNY 18 billion for the full year. We expect to more than double these investments in 2026.”

The stock fell 6.5% on the day. It has since declined to HKD 499.60, 23% below where we wrote the AI Dividend piece, 25% below where we wrote the Patience Premium.

Two things need to be said before any analysis proceeds.

The business validated the thesis. Gross margins expanded to 56.2% for the full year. Marketing Services grew 19% against a 14% China industry benchmark. Business Services accelerated to 22% year-on-year in Q4. The AI dividend was real, structural, and compounding exactly as described.

The capital allocation thesis was wrong. We argued Tencent would extract AI value without entering the infrastructure cycle. They entered the infrastructure cycle. The question is whether that changes the investment case or deepens it.

Our answer: it deepens it. But only if you’re willing to update the framework honestly. Which is what this piece attempts.

What Held, The Dividend Is Real

Start with what the data actually shows, because the narrative around Q4 FY25 obscured something important: the operating results were excellent.

The “Patience Premium” piece argued Tencent had built monetisation levers it was deliberately not pulling. Video Accounts at 3–4% ad load versus Douyin’s 10–12%. Weixin Search below 10% of advertising revenue. Mini Shops take rates near zero. The argument was that these levers existed, were proven by competitors, and were controlled by management with a decade-long track record of patient execution.

Every one of them is still turning:

The AI Dividend piece called the 700 basis point spread between gross profit and revenue growth the financial signature of AI-powered monetisation working. FY2025 delivered 21% gross profit growth against 14% revenue growth. The spread held. The harvest the Patience Premium described, levers proven by competitors, controlled by management, available when needed, is proceeding exactly as argued.

What changed is not the harvest. What changed is that Tencent decided not to wait for it.

What Changed, The Dividend Is Being Reinvested, Not Collected

Here is the sentence from the AI Dividend piece that requires the most honest revisiting:

“President Martin Lau made the strategic choice explicit: ‘We actually believe that there’s no insufficiency of GPUs for us at this moment for our internal use.’”

We cited that quote as evidence of strategic discipline, Tencent pointing AI investment inward at advertising and content rather than outward at low-margin infrastructure. Three months later, the same management team committed to more than doubling infrastructure investment. Both statements were true. What changed wasn’t the compute availability. It was the opportunity.

Between Q3 FY25 and Q4 FY25, OpenClaw emerged as a commercially viable agentic AI paradigm. The GPU sufficiency statement reflected Phase One reality, deploying AI within existing businesses to raise yield. The RMB 36 billion commitment reflects Phase Two ambition, building the execution layer for AI agents operating inside WeChat on behalf of 1.42 billion users.

These are not the same strategic decision. Phase One was internally directed, margin-accretive, and required no additional infrastructure beyond what Tencent already owned. Phase Two is externally contested, requires building capability that doesn’t yet exist at commercial scale, and demands capital investment before the return profile is visible.

Lau framed the distinction explicitly:

“We are breaking out our investment in new AI products because we view these strategic investments conceptually similar to investment in an affiliate or to CapEx. These are upfront investments required to build the necessary foundation to unlock new value as opposed to ongoing operating expenses.”

This is management asking investors to hold two equations simultaneously: Phase One returns already flowing through the P&L, and Phase Two costs front-loaded in 2026 with returns arriving later. The market’s 6.5% reaction was not irrational. When a company reduces mechanical EPS accretion while increasing spend on unmonetised products, the short-term earnings trajectory gets worse before it gets better. What the market got wrong was treating the investment as evidence of strategic weakness rather than strategic expansion.

The prior articles asked: how does AI make Tencent’s existing business more valuable? The correct question after Q4 FY25 is different: what is Tencent worth if WeChat becomes the execution layer for AI agents in China?

Those are not the same question. The first has a bounded answer, better margins, higher eCPM, more recurring gaming revenue. The second has a potentially unbounded answer. And the answer to the second question begins with this:

“We’re building AI agents which autonomously interact on behalf of users within Weixin functionalities, especially Mini Programs.”, Martin Lau

Read that sentence slowly. Not a chatbot. Not a search product. Agents that act, inside the surface where 1.42 billion people already conduct their digital lives, paying bills, messaging friends, discovering merchants, booking services. The agent doesn’t need to win a benchmark. It needs to be present when the user delegates a task. And the user’s tasks are already happening inside WeChat.

The product is not Yuanbao. The product is Weixin with AI in it.

What the Market Is Getting Wrong

The consensus view on Q4 FY25 can be summarised in one sentence: quality compounder, temporary AI investment headwind, buy on weakness, 12–18 month payoff as margins recover and buybacks resume.

This is probably right about the near-term stock trajectory. It is asking the wrong strategic question.

Lau was direct about why:

“We actually believe that AI chatbot applications are largely competing with search applications rather than with every other application.”

The market is valuing Tencent’s AI position by measuring Yuanbao’s user numbers against Doubao and Qwen. Lau is saying that’s not the contest. The chatbot category is a search substitute. The agent category is something different, and the agent category is where WeChat’s structural advantages are overwhelming rather than merely competitive.

To understand why, consider what an AI agent actually needs to complete a real-world task. It needs to verify who it’s acting for. It needs to access the services the user wants. It needs to process payment. It needs to report back through a channel the user trusts and monitors. Four requirements. Every commercially valuable agent task needs all four simultaneously.

WeChat owns all four in a single surface. The social graph provides identity. 1.3 million Mini Programs provide service access. WeChat Pay processes the transaction. The messaging layer closes the loop. No other entity in China owns all four. Alibaba has payments and merchant access but not communication. ByteDance has attention but not trusted identity or full merchant integration.

This is the tollbooth. When an agent books a restaurant, whether built by Tencent, ByteDance, or a startup, it needs to authenticate the user, access the restaurant’s booking system, and process the payment. Three of those four requirements route through WeChat infrastructure. Tencent doesn’t need to build the best agent. It needs to own the rails the agents run on.

The Patience Premium piece made an analogy that becomes more precise here. In 2014, Alibaba counted red envelopes. Tencent counted relationships. Alibaba optimised for the transaction volume of a single campaign. Tencent embedded payments into the social fabric of the holiday itself. A decade later, WeChat Pay processes more daily transactions than Alipay managed in that entire campaign.

The same strategic moment is occurring one layer up. Competitors are counting chatbot DAUs. Tencent is embedding agents into the surface where Chinese digital life already happens. The question is not who wins the model race. It is who owns the execution layer when the model race is over.

And then there is this, the most telling signal in the entire earnings package, reported almost as an aside by Bloomberg:

The normally publicity-shy Ma used his own WeChat feed to promote OpenClaw-inspired agentic tools.

Founders don’t go public about products they’re uncertain about. Pony Ma spent three years in near-total public silence through the regulatory crisis, the gaming licence freeze, the revenue decline. His re-emergence to personally promote agentic tools through his own social feed is not a marketing decision. It is a founder communicating conviction about an architectural moment. That conviction doesn’t appear in Bloomberg consensus estimates. It should.

Variant Perception and Three-Year Scenarios

The Patience Premium framed Tencent as an option on Chinese consumer attention priced like the options had expired. At HKD 662, we argued the market was offering utility multiples for a company with social network monetisation potential.

The updated framing is both more precise and more ambitious. Tencent is an option on the execution layer of China’s agent economy, currently priced like an attention company in moderate decline.

At HKD 499.60, against a Bloomberg consensus price target range of HKD 650–787 and a five-year average forward P/E of approximately 22x, the 14.23x FY2026E multiple embeds a specific worldview: the AI investment cycle is a drag without a visible return path, the chatbot products are losing to better-funded competitors, and the buyback reduction removes the mechanical EPS support that justified premium pricing.

Each of those concerns is legitimate. None of them addresses the tollbooth.

The prior articles’ price targets require honest revision given the capital allocation shift. The Patience Premium’s base case of HKD 859 is still achievable but the timeline extends, the Phase Two investment cycle pushes it toward FY2028 rather than FY2027. The AI Dividend’s 18-month target of HKD 1,100–1,150 was too aggressive on timing; that outcome, if it occurs, is a three-year view.

There is also one new bear case risk that didn’t exist when we wrote either prior piece: the OS-layer bypass. If a hardware manufacturer, Huawei, Apple, or a device maker not yet prominent, establishes an OS-level agent interface before WeChat agents reach habit-forming scale, the tollbooth is bypassed at the instruction layer. Users instruct the OS, which accesses apps directly without routing through WeChat’s Mini Programs ecosystem. The four-layer execution stack that makes WeChat’s position so strong becomes irrelevant if the agent never enters WeChat in the first place. This risk gets 25% probability, higher than anything in prior articles, because the capital allocation shift means Tencent is now racing against a timeline it didn’t choose.

One specific input worth noting: Lau made the clearest possible statement about why the subscription model, the primary AI monetisation path in the US, is unavailable in China:

“Unlike in the U.S. where you can actually get consumers to pay subscriptions... In China, those are not sort of that available.”

This is not a weakness in the Tencent thesis. It is the answer to why the execution layer thesis is the only viable AI business model at scale in China, and why Tencent’s specific asset combination is worth more than a 14x multiple implies. If subscriptions are off the table, the only path to AI monetisation is transaction economics. Transaction economics require the rails. Tencent owns the rails.

The Updated Dashboard

The Patience Premium tracked five metrics. Six now, one addition that didn’t exist as a concept when we wrote the prior pieces.

Gross margin trajectory. Above 57% = Phase One still compounding, mix shift structural. Below 55% = competitive pressure winning. Bloomberg consensus at 58.3% by FY2028 is the bull embedded in the fundamental model that the sentiment narrative ignores.

Marketing Services versus China ad industry. Above 400 basis points outperformance = AI yield enhancement intact. Below 200 basis points = the eCPM advantage is narrowing and the AI dividend thesis requires revision.

Business Services growth. Above 20% sustained = execution layer monetisation beginning to show in reported numbers. Below 15% = the Phase Two ramp is slower than the Tencent Cloud analogy implies, and the timeline for returns extends further.

Mini Shops disclosure. Any GMV number with a stated take rate above 1% = the harvest has begun. Continued silence = still planting. The Patience Premium’s best observation holds here: the silence is data. When management starts disclosing the number, the monetisation has started.

Management language on agents. Specific task completion metrics or agent usage data = product-market fit emerging from theory into practice. Continued product description without metrics = still in the discovery phase, which is not necessarily bad but extends the timeline for the scenario table above.

New, any metric Tencent introduces that didn’t exist in prior quarters. This is the most reliable leading indicator in the entire dashboard. Companies do not invent new disclosures for businesses that are struggling. The quarter management introduces agent task volume, closed-loop transaction share, or Mini Shops take rate is the quarter the transition from thesis to evidence becomes legible.

The Patience Premium ended with this: Some companies monetise because they must. Tencent monetises when they choose.

That framing still holds, but Q4 FY25 added a layer. Tencent chose to stop collecting the dividend and reinvest it instead. The question the next four quarters answer is whether that choice was disciplined capital allocation or a costly detour.

The evidence so far suggests the former. The business is stronger than the stock price implies. The structural position is more defensible than the chatbot narrative acknowledges. And a founder who spent three years publicly silent just promoted his company’s agentic tools on his personal feed.

That’s not nothing.

Subscribe now

• Alibaba is uniquely positioned to turn AI into execution, not just recommendation

• But the legacy engine (ads + commissions) is stalling despite strong user growth

• The outcome depends on a narrow bridge: cloud scaling into a profit engine before commerce fades

In 2003, Taobao had a problem that had nothing to do with technology. Chinese consumers would not send money to strangers on the internet, and sellers would not ship goods before receiving payment. Every transaction was a standoff. The solution was not a better search algorithm or a wider product catalogue, it was Alipay, which held the buyer’s money in escrow until the package arrived. What made this work was not the software. It was the recognition that in a market where nobody trusted anybody, the company that inserted itself into the middle of the transaction would not merely participate in commerce. It would become the infrastructure that commerce ran on. By 2013, Taobao had 70% of the market. The lesson was not that trust mattered. The lesson was that solving the trust problem made you the market.

What we argued in January

Two months ago, we wrote that Alibaba was facing a second version of that problem, extending trust from human-to-human payments to human-to-software execution. We framed three systems being built simultaneously: a compute spine in cloud infrastructure, a frequency layer in quick commerce, and a monetization engine in advertising and merchant tools. We said everything hinged on whether the bills would get smaller. We were watching three signals: sales and marketing trending toward 23%, customer management revenue holding above 8%, and cloud sustaining above 25% growth. One came in well above expectations. The other two missed.

The quarter that split the story

Alibaba’s December quarter did not settle the debate about the company. It separated it into two debates.

Cloud Intelligence Group revenue grew 36% year-over-year, accelerating for the fourth consecutive quarter. AI-related product revenue posted triple-digit growth for the tenth quarter in a row. Token consumption on the model studio platform grew sixfold in three months. Management announced a new business group called Alibaba Token Hub, set a five-year target of $100 billion in combined cloud and AI external revenue, and disclosed that T-Head, the chip design subsidiary, had shipped 470,000 cumulative AI chips. On the AI and infrastructure side, the momentum is not incremental. It is compounding.

Then there was the other side. Customer management revenue, the advertising and commission income that has historically been the company’s profit engine, grew 1%. Adjusted EBITA fell 57%. Free cash flow was RMB 11.3 billion, down 71% from a year earlier. Sales and marketing remained at 24.9% of revenue. Management dropped the phrase “peak investment” entirely and replaced it with language about “sustained reinvestment” and a “new phase of entrepreneurial reinvention.” The quick commerce business, which grew revenue 56%, was cited as a primary driver of the profit collapse, with profitability now targeted for fiscal year 2029.

Two companies reported under one ticker. The question the rest of this piece tries to answer is which one will define Alibaba three years from now.

Why AI makes Alibaba’s old strengths new again

Most AI products today sit beside economic activity. They recommend a restaurant but do not book the table. They compare prices but do not complete the purchase. The gap between suggestion and execution is where most applications stop, because crossing it requires access to payments, identity, logistics, and the commercial surfaces where money changes hands.

Alibaba controls all of those surfaces. When Qwen was upgraded in January 2026, users could order food through Taobao Instant Commerce, book travel through Fliggy, process payments through Alipay, and navigate through Amap, all without leaving the AI interface. By the end of February, roughly 140 million users had completed a real purchase through Qwen’s agentic features.

It is worth being specific about who else in China could build this kind of system, because the answer sharpens why Alibaba’s position matters. ByteDance has massive consumer distribution and competitive AI models, but no meaningful external cloud business, Douyin commerce routes payments through third parties rather than owning the rails. Tencent has WeChat, which could become a default agent interface, and a respectable cloud business, but lacks proprietary chips and its commerce surfaces are fragmented across mini-programs rather than integrated. Huawei has chips and cloud but no consumer ecosystem at all. No single competitor currently controls the complete chain from silicon to models to cloud to the endpoints where transactions happen.

That integration matters because of something CEO Eddie Wu said on the earnings call that deserves more attention than it received. He argued that enterprises are beginning to treat token consumption not as part of their IT budget, but as an operational cost, a production input, like raw materials or energy. If that shift takes hold broadly, the implications for cloud economics are significant. IT budgets have traditionally represented around 5% of corporate revenue. They are discretionary, capped, and among the first expenses cut in a downturn. Production inputs are different. Companies consume more of them as they grow, not less. If AI tokens follow that path, from experimental to operational to simply part of how work gets done, then the addressable market for AI infrastructure is not the current cloud TAM. It is considerably larger.

This should raise a natural question: if cloud revenue is growing 36% and the addressable market is expanding, why hasn’t operating leverage appeared? The segment’s adjusted EBITA margin was 9%, essentially flat year-over-year despite accelerating revenue. Three factors explain the gap. First, the China cloud market is in a land-grab phase where Alibaba, Huawei, Tencent, and ByteDance are all competing aggressively on price to lock in enterprise AI workloads, compressing margins even as volumes surge. Second, the current revenue mix skews toward GPU-intensive AI training workloads, which carry lower margins than the inference and model-as-a-service offerings that should eventually dominate. Third, T-Head’s proprietary chips, which could structurally reduce inference costs, have shipped 470,000 units with annual revenue at the CNY 10 billion level, meaningful but not yet at the scale where they materially alter the cloud segment’s cost structure. Management hinted that cost improvements from T-Head may not arrive gradually but as a step-change when production volume reaches a sufficient threshold. That is worth watching closely, but it is a claim about the future, not a feature of the present.

And a note of caution on the Qwen figures. The 140 million transaction number came during a Chinese New Year campaign with significant promotional support, the kind of environment designed to produce impressive launch metrics. What matters more is whether users return to Qwen for purchases in March and April without the promotional nudge. Repeat rate, incremental GMV relative to traditional Taobao flows, and average order value through the AI interface are the metrics that will separate a successful product from a subsidized launch. Those numbers do not exist yet.

The part that got harder

If the previous section described the destination, this section describes the bridge. And the bridge is narrower than we assumed in January.

Start with customer management revenue. Growing 1% sounds like stagnation, but it is actually worse than that in context. Taobao’s monthly active consumers grew by double digits. The 88VIP membership program surpassed 59 million members. Annual active consumers increased by 150 million in calendar 2025. All of those engagement metrics improved, and yet the revenue Alibaba earns from advertising and commissions on those users barely moved. The explanation is compositional: the new users arriving through quick commerce carry structurally lower spending levels and purchase frequency than existing users. Management acknowledged this on the call. The software service fee that had provided a tailwind to CMR in prior quarters is now fully lapped. Management said the March quarter is showing “significant recovery” in both physical goods GMV and CMR trends. That forward guidance is the single most important near-term test of the thesis. But guidance is a claim, not evidence.

Then there is quick commerce itself. Management gave the most specific financial targets yet: over RMB 1 trillion in GMV by fiscal year 2028, positive cash flow at that scale, and profitability in fiscal year 2029. That timeline is more concrete than what we had before, and it is also later than our base case assumed. Quick commerce may well be strategically necessary, if Meituan owns the daily “I need this now” intent and Alibaba does not contest it, the entire ecosystem gradually loses relevance in consumers’ daily routines. But strategic necessity does not make the economics pleasant. The business is still consuming cash, and it will continue to do so for years.

These two problems connect through a specific equation. Alibaba’s transformation depends on commerce cash flow holding long enough for cloud to reach the scale and margin structure where it becomes self-funding. If customer management revenue stabilizes in the 3-5% growth range and cloud margins gradually improve toward the mid-teens as T-Head scales and the revenue mix shifts toward higher-margin model-as-a-service offerings, the math works. Commerce funds the bridge; cloud becomes the destination. If customer management revenue stays near zero growth and cloud margins remain flat, the bridge is too narrow to support the crossing. That rate equation, how fast commerce decays relative to how fast cloud margins improve, is the single most important variable for the stock over the next two years.

What we think now

We are more confident than we were in January that the AI and cloud demand story is real and durable. Qwen’s distribution inside the ecosystem, the cloud acceleration, and the early evidence of agentic transaction execution are stronger signals than we had two months ago, with the caveat that the execution evidence is still early and promotional in nature.

We are less confident that commerce will recover quickly enough to make the transition financially comfortable. The 1% CMR print, the elevated marketing spend, and the pushed-out quick commerce timeline all suggest the funding side of the equation is weaker than we modeled.

Where we differ from the prevailing market view is on what matters more: the level or the direction. The market is looking at a commerce company with 1% CMR growth and a 9% cloud margin and seeing a structurally impaired conglomerate. We are looking at a cloud business that has accelerated from 7% to 36% in four quarters and asking what this company’s earnings mix looks like when cloud is 25% of revenue instead of 15%. If we are wrong, it will be because commerce decayed faster than cloud could compensate, and the March quarter CMR print is the first real test of that.

Three years from now

We model three scenarios for roughly fiscal year 2029.

The bear case values the stock at roughly HK$95, or around US$85. Revenue grows 5-6% annually, CMR stagnates near 0-2%, cloud decelerates to the mid-teens, and cloud margins remain around 10-11%. Quick commerce misses the FY29 profitability target. EBITDA margins stay at 12-13%. The market applies 7-8x. We assign this 20% probability.

The base case reaches roughly HK$190, or around US$170. Revenue grows 9-10% annually. CMR recovers to 4-6% as new user cohorts mature and AI-powered merchant tools improve returns. Cloud sustains 22-25% growth with margins expanding to 14-16% as T-Head scales and MaaS becomes a larger share. Quick commerce reaches breakeven by FY28. Sales and marketing normalizes to 20-21%. EBITDA margins recover to 17-18%. The market applies 10-12x. We assign this 55% probability.

The bull case reaches roughly HK$300, or around US$260. Revenue grows 12-14% annually. The token-as-production-input thesis materializes broadly. Cloud sustains 28-30% growth with margins reaching 18-20% as T-Head surpasses one million chips and inference cost advantages compound. CMR re-accelerates to 8-12%. EBITDA margins reach 20-22%. The market applies 14-16x. We assign this 25% probability.

What we are tracking

In the next 90 days, we are watching two things. Customer management revenue growth in the March quarter: above 5% confirms the December trough was seasonal; below 3% suggests the impairment is structural. And sales and marketing as a percentage of revenue: a move below 23% would be the first real signal that spending is normalizing.

Over the next six months, the signposts shift to cloud profitability and capital allocation. Cloud EBITA margin crossing above 11% in any quarter would indicate that operating leverage is beginning to emerge. A return to share buybacks above $1 billion per quarter would signal that management sees the free cash flow turn approaching. And the first profitable quarter from the international commerce segment would demonstrate that Alibaba can actually close an investment cycle.

Over the next twelve to eighteen months, the variables that matter most are the ones the market is not yet tracking. T-Head chip shipments and revenue will determine whether the non-linear margin improvement management hinted at is real. Model-as-a-service as a percentage of cloud revenue will validate or undermine the production-input thesis. And quick commerce unit economics at the city level will provide the first hard evidence of whether the frequency layer can be more than a necessary cost.

We would change our view if we saw two consecutive quarters of cloud growth below 20% with flat margins, customer management revenue declining year-over-year, or management language shifting from “sustained investment” to “accelerated investment”, which would indicate the spending cycle is extending again rather than resolving.

From escrow to execution

In 2003, Alibaba helped strangers trust each other enough to complete a transaction. In 2026, it is trying to help people trust software enough to act on their behalf. That is not a small shift. It changes what the company is, what it is worth, and how it should be evaluated.

After the December quarter, the destination looks more valuable than it did in January. Cloud is compounding. Qwen is executing real transactions at scale, though how durable that behaviour proves without promotional support remains to be seen. The full-stack integration from chips to models to consumer surfaces is, for now, intact and widening. But the bridge from here to there is narrower than we assumed, commerce is weaker, quick commerce is later, and the spending cycle is longer. Whether the bridge holds is not a question of strategy. It is a question of arithmetic, and the next two quarters will provide the numbers that answer it.

Subscribe now

• Q4 proved the shift is real: identity recovery is exploding (400 → 900 customers) and security now drives nearly half of expansion revenue.

• Rubrik’s moat is architectural: it maintains a time-indexed map of enterprise state, something detection vendors and hyperscalers cannot easily replicate.

• The opportunity is a category error: the market values Rubrik like storage while the business is compounding like security infrastructure.

In our last piece, we called Rubrik the janitor who became the bodyguard. The janitor spent a decade storing copies of enterprise data as insurance, unremarkable work, mostly invisible, occasionally essential. The bodyguard, we argued, was emerging: the same immutable data archive that powered backup was becoming the only mathematical way to restore a business to the exact second before a breach. The mechanism was right. What Q4 revealed is that we were describing a transition that was already further along than we knew, and asking a question that Q4 has now largely answered, which means the interesting question has moved.

The market, meanwhile, is arguing about a different question entirely. Rubrik reported its strongest operational quarter since going public, record net new ARR, first positive non-GAAP EPS, $238 million in full-year free cash flow, and the stock sits at $54, 44% below its peak. The explanation is not complicated: reported revenue growth will decelerate from 48% to roughly 21% in FY27, and algorithmic screens do not read footnotes. But the explanation, while accurate, is not the point. The point is what the dislocation reveals about how the market is categorizing this company, and why that categorization is wrong in a way that matters for the next three years.

The Wrong Category

Backup is about storing copies. That is what Rubrik was built to do and what most investors still think it does. The market’s mental model has updated to “backup company with a security story,” which is better than “backup company” but still wrong in a structural sense.

What Rubrik actually maintains is an enterprise’s relationship with its own past. Every file, every directory, every identity state, every configuration, indexed, immutably stored, continuously updated, queryable across time. This is not semantically equivalent to backup. Backup answers “do I have a copy?” Rubrik’s system answers “what was the enterprise, precisely, at any moment in its history?” The distinction sounds philosophical until you understand what cyber attacks actually target.

The Scattered Spider attacks that have dominated enterprise security discussions over the past two years did not encrypt files. They compromised identities, corrupted Active Directory, and locked administrators out of the systems they needed to respond. CrowdStrike can tell you this happened. It can identify the attacker’s methodology and map the intrusion. What it cannot do is reconstruct the Active Directory as it existed before the attack, because endpoint telemetry does not store the historical state of the enterprise. Rubrik does. This is not a feature gap CrowdStrike can close by hiring more engineers. It is an architectural reality: you cannot reconstruct what you did not index.

The shift from perimeter defense and endpoint detection to identity-based attacks and assumed breach means the scarce capability in enterprise security is no longer detecting what happened, it is knowing what the enterprise used to be and being able to get back there. That is a different product category than backup, and it is a different product category than traditional cybersecurity. It is, for lack of an established term, state integrity. And Rubrik is the only company with the architecture to own it at enterprise scale.

Why Owning State Compounds

The reason this matters beyond a single product category is that Rubrik’s value grows with the completeness of its map. Every additional workload, identity system, SaaS application, and cloud environment that Rubrik indexes makes its map more accurate, its recovery more precise, and its governance more relevant. This is a network effect, not across customers, but within them. Each enterprise Rubrik serves becomes a progressively more locked-in customer not because switching costs are high in the traditional sense, but because the value of the map grows with its coverage.

This is the structural dynamic that the “backup company” frame misses entirely. Point products get commoditized as complexity rises, the more fragmented the enterprise environment, the harder it is for any single security tool to maintain relevance. System-of-record products do the opposite. The more fragmented the environment, the more valuable the system that maintains a coherent picture of what that environment is supposed to look like. Rubrik gets stronger as enterprises get more complex. That is not true of most software companies, and it is certainly not true of backup vendors.

The zero-marginal-cost cross-sell is the financial expression of this dynamic. Once Rubrik has indexed an enterprise’s data estate, which happens in the course of selling backup, deploying an identity resilience module or a data security posture product requires no new infrastructure, no new implementation project, no new security stack integration. The telemetry already exists. The customer acquisition cost for Act II is, in the limit, zero. This is why $238 million in free cash flow is not an aberration. It is what the Trojan Horse looks like when it opens.

Q4 as Proof, Not Update

We argued in prior pieces that identity and security products were the plausible next adjacency. Q4 closed that debate.

Identity customers grew from 400 to more than 900 in a single quarter, the fastest product growth in company history. More importantly, 40% of those new identity customers were net-new logos: enterprises that came to Rubrik specifically for identity recovery, not as a cross-sell from backup. That means the CISO buying motion is real and it is self-sustaining. Rubrik is winning in security evaluations against security vendors, on security criteria, in the security budget. Security products now account for more than 45% of net revenue retention expansion, up from 34% a year ago and from essentially zero three years prior.

These are not incremental data points. They are evidence that the compounding mechanism has activated beyond backup. The map is extending into identity. The zero-CAC cross-sell is working. The buyer has shifted. When we said the janitor was becoming the bodyguard, we were describing a hypothesis. Q4 makes it a fact.

The deferred revenue balance makes the durability concrete: $1.845 billion in contracted, already-signed revenue against $1.316 billion in trailing twelve-month revenue. The pipeline covers more than a year of recognized revenue with zero new bookings. Rubrik’s contracted future is larger than its recent past. That is not a statistic you associate with a category in decline.

The FCF as Fingerprint

There is a precise financial signature for a company whose second act is monetizing infrastructure it built in its first act at near-zero marginal cost. Revenue grows at a meaningful pace. Free cash flow grows much faster. Margins expand structurally rather than cyclically. The GAAP-to-non-GAAP reconciliation narrows as stock-based compensation normalizes.

Rubrik’s FY26 financials are this fingerprint. Full-year free cash flow of $238 million on $1.316 billion in revenue, an 18% FCF margin, against $22 million and 2% in FY25. ARR contribution margin swung from negative 12% in FY24 to positive 12% in FY26. Non-GAAP EPS turned positive in Q4 for the first time. These are not the numbers of a backup vendor straining to grow. They are the numbers of a platform in the early innings of a compounding expansion loop.

The one caveat worth stating clearly: FY27 free cash flow guides to only $265-275 million, or roughly 14% growth despite the structural leverage. Management is investing the incremental dollar rather than dropping it to cash, capex tripled in Q4. That is not a red flag, but it warrants watching. The question is whether the investment is building capacity for the next S-curve or filling gaps in the existing one.

AI Is Not the Thesis. It Is the Destination of the Same Mechanism.

Rubrik Agent Cloud reached general availability in Q4 with integrations into Amazon Bedrock and Microsoft Copilot Studio. No paying customers were disclosed. The AI chapter is not open yet.

But understanding why RAC exists matters for understanding what Rubrik is becoming. AI agents access data, modify directories, and make decisions inside production environments with trusted credentials. When one behaves incorrectly, through a bug, a compromised prompt, a misconfigured policy, the damage is not a file. It is a workflow. It might propagate across forty downstream systems before anyone notices. The enterprise needs to know what state it was in before the agent acted, and it needs to be able to get back there.

This is the same problem Rubrik already solves for human attackers who compromise identities. The mechanism is identical: maintain a time-aware map of enterprise state, detect deviation, restore clean state. The only difference is the actor, human versus agent. RAC is not a new thesis bolted onto Rubrik’s platform. It is the existing mechanism applied to the next domain of enterprise risk.

That is why AI is the destination rather than a separate bet. The same architecture that makes Rubrik valuable for identity recovery makes it valuable for agent governance. The map just needs to extend. And the more the enterprise deploys AI agents, the more complete and more valuable the map becomes. This is the loop extending into its next iteration.

What It Is Worth

The market is pricing RBRK at roughly 7x forward ARR as a decelerating backup vendor. The question is which misperception clears first, and when.

In the bear case, it never does. Cloud providers, AWS, Azure, GCP, build credible cross-cloud identity recovery and governance tools that enterprises accept as good enough. The CISO buying motion stalls. Rubrik remains a premium recovery company with excellent economics but a capped multiple. Revenue grows at a 16-18% CAGR through FY29, FCF margins plateau at 15-16%, and the stock trades at 5-6x ARR on a $2.0 billion ARR base. Call it $45-55. It is still a fine business. Private equity notices at this price. That creates a floor, not a recovery.

In the base case, the accounting noise clears, the CISO transition continues, and identity becomes a scaled second engine. Security products exceed 55% of NRR expansion by FY28. Revenue grows at a 22-24% CAGR. FCF margins expand toward 22-24% as the cross-sell loop matures. The multiple re-rates from 7x to 10-12x ARR as sell-side analysts are forced to reclassify the company from storage to security. On a $2.3-2.5 billion ARR base at those multiples, the stock reaches $90-110 over three years.

In the bull case, identity continues compounding and AI begins contributing. RAC generates meaningful revenue by FY28. The company gets reclassified not as security but as AI data infrastructure, the control plane for enterprise operating state in a world of agentic systems. Revenue grows at a 27-30% CAGR. FCF margins approach 27-28%. The multiple expands to 14-17x ARR on a $2.7-3.0 billion ARR base. The stock reaches $160-190.

The base and bull cases are not dependent on management executing perfectly. They are dependent on the mechanism being real, which Q4 confirmed, and on the market eventually valuing it correctly. The bear case is dependent on hyperscalers solving a problem that runs counter to their incentives: vendor-neutral, cross-cloud recovery that works as well on AWS as on Azure as on GCP. Their incentive is lock-in. Rubrik’s value proposition is the opposite.

What to Watch

Four questions will determine which scenario unfolds. Does identity reach 1,500 customers by the end of Q2 FY27, or does it plateau after the initial CISO wave? Does the June 10 Investor Day produce a disclosed RAC paying customer, or does AI remain a product without revenue? Does normalized revenue growth, adjusted for the $60 million material rights headwind, hold above 25% for two consecutive quarters, collapsing the deceleration narrative? And does security’s share of NRR expansion cross 50%, forcing a categorical reclassification?

The answers will arrive over the next four to six quarters. The optical distortion will make every quarter look like deceleration to anyone reading headlines. That gap, between what the screener sees and what the mechanism is doing, is the opportunity.

The most valuable system in the AI enterprise may turn out to be the one that knows what the enterprise used to be. Rubrik spent a decade building that system while the market thought it was building backup. Q4 was the quarter the system proved it extends beyond backup. The question now is not whether that matters. It is how large the surface area becomes.

Subscribe now

• The FCF debate is closed. €1.2B of sovereign prepayments funded Rheinmetall’s working capital and flipped the balance sheet from €1.3B net debt to €369M net cash.

• The moat upgraded. This is no longer just regulatory entrenchment, governments are pre-funding Rheinmetall’s expansion because the alternative is a capability gap.

• The business improved, the valuation didn’t. The ceiling for the thesis rose, but at €1,542 (~40× FY26e EPS) the probability-weighted return is still ~8% annually.

Rheinmetall’s shares fell nearly 8% after the company’s 2026 outlook on profit margin and free cash flow came in below elevated expectations, with the ~19% operating margin guide below analyst consensus of 19.6% and free cash flow conversion guidance of more than 40% of EBIT well below the 70-90% some investors were looking for.,

• Reuters, March 11, 2026

The Frame We Brought In

Coming into this print, the thesis rested on three pillars. First, Rheinmetall owned the only fully vertically integrated European defense production stack, from propellant through complete weapons systems, vehicles, electronics, and now naval, meaning every euro of NATO defense spending that had to be sourced in Europe flowed through a funnel where Rheinmetall was either prime contractor or mandatory subsystem supplier with no credible continental substitute. The EDIP regulatory framework and European sovereign-sourcing pressure made this entrenchment progressively harder to route around. Call it a regulatory moat with vertical integration on top.

Second, FCF was the unresolved question that could break the story. Through the first nine months of 2025, Rheinmetall had burned through €813M in operating free cash flow while posting record operating results. A company that converts 29% revenue growth and 18%+ operating margins into negative cash generation is either building something genuinely large or papering over structural weaknesses with accounting. We didn’t know which.

Third, €1,542 was full price for even a strong business. Our 5-year compounding analysis was blunt: at current prices, you needed the bull case for 15% CAGR. The high-certainty entry for that return on conservative assumptions was €700-870. We called it a strong business at full price, the kind of situation where you respect the quality and wait for the market to give you a better entry.

The central tension was this: a structurally compelling business case sitting inside a valuation that left no room for error. The question this quarter had to answer wasn’t whether demand was real. That debate was already won. The question was whether the mechanism, the actual source of Rheinmetall’s advantage, was as strong as we thought. It turned out to be stronger. And differently strong, which matters more.

What the Quarter Actually Resolved

The FCF question closes here, but the how matters more than the that.

Full-year 2025 operating free cash flow came in at €1.218B. The net financial position, net debt of -€1.292B at end-2024, flipped to +€369M net cash. That’s a €1.66B swing in twelve months. The bear thesis needed a new argument before Q&A started.

But here is what the press release buried and the transcript revealed: €1.2B of customer prepayments essentially funded the entire €1B inventory build needed to execute 2026 orders. CFO Klaus Neumann was explicit: “the customer contributed €1.2 billion to our cash flow working capital improvement.” Rheinmetall didn’t generate cash by becoming more operationally efficient. It generated cash because sovereign governments with urgent capability gaps agreed to finance the factory expansion in advance.

That is a qualitatively different mechanism from what the EDIP moat framing implied. A regulatory moat says: you can’t easily route around us. What this quarter revealed is something more offensive: our customers are writing advance checks to build the infrastructure they cannot exist without. The distinction matters because it changes the terminal multiple conversation. A company that is hard to replace is worth 25-28x earnings. A company whose customers pre-fund its own competitive moat, because the alternative is a capability gap they cannot accept, is worth more. Exactly how much more is the central question of the three-year scenario work below.

CEO Armin Papperger made the subtext explicit in Q&A. Pushed on cash conversion guidance of “>40% of EBIT”, well below the 70-90% some investors wanted, he said: if Arminius and the large ammunition contracts arrive with expected advance payments, “we must not longer speak about cash conversion rates.” That line deserves arithmetic, not just quotation. The Arminius program carries a first fixed order of approximately €12.5B expected in Q2/Q3 2026, with 20-30% advance payment terms under negotiation. That’s €2.5-3.75B in cash before Rheinmetall produces a single vehicle. The concurrent ammunition framework, 600-700K artillery rounds, another multi-billion contract, carries similar terms, another €600-800M. Combined: north of €4B in sovereign advance payments in a single quarter, against full-year 2025 OFCF of €1.218B. The “>40% guide” is a floor. The ceiling is deliberately unguided because the upside is binary and management knows it.

The prior frame, EDIP moat, was defensive. The correct frame after this quarter is offensive: your customers don’t fund your competitors.

The Weekend Nobody Had in Their Model

On the morning of this earnings call, Papperger had already spent his weekend taking ministerial-level phone calls from governments across Europe and the Middle East. Iranian attacks had triggered urgent inquiries about Skyranger, Skynex, and Millennium Gun platforms, gun-based air defense systems already deployed across Gulf states, proven effective against drone swarms at a fraction of the cost of missile-based alternatives.

The economics Papperger cited are not marginal. In the first 72 hours of the conflict, the US and its allies expended roughly 2,000 munitions at approximately $4B total cost. Against that, a gun-based system shooting down a $50,000 Shahed drone with a $25,000 round isn’t just arithmetic, it’s a doctrine change. The Western world built its air defense architecture around missiles. The cost curve of drone warfare has made that architecture financially unsustainable at scale. Rheinmetall, which has been producing gun-based air defense systems for 25 years, is the primary industrial beneficiary of that realization.

And because Rheinmetall controls the weapons systems, the feeding mechanisms, and increasingly the fire control software internally, it can scale faster than any competitor relying on external subcomponents. Papperger confirmed that deliveries to the Middle East are possible by August, systems already in production, without signed contracts. He also stated that expanding beyond 400 systems per year doesn’t require significant additional investment given the level of internal integration. “There is more in,” he said of the Air Defence guidance, and he said it twice.

The Air Defence segment was guided at €0.9-1.0B for 2026 at approximately 19% margin. I think that number is the most deliberately understated line in the entire guidance package, and it’s understated for the right reason, Rheinmetall doesn’t yet have signed contracts from the weekend’s conversations. When those contracts arrive, they will be incremental to a guidance figure already set as a floor.

This matters for the prior thesis in a specific way. Every scenario we built before this quarter was anchored to European rearmament as the demand driver. The Middle East development isn’t a replacement for that thesis, European sovereign urgency is still the structural base. It’s an additional demand vector that operates on its own logic, its own timeline, and its own contract structure. Our prior TAM assumption was too narrow, and the Air Defence segment is where that error shows up most directly in the numbers.

What the Segment Restructuring Is Really Saying

Rheinmetall introduced a five-segment reporting structure effective January 1, 2026: Vehicle Systems, Weapon and Ammunition, Digital Systems, Air Defence, and Naval Systems. This is not cosmetic.

Under the old Electronic Solutions umbrella, it was structurally impossible to evaluate whether Rheinmetall’s digitization narrative, TaWAN, Battlesuite, D-LBO, the tactical core, was generating real economics or just large one-time project revenue dressed in platform language. Now Digital Systems and Air Defence have standalone P&Ls that management must defend every quarter. The platform narrative now has accountability attached to it, which is exactly what a sophisticated investor should want and exactly what management should be willing to provide if the economics are real.

One important nuance that Reuters and most initial sell-side coverage missed entirely: a €48M one-time divestment gain was embedded in the 2025 Electronic Solutions numbers, surfaced only because Warburg Research pushed on it in Q&A. Strip that out and the 2026 Digital Systems margin guide of ~12% represents essentially flat underlying performance rather than the compression the headline implies. Papperger was direct about the why: F-35 integration, satellite programs, and drone capacity are running costs ahead of their associated revenues. Digital Systems in 2026 is absorbing setup costs for programs that should generate revenue at materially higher margins from 2027-2028. The segment margin guide isn’t deterioration. It’s the income statement signature of deliberate investment front-loading.

*Includes €48M one-time divestment gain; underlying ~12%

The margin line deserving the most attention is Weapon and Ammunition at 29.3% for the full year. The Q3 2025 result of 23.2% had rattled bulls, it was the first genuine evidence that the ammunition scarcity premium might be normalizing as new European capacity came online. The full-year recovery to 29.3%, with 29-30% guided for 2026 knowing that capacity is expanding, closes that debate definitively. Management is guiding the same range for 2026 with full visibility into competitor capacity additions. That is not what margin normalization looks like.

One additional item that deserves mention precisely because it was asked about directly and explained clearly: Vehicle Systems Q4 margin came in light. The reason, per Papperger, was specific and structural, approximately twelve heavy weapon carriers completed and sitting in the yard because the German MoD lacked capacity to perform final acceptance checks. Software issue on one vehicle variant, now resolved. Neither is a manufacturing problem. Both represent deferred Q1 2026 revenue rather than lost revenue. Without that explanation, a reader of the press release would see a soft Q4 Vehicle Systems result and have no framework for interpreting it. The transcript is where this became clear.

What the Upgraded Moat Is Worth, and What We Still Can’t See

Let me be precise about what changed and what didn’t, because conflating the two leads to the wrong investment conclusion.

What changed: The moat framing. The prior thesis described EDIP regulatory entrenchment as the primary competitive advantage, a defensive position that made Rheinmetall hard to route around. The correct post-earnings framing is more aggressive: Rheinmetall’s sovereign customers are pre-funding its expansion because the alternative, a capability gap they cannot remedy on any acceptable timeline, is worse than writing the check. That shifts the quality of the advantage from regulatory to structural dependency. It also shifts the terminal multiple conversation. A regulatory moat erodes when regulations change. A dependency moat deepens every time a new Boxer is delivered, a new Skyranger deployed, a new D-LBO software integration completed, because each event increases switching costs, extends lifecycle revenue, and makes the operational interoperability argument for the next procurement cycle. The difference between these two framings is approximately €500-700 in fair value at current earnings levels, which is material when you’re deciding whether €1,542 is full price or merely full price for the old frame.

What didn’t change: The valuation concern. The stock is at €1,542 against the same earnings base we were analyzing before the print. The multiple absorbed most of the good news in real time, which is what a market-cap-weighted momentum vehicle at 40x forward earnings does when a strong print confirms but doesn’t dramatically exceed consensus. The probability-weighted expected return from here is approximately 8% annually. Unexciting for the complexity and binary risk carried. The high-certainty entry for 15% annual compounding remains €900-1,100. The upgraded thesis raises the ceiling. It does not lower the entry.

What we still can’t see: The platform narrative remains unverified by financials. Management talks convincingly about TaWAN, Battlesuite, and the tactical core. But the reporting still provides no software or services revenue split, no attach rate, no lifecycle revenue per platform, and no recurring-economics bridge for Digital Systems. The segment structure improvement is genuine. The accountability gap is real. If by Q3 2026 Digital Systems is still reporting only project revenue with no platform-specific KPIs, the narrative is running materially ahead of the economics, and the 38x terminal multiple in the bull case needs revisiting.

Two additional risks became louder this quarter rather than quieter. First, M&A: the net cash position and Papperger’s explicit statement that the balance sheet “paves the way for more acquisitions” means this is an active risk, not a theoretical one. One manageable deal under €1.5B in a core domain is fine. A €3B+ deal in an unproven domain, space, cyber, commercial logistics, while simultaneously integrating NVL, executing Arminius, and ramping D-LBO is a management bandwidth problem that sits at the center of the bear case. Second, the Lockheed JV: Papperger was unusually candid, “I must say it’s too slow”, and confirmed parallel conversations with other missile manufacturers. The rocket motor factory, qualifying end of 2026, will serve “whatever missile it is.” The Lockheed relationship has moved from strategic anchor to one option among several. That’s a disclosure change worth tracking.

Three Years, Three Scenarios

The thesis upgrade has a price tag, and the question is whether you’re paying for the old thesis or the new one. Three years gets us to end-2028, when major capacity investments are substantially complete, D-LBO and TaWAN are either generating recurring revenue or they aren’t, NVL is integrated, and the Arminius first tranche is in full delivery. By then you know which thesis was right.

Bull case, 35% probability: “The Infrastructure Layer Wins”

Arminius closes Q2 2026 with full prepayments. Middle East air defence orders add €500M+ incrementally to guidance. W&A margin holds 29-30%, new capacity absorbed by growing demand, not normalization. Digital Systems scales to €3.5B+ at 15%+ margin as D-LBO and TaWAN convert to recurring economics; management begins disclosing platform-specific metrics that justify a re-rating. NVL frigate orders arrive. US optionality, XM30, howitzer, Navy, converts partially, none of it currently in the business case. One manageable M&A deal. No dilution.

FY2028 revenue reaches €27.5B at 21% operating margin. Adjusted EPS approximately €82. At 38x, justified if Digital Systems recurring revenue is visible, US exposure has partially converted, and the sovereign prepayment model is proven durable, the stock is worth €3,100. Adding approximately €110 in cumulative dividends over three years, total return is roughly 107%, or 26% annually from today’s price.

Base case, 50% probability: “Execution Meets Reality”

Arminius slips to Q3 on German procurement timing. W&A margin drifts modestly to 27-28% as European capacity expands, the scarcity premium partially but not fully normalizes. Middle East adds incrementally, not transformatively. Digital Systems stays in investment mode; recurring revenue visible but below 10% of segment by 2028. NVL integrates successfully but synergies remain a 2030 story. One manageable M&A deal, €1-2B range. FCF conversion of 55-65%, prepayments real but lumpy.

FY2028 revenue reaches €24B at 19.5% operating margin. Adjusted EPS approximately €67. At 28x, appropriate for a high-quality defense prime with growth decelerating toward 20-25% annually and no proven platform re-rating, the stock reaches €1,875. Total return including dividends approximately 27%, or 8.3% annually. Serviceable, but at this valuation you are not being paid for the complexity or the binary risks you’re carrying.

Bear case, 15% probability: “Industrial Physics Win”

Arminius slips into 2027 on German parliamentary budget disputes. W&A margin normalizes to 22-24% as KNDS, Hanwha, and BAE capacity erodes the scarcity premium faster than demand grows. Digital Systems disappoints, integration costs persistent, D-LBO delayed past contract milestones. A large dilutive acquisition in 2026-2027 spreads management thin across too many simultaneous programs. Ukraine ceasefire in 2027 triggers a 20-25% sector-wide de-rating as European governments reprioritize. Net cash position consumed by M&A, FCF normalization delayed.

FY2028 revenue reaches €19.5B at 16.5% operating margin. Adjusted EPS approximately €46. At 18x, full normalization to mature European industrial multiple, the range where BAE Systems and Leonardo trade in a normalized cycle, the stock reaches €825. Total return including dividends approximately -40%, or -19% annually.

Upside/downside ratio: approximately 2.5:1. The asymmetry is real. But at €1,542, 40x FY2026e EPS, you are already paying for base-to-bull execution. The probability-weighted expected return of ~8% annually is unexciting for a stock carrying geopolitical binary risk at a growth multiple. The high-certainty entry for 15% CAGR is €900-1,100, unchanged from before the print. What changed is the ceiling. The bull case got higher because the mechanism got stronger. The entry requirement did not.

What to Watch

Five signals, each observable before annual results confirm them.

Arminius timing and prepayment rate. The single most important event of 2026. Q2 closure with 25%+ advance payment is the bull signal, and the arithmetic is dramatic enough that it will be visible in the Q2 OFCF number before management formally announces anything. A slip to “H2” in management language is the tell that the base case is becoming the bear case. Watch specifically for any change from “Q2/Q3” to vaguer calendar framing in Q1 results commentary.

W&A quarterly margin. The most important operational number in the entire portfolio. If it holds 28-30% through mid-2026 as new capacity comes online, the scarcity premium is structural and the bear case on margin normalization is losing. A drift toward 25% by H2 2026 means competing European capacity is arriving faster than demand is absorbing it. This one number, tracked quarterly, separates the bull from the base case more reliably than any other metric.

Air Defence order announcements. The weekend’s ministerial conversations should convert to signed contracts in Q1/Q2 2026 if the demand is real. If the segment is tracking above the €0.9-1.0B full-year guide by Q2 earnings, the Middle East vector has confirmed. If not, the phone calls were exploration. Papperger’s credibility on this specific claim is high, he was specific about August delivery timelines on systems already in production, which makes non-confirmation a meaningful signal rather than routine guidance conservatism.

Digital Systems KPI disclosure. The segment exists as a standalone entity starting January 1, 2026. This is its first full year of accountable reporting. By Q3 2026, any sophisticated investor should expect at least a directional disclosure of software or services as a percentage of segment revenue, or a first TaWAN attach rate, or a D-LBO vehicle completion milestone number. If by year-end 2026 the Digital Systems disclosure still contains only revenue and margin with no platform-specific metrics, the narrative is running ahead of the economics and the 38x terminal multiple assumption in the bull case requires a significant haircut.

M&A announcement size and domain. Papperger said explicitly the balance sheet “paves the way for more acquisitions.” The question is not whether another deal comes, it almost certainly will. The question is whether it is additive or dilutive to the thesis. Under €1.5B in a core domain with clear integration rationale: manageable, track the integration execution. Above €3B, or in an unproven domain while simultaneously integrating NVL and executing Arminius: the bear case just got more probable, and the position sizing conversation changes.

What Changed

The prior thesis was: Rheinmetall has a regulatory moat, FCF is the unresolved question, and the stock is full price for a strong business.

The current thesis is: the FCF question is closed, and what it revealed underneath is more durable than a regulatory moat. Rheinmetall’s sovereign customers are pre-funding their own dependency on its industrial capacity. The Middle East development extends that dependency globally before the European thesis has even fully converted. The company is becoming mandatory infrastructure for rearmament, not because it lobbied for the right regulations, but because no alternative exists at the scale and breadth required, and the cost of a capability gap has become politically unacceptable.

That thesis has a higher ceiling than the prior one. The bull case at €3,100, 26% annual compounding, was not available under the EDIP moat framing. It is available under the sovereign dependency framing, but only if Digital Systems converts its narrative to recurring economics, W&A margins prove structural rather than cyclical, and Arminius arrives with the prepayments management has effectively promised.

The entry price has not changed. The high-certainty entry remains €900-1,100. At €1,542 the market has already partially priced the thesis upgrade. What you’re deciding today is whether the remaining upside, the difference between the base case at €1,875 and the bull case at €3,100, is worth the binary risks you’re carrying at a 40x multiple.

Papperger closed the call with a line that was either corporate modesty or strategic confidence: “We are responsible for the performance. You are responsible for the share price.” My read is that he believes the business has already made the harder argument, that the mechanism is proven, the demand is funded, and the industrial capacity is being built on the sovereign customer’s advance payment. The question for investors is whether you are paying for the argument or waiting for the proof.

At €1,542, you are paying for most of the argument. The proof arrives quarterly. The first test is Q2 2026, when Arminius either closes or it doesn’t. Everything else follows from that.

Subscribe now

• OCI growth confirmed the demand story: cloud infrastructure accelerated to 84% growth, while AI and multi-cloud database adoption continued scaling rapidly.

• The real change was structural: large AI contracts increasingly use customer-funded GPUs or upfront prepayments, shifting capital intensity away from Oracle.

• If BYOH scales: Oracle evolves from capital-constrained builder to AI infrastructure orchestrator, with backlog converting into revenue without the same balance-sheet strain.

Q2 was the quarter where Oracle’s AI story became impossible to dismiss and impossible to own comfortably. The demand was real; the balance-sheet burden suddenly was too. Q3 matters because Oracle finally showed what the bridge across that gap might look like.

That is the real update to the view.

What We Said Going In

Our prior thesis had three chapters. The AI Tollbooth: Oracle sat at the control point between decades of mission-critical enterprise data and the AI models that needed it. Strategic Vindication: the architecture Oracle built between 2016 and 2020, optimized for high-performance, data-intensive workloads, turned out to be precisely what the AI era demanded. And the Capital Chasm: the demand was unmistakably real, but Oracle had begun to look like a company that found the future and might strain its balance sheet building it.

Our variant going in was specific: the market was pricing a funding crisis, but the $553 billion RPO was effectively collateral for financing not yet announced. We expected a specific Stargate joint venture or sovereign wealth fund disclosure to be the binary catalyst.

We got the financing. The $30 billion raise was oversubscribed before the quarter even closed, at investment grade, with bond investors, who are smarter than equity markets about credit risk, effectively voting on that backlog’s collateral value. The cleanest version of the bear thesis weakened sharply. On that, we were right.

What we got wrong was the structure of the answer. We expected a one-time JV announcement. What Oracle revealed was a recurring business model shift: customers prepaying for capacity, or supplying their own GPUs, with Oracle providing the operating layer on top. That is more durable than a single financing event. A JV solves one capital problem. A structural model change means the capital problem recurs less often. We will take being wrong in that direction.

The Numbers

The subtitle above is the quarter. What follows is the evidence.

OCI growth accelerated to 84%. Free cash flow deteriorated to negative $24.7 billion trailing. Both numbers matter. Only one was surprising.

Oracle kept FY26 revenue guidance at $67 billion, kept FY26 capex at $50 billion, and raised FY27 revenue guidance to $90 billion. This was the first quarter in more than 15 years where both organic revenue and non-GAAP EPS grew 20% or more in USD. The growth story is no longer in question. What this quarter needed to answer was whether the model underneath that growth was becoming more or less sustainable. That is a different question, and it got a different kind of answer.

The Problem Was Never Demand

Our previous arc was right on strategy and too optimistic on economics. Oracle clearly found a privileged place in the AI stack: close to enterprise data, useful to the model layer, more relevant in the age of agents than the market appreciated when it was still grading the company as a late cloud entrant. Q2 proved that. The problem was that Oracle then looked like a company that had discovered the future and might strain the balance sheet building it.

Q3 did not answer “is demand real?” That debate is over. Multi-cloud database revenue grew 531% year over year. AI infrastructure revenue grew 243%. Oracle now has 33 Microsoft regions live, 14 Google regions live, and AWS went from 2 live regions at the start of Q3 to 8 by quarter-end, with 22 expected by Q4-end.

The market still frames Oracle as if the relevant question is whether it can become the fourth hyperscaler. That is the wrong frame. Oracle’s real opportunity is narrower, more strategic, and more defensible: the multi-cloud database layer, the enterprise data control point, and the operating environment where private data meets frontier models. Oracle is not building another AWS. It is building the place enterprises go when they want to use the best models without surrendering control of their most valuable data, where vector embeddings, MCP access, advanced security controls, and data co-located with agents are the requirements, not the differentiators. That is a use case Oracle’s architecture is unusually well aligned to serve, which is different from saying competitors cannot address it, but does suggest Oracle holds structural advantages that most coverage has underpriced.

A Better Kind of Backlog

A $553 billion backlog is no longer interesting because it is huge. It is interesting if it is becoming better.

That is what changed this quarter. The most important sentence in the press release was not about OCI growth. It was this: most of the increase in Q3 RPO came from large-scale AI contracts where Oracle does not expect to raise incremental funds, because the equipment is either funded upfront through customer prepayments or bought directly by customers and supplied to Oracle.

That sentence is the quarter.

Since the last call, Oracle signed more than $29 billion in contracts under this structure, bring-your-own-hardware and upfront customer payments, enabling expansion, in Magouyrk’s words, “without any negative cash flow from Oracle.”

Think about what this means structurally. In the old model, Oracle finances the GPUs, builds the data center, sells cloud capacity, carries all the capital risk. In the new model, the customer supplies the GPUs, Oracle provides the facility, network, database, and software stack. The customer has equity in the hardware sitting inside Oracle’s building. Oracle captures the operating and software layer without deploying the most capital-intensive component. Lower margin per contract, but dramatically higher return on Oracle’s own capital. That distinction has not made it into consensus models, and it is the intellectual center of why Q3 represents an inflection rather than just an acceleration.

The bears were right that this affects margins. What they missed is that it changes Oracle’s return on its own capital entirely. Those are not the same thing, and treating them as equivalent is where the bear case has been analytically imprecise.

The Delivery Quarter

The smartest thing Oracle did on this call was not claim demand exceeds supply. Everyone knew that. It gave the market an actual operating bridge from demand to revenue.

More than 10 gigawatts of partner-backed capacity secured over three years. More than 90% of that capacity funded through partners. Manufacturing sites tripled. Rack output up 4x. Time from rack delivery to revenue down 60%. More than 400 MW delivered in Q3. 90% of committed capacity on or ahead of schedule. Delivered AI capacity at 32% gross margin, above the prior 30% guidance.

Q2 proved Oracle could sign. Q3 is the first quarter that suggests Oracle can deliver. The stock no longer needs promotion. It needs sequencing. Oracle finally provided it.

One financing detail deserves more attention than it has received. The $30 billion raise comprised two instruments that are not equivalent. The investment-grade bonds carry no dilution. The $4.95 billion mandatory convertible preferred stock will convert to common equity on a fixed timeline, regardless of stock price. At $151 heading into this print, the conversion terms are manageable. At $180-200+, the dilution becomes more meaningful. The market celebrated the oversubscribed raise without distinguishing between the two instruments. That distinction matters, and it belongs in any honest accounting of what the financing actually cost.

The Hidden Second Leg

The easiest mistake after this quarter is thinking it was all OCI. It was not.

Cloud applications revenue grew 13% to $4.0 billion. Fusion ERP grew 17%. NetSuite grew 14%. More tellingly, cloud applications deferred revenue grew 14% in constant currency, faster than in-quarter revenue growth of 11%, which is the forward signal most coverage skipped. More than 2,000 customer go-lives in the quarter. Management repeatedly framed these not as single-app wins but as “ecosystem automation” conversations combining OCI, AI Data Platform, Fusion Applications, industry suites, database, and AI tooling together.

If OCI pulls more enterprise data into Oracle’s orbit, Oracle doesn’t just monetize the compute. It monetizes the database, the tooling, the applications, and the workflows around them. That is a materially better business than renting GPUs, and the applications trajectory suggests it is beginning to compound.

There is a second margin lever that one plausible reading of the numbers supports: Oracle’s restructuring charge is $961 million year-to-date, up 337% YoY, without a major acquisition to explain it. Management said AI code generation is allowing smaller engineering teams to deliver more software more quickly. If that is what is happening at scale, and the new product output this quarter is at least consistent with the claim, then AI may not just be driving Oracle’s infrastructure demand. It may also be improving Oracle’s own software economics in ways that don’t yet appear in any forward model. That is inference, not proof. But it is the kind of inference worth tracking.

Variant Perception

Consensus sees Oracle as a cloud infrastructure winner with a recovering balance sheet, still hostage to capex intensity, customer concentration, and backlog-conversion uncertainty. The more interesting framing is that the market remains anchored to Oracle-as-capital-constrained-builder and has not fully updated to Oracle-as-orchestrator, a company where more of the infrastructure burden may increasingly sit with customers and partners while Oracle captures the recurring cloud, database, and software economics on top. If that is right, the FY2027 $90 billion guide carries more management conviction than the market realizes, because it was set after the BYOH model was already in place for $29 billion of new contracts. The silent risk to that framing is interest expense: $1.18 billion in Q3, annualizing toward $5-6 billion against $7.4 billion in non-GAAP operating income, growing with every quarter the existing debt load compounds. The BYOH model helps on new contracts. It does not reduce the cost of servicing what is already on the balance sheet.

Three-Year Scenarios

The bear doesn’t require Oracle to fail outright. It requires the BYOH model to prove narrower than management suggests, effective for a handful of mega-deals but not scalable as a standard contract structure, forcing Oracle back to self-funded capex just as $134 billion in debt demands full servicing. Add AI inference efficiency gains eroding compute demand faster than new use cases emerge, and the equity story compresses toward a leveraged utility multiple. The bear case is narrower than it was six months ago. It is not gone.

The base assumes BYOH becomes roughly 40% of new RPO, OCI sustains above 60% growth through FY2027 then moderates, multi-cloud database reaches $8-10 billion annualized by FY2029, and software margins expand 300-400 basis points from AI-driven R&D efficiency. This is a good business with a complicated balance sheet that slowly resolves. The re-rating is real but partial.

The bull requires OpenAI to remain the dominant AI model company, begin its $300 billion commitment on schedule, and the BYOH structure to become the standard contract form for large AI deals, not the exception. If those three things happen simultaneously, Oracle re-rates from legacy software toward a full-stack enterprise AI platform multiple. The probability-weighted expected return across all three scenarios is approximately +85% over three years, or ~23% annualized, with upside/downside skew of roughly 3x.

What We Think Now

The single number that will tell us whether this thesis is holding is BYOH as a percentage of sequential RPO adds. If it grows quarter over quarter, the model is scaling and the variant is strengthening. If it shrinks, if Oracle reverts to self-funding incremental growth in Q4 or Q1 FY27, the capital intensity concern reopens immediately and the base case becomes the ceiling rather than the midpoint. The number that could quietly undermine everything else, regardless of BYOH trajectory, is interest expense. It doesn’t announce itself. It just compounds, quarter after quarter, against an operating income line that needs OCI to keep accelerating to stay ahead of it. Watch both numbers. One tells you whether the model is working. The other tells you whether there is enough time for the model to work before the balance sheet demands its answer.

What we thought before was that Oracle had discovered a strategically privileged position in AI but risked letting capital intensity overwhelm the equity story. What we think now is that risk is still present, but Q3 is the first quarter where Oracle showed a believable way around it. The story is no longer simply “Oracle has huge demand.” It is now “Oracle may have found a way to keep the upside while outsourcing more of the funding burden.” That is a real change. The bear case has narrowed from “the balance sheet may not survive the build phase cleanly” to “the new funding model may not scale as broadly as management suggests.” That is still a serious question. But it is a much narrower one.

And that is why Q3 mattered. Not because Oracle beat. Because Oracle finally made the path from backlog to revenue feel more like a business model and less like a hope.

Subscribe now

The story of distressed credit, though, is far more nuanced — and the mechanics behind it are worth understanding.

What is distressed credit, exactly?

When a company struggles to meet its debt obligations, its bonds and loans often begin trading at a significant discount in secondary markets. Investors who buy this debt are making a bet on how the story ends — through restructuring, recovery, or in some cases, liquidation. It’s one of the most complex corners of credit markets, and one of the least understood outside of it.

The central question any distressed investor must answer isn’t purely financial.

It’s almost diagnostic: is this business broken beyond repair — or just broken right now?

A company might be distressed because of a poorly structured leveraged buyout, a temporary industry downturn, or a management misstep — none of which necessarily reflect the underlying health of the business itself.

Curious where this leads?

Akshay Shah - CIO of KYMA Capital and former senior investor at Blackstone’s GSO Capital Partners- is exploring exactly this, live on 12th March with Kristal.

Register Now!

Two very different playbooks

How an investor answers that diagnostic question determines everything about how they engage with a distressed situation.

The more aggressive approach treats distressed debt as a liquidation exercise — recover as much as possible, as fast as possible. It works. But it often destroys value in the process. Prolonged bankruptcy proceedings burn cash. Key employees leave when the future looks uncertain. Customers defect. Suppliers tighten terms. The business that emerges, if it emerges at all, is frequently worth far less than it could have been.

The alternative — sometimes called constructive or collaborative restructuring — starts from a different premise. If what a business is worth alive significantly exceeds what it’s worth in pieces, then preserving it isn’t just good ethics. It’s good investing. This means working with management rather than against them, stabilising operations, and reaching restructuring agreements before they become protracted courtroom battles. It requires more than credit skills — operational judgment, relationship capital, and the conviction to back a business through its most difficult chapter.

Why this matters right now?

Distressed and special situations funds have collectively raised $100 billion over the past two years, with the 10 largest currently targeting almost $50 billion more, according to Hedgeweek. Institutional appetite for this asset class is significant — and the opportunity set is only growing.

Understanding how distressed credit works, and the different philosophies that shape how it’s practiced, is increasingly relevant for anyone navigating today’s credit markets.

See you there, on:

📅 Thursday, 12th March

🕐 5:00 PM SGT / HKT | 2:30 PM IST | 1:00 PM GST

Register Now!

• Broadcom’s $100B AI revenue “line of sight” for 2027 isn’t guidance, it’s a logistics statement backed by supply chain commitments and binding customer deployments.

• Hyperscaler “insourcing” doesn’t displace Broadcom; it deepens dependence because Broadcom industrializes custom silicon at scale.

• The market still values Broadcom as a semiconductor cyclical while it is increasingly the infrastructure layer underneath hyperscaler AI.

From Bloomberg:

“Broadcom reported first-quarter revenue of $19.31 billion, beating analyst estimates, with AI chip sales doubling year-over-year to $8.4 billion. The company guided for second-quarter revenue of approximately $22 billion, nearly $1.5 billion above Wall Street expectations, and disclosed that it has line of sight to more than $100 billion in AI chip revenue in fiscal 2027.”

In December, Broadcom reported a strong quarter, guided conservatively, and watched $370 billion of market cap disappear. The reason was a single clarification: the OpenAI deal was a framework for 2027-2029, not near-term contracted revenue. Hock Tan had separated what was real from what was hoped for, and the market, drunk on AI narrative, treated honesty as a miss.

We wrote at the time that the selloff was a category error. Broadcom lived in the physics economy: $73 billion in contracted backlog, $21 billion in Anthropic binding purchase orders, five confirmed XPU customers with delivery schedules. The stock wasn’t pricing any of that accurately because the market was busy pricing a different company, one that played the narrative game, and Broadcom kept refusing to be that company.

What we got right: the discipline discount was real and it was an opportunity. What we underestimated: discipline, once established as the operating mode, doesn’t just correct a mispricing. It creates a new kind of leverage. Tonight Hock Tan used twelve quarters of credibility to say something that would have been dismissed as promotional coming from anyone else: Broadcom has line of sight to more than $100 billion in AI chip revenue in a single fiscal year. In prepared remarks. Reviewed by legal. Backed by a supply chain already booked through 2028.

The market believed him. That is the whole story.

What “Line of Sight” Actually Means

The phrase is doing a lot of work. “Line of sight” is not a forecast. It is not guidance. It is Broadcom’s way of saying: we can see the delivery schedule from here, and the delivery schedule adds up to $100 billion.

The mechanics are worth understanding because they explain why this number is more grounded than it sounds. When Bernstein’s Stacy Rasgon walked through the gigawatt math on the call, Anthropic 3 GW in 2027, OpenAI 1+ GW, Meta multiple GW, Google 3+ GW, Tan confirmed the aggregate is tracking toward 10 GW of customer deployments. Each gigawatt of compute represents roughly $20 billion in chip content. The arithmetic closes. The $100 billion is the floor, not the midpoint, and Tan’s language was explicit: “we have also secured the supply chain required to achieve this.”

That last sentence is the one nobody followed up on, and it is arguably more important than the number itself. Broadcom has pre-committed CoWoS capacity, High-Bandwidth Memory allocation, and substrate supply through 2028. In a world where every AI infrastructure company is competing for the same constrained manufacturing inputs, Broadcom has already reserved the majority of the relevant capacity for three years. The $100 billion disclosure is not a goal. It is a logistics statement. The supply chain is booked. The customers have committed. The question is execution, and Broadcom has beaten its AI revenue guide every quarter for nine consecutive quarters.

The reason this credibility exists, and this is the part the market still underweights, is precisely because of December. The CEO who told investors the OpenAI deal wasn’t near-term revenue, at personal cost, is the same CEO now saying $100 billion in 2027 is visible. That’s not a coincidence. Discipline is the currency. Tonight Tan spent some of it, and the market’s willingness to take the number seriously is the return on twelve quarters of refusing to oversell.

Why Insourcing Makes Broadcom More Valuable, Not Less

The bear case that has followed Broadcom for three years goes like this: Google has TPUs. Meta has MTIA. Amazon has Trainium. OpenAI is building its own silicon. Eventually these internal programs mature, volume moves in-house, and Broadcom’s economics compress. The customer is the competition.

I think this has the causality exactly backwards, and the reason it’s backwards explains why Broadcom’s moat is structurally different from what most people think it is.

The mistake is assuming the hard part of custom silicon is the architecture. It isn’t. Google has been designing TPUs since 2016. They are genuinely good chips. But Google still uses Broadcom, not because they can’t design the silicon, but because the hard part is everything that comes after the architecture: SerDes IP optimized for their specific cluster topology, advanced packaging at yields that make the economics work, 100,000-unit production runs without defect cascades, networking integration across the entire compute cluster, and the ability to do all of this on a roadmap that doesn’t slip. Tan described it plainly when asked about customer-owned tooling competition: “Can you produce 100,000 of those chips quickly at yields you can afford? We don’t see too many players in the world that can do that.”

This is a 20-year learning curve. It cannot be acquired by hiring engineers or replicated by running a program for three years. And here is the structural implication: the more strategically important custom silicon becomes to hyperscalers, the more their competitive position in AI depends on their XPU program working, the less they can afford to experiment with alternatives to Broadcom. Failure at cluster scale is not a bad quarter. It is an existential setback. When Google’s Ironwood TPU program is co-designed with Broadcom, replacing Broadcom doesn’t mean finding a better chip vendor. It means rebuilding the entire development relationship, losing the SerDes IP, starting from scratch on packaging yield, and falling 12-18 months behind on the training infrastructure the business depends on.

This is why the customer list expanding from three to six in four quarters is the opposite of a concentration risk. Each new customer, Anthropic, then the implied fifth, now OpenAI confirmed tonight, adds revenue while deepening the organizational muscle required to run parallel deep-engagement programs simultaneously. Broadcom now has six hyperscale customers who each view their XPU program as strategic infrastructure rather than a vendor relationship. The switching cost is the entire AI roadmap. That is not a chip vendor. That is a structural dependency.

The sixth customer confirmation also directly addresses the most specific bear argument from last quarter. Meta’s MTIA program, cited in several analyst notes as evidence of hyperscaler insourcing displacing Broadcom, was defended explicitly tonight: “alive and well, we’re shipping now.” The insourcing thesis didn’t displace Broadcom. It hired Broadcom as the industrial partner to make insourcing real.

The Asset Nobody Is Pricing

There is a piece of tonight’s print that received almost no analyst attention, and it may be the most important element for anyone thinking about where the stock goes over three years.

AI networking, Tomahawk switching, SerDes, DSPs, represented 33% of Q1’s $8.4 billion AI revenue, approximately $2.8 billion. In Q2, it is trending to 40% of $10.7 billion: roughly $4.3 billion in a single quarter. Annualized from Q2, that is $17 billion of AI networking revenue, growing faster than the XPU business, captured by a product, Tomahawk 6 at 100 terabits per second, that has no competitive equivalent in market today.

Arista Networks is the closest comparable. Arista trades at approximately 38 times forward earnings on $8 billion of annual revenue growing at 25%. Broadcom’s AI networking business is more than twice as large, growing faster, and receives zero standalone valuation credit because it sits inside the semiconductor segment disclosure. If you apply Arista’s multiple to Broadcom’s AI networking business, networking alone is worth roughly $500 billion. Add the XPU business at 22 times, infrastructure software at 20 times, and the sum-of-parts reaches $2.1-2.2 trillion against a current enterprise value of $1.54 trillion. The gap is not a precision argument. It is a classification argument. Broadcom is being valued as a semiconductor cyclical when it is simultaneously a networking monopolist growing faster than any pure-play networking company, an XPU franchise with $100 billion in confirmed 2027 visibility, and a software annuity generating 78% operating margins.

Charlie Kawwas disclosed something in Q&A that made this gap wider still. Scale-up networking, connecting chips within a rack, previously dominated by NVIDIA’s NVLink and InfiniBand, is now being specified as Ethernet by Broadcom’s XPU customers. “A lot of the XPU designs we’re doing, we’re being asked to scale up through Ethernet.” This extends Broadcom’s networking moat into the one architectural layer where NVIDIA had been structurally protected. If Ethernet wins scale-up as definitively as it has won scale-out, the networking business grows faster and commands a higher multiple than current estimates support.

The VMware piece is simpler. Software grew 1% year-over-year in Q1, the one number that wasn’t clean, but ARR grew 19% and bookings exceeded $9.2 billion, consistent with subscription conversion mechanics that defer reported revenue. Q2 is guided at $7.2 billion, implying 9% growth. The monitoring item is straightforward: above $7.4 billion confirms the ARR-to-revenue conversion is working; two consecutive misses requires genuine re-examination of whether VMware pricing pressure is generating quiet churn rather than the delayed recognition management describes. VMware is not the thesis. It is the 78%-margin cash engine that funded $10.9 billion in capital returns last quarter and allows Broadcom to run the AI program without treating the balance sheet as a constraint. It finances patience.

Three Futures

The honest version of a three-year scenario analysis acknowledges that the range is wide and the assumptions are explicit.

The bull case requires the $100 billion 2027 disclosure to deliver as described, not perfectly, but within range, and for the market to re-rate Broadcom toward infrastructure platform multiples rather than semiconductor cyclical multiples. OpenAI ramps to 3+ GW in 2028, a seventh customer is added, networking mix holds above 40%.

The base case is what happens if the $100 billion is real but lumpy, VMware grows modestly, gross margins hold at 77%, and the market remains in “show me” mode through FY2027 before partially re-rating on delivery. This is the scenario where Broadcom keeps executing and the stock closes roughly half the valuation gap implied by the SOTP analysis.

The bear case requires a genuine AI capex pause, not a sentiment wobble but a real pullback in hyperscaler compute spending, combined with at least one major customer program slipping a generation. This is the scenario where Broadcom’s supply chain commitments become inventory exposure and the multiple reverts toward cyclical semiconductor territory. It requires several things going wrong simultaneously against the backdrop of $21 billion in Anthropic binding purchase orders and supply chain committed through 2028. Possible. Not the base case.

The upside/downside skew at 2.2x argues for the position. The bear case floor at $240 assumes the supply chain lockup and binding POs provide no protection. The bull case ceiling at $650 assumes only modest multiple expansion relative to what the networking business alone would justify on a standalone basis.

Four things will tell you which scenario is unfolding: whether AI networking mix holds above 40% of AI revenue (rising above 50% signals the network is becoming more valuable than the compute); whether Q2 and Q3 software revenue confirms the ARR-to-revenue conversion story; whether OpenAI’s relationship migrates from framework language to binding purchase order language with delivery dates; and whether gross margin holds at Hock’s explicit 77% floor, the line he has now publicly staked his credibility on.

The Asset Nobody Expected: Credibility

Our original view was that Broadcom was the disciplined adult in an AI market intoxicated by narrative, being penalized for refusing to blur the line between real orders and aspirational frameworks. That view was correct and the buying opportunity was real.

The updated view is more interesting. Discipline wasn’t just a personality trait that would eventually be rewarded by a market correction. It was the precondition for something more durable: the credibility to make a claim as large as $100 billion in a single year and have institutional investors treat it as a delivery schedule rather than a forecast. That credibility is now a competitive asset. It is why the six customers are willing to share multi-year roadmaps with Broadcom. It is why supply chain partners lock up capacity years in advance. It is why the COT programs that were supposed to displace Broadcom end up running through Broadcom.

The question is no longer whether Broadcom wins AI. It is whether the market understands that winning looks like becoming the industrial infrastructure that makes everyone else’s AI real, not a better chip vendor, but the prime contractor that turns ambition into physics. The $100 billion is not a promise. The supply chain is already booked.

Subscribe now

• Fundamentals fully recovered: record net new ARR, stable retention, and accelerating platform modules.

• Falcon’s advantage is data scale: hundreds of millions of endpoints feeding a single telemetry pipeline competitors can’t replicate.

• Valuation depends on the narrative: security vendor = fairly priced; security data infrastructure = significant mispricing.

There are two ways to understand CrowdStrike.

The first version: a very good security company. Best endpoint detection in the market, strong gross margins, loyal enterprise customers who survived the worst IT outage in history and stayed anyway. Growing at roughly 20% with expanding profitability and a platform that is clearly working. A quality business.

The second version: a data infrastructure company that operates inside the security industry. One that owns the telemetry layer of the enterprise security stack, generated by a single agent sitting on hundreds of millions of endpoints, and is becoming the architectural substrate that every other security decision gets made on top of. Not a security product. Infrastructure.

The difference between these two framings is not academic. If the first version is true, CrowdStrike at today’s price is fairly valued and the stock goes roughly nowhere for the next three years. If the second version is true, the market is still pricing a security company and you are looking at a substantial mispricing.

Q4 FY26 did not resolve which version is correct. But it moved the probability toward the second one more decisively than any prior quarter. And there is a specific reason why the dislocation persists, which makes the question worth asking precisely now.

Why Now

The July 2024 outage created a narrative discount on CrowdStrike that was entirely rational at the time. A company that caused a global IT catastrophe deserved skepticism, elevated risk premium, and multiple compression. The question was whether the damage was structural, permanent churn, retention deterioration, competitive displacement, or operational, meaning recoverable.

Four quarters of data have answered that question completely. Gross retention held at 97% every single quarter of FY26. Net new ARR crossed $331 million in Q4, 18% above the pre-outage record. Full-year net new ARR exceeded $1 billion for the first time in company history. The customers who stayed through the worst-case scenario are the stickiest cohort this company will ever have, and they are expanding, not contracting.

The outage proved something most investors read as negative but was actually strongly positive. It proved the switching cost is not about product quality. CISOs who considered leaving discovered that the years of behavioral telemetry inside their Falcon deployment, the forensic baseline that makes post-incident board reporting possible, was not portable. You could buy a competitor’s product. You could not buy your own history.

The fundamentals recovered. The narrative discount did not. At $391, after a 17% year-to-date decline from a $517 peak, the market is still pricing CrowdStrike as if the outage revealed something bad about the business. It revealed something good. That gap is the opportunity, or the trap, depending on which version of CrowdStrike is real.

What I Got Right and What I Missed

My Q3 piece called the 73% net new ARR growth a comp illusion. That was correct. The underlying growth rate was approximately 20%, not 73%, and the right baseline was the pre-outage trend rather than the disaster quarter. I was also right that at $517 and 24x forward revenue there was no margin of safety. The 17% decline since December confirms it.

What I missed was the behavioral mechanics of Falcon Flex. I treated the re-Flex statistics as anecdotal color. They were structural evidence. Over 380 customers have returned to renew Flex commitments early, at approximately 50% higher ARR. Nearly 100 have done this multiple times. I knew those numbers and classified them as marketing. I should have classified them as a model change.

I also misread the SIEM silence. I wrote in Q3 that if SIEM had grown to $600 million, Kurtz would have said so. He said so this quarter, $585 million ARR, growing above 75% year-over-year, with implied Q4 net new ARR of roughly $155 million exceeding any prior full-year figure. The silence in Q3 was strategic, not defensive. Management was waiting for a trend rather than a data point. That is a meaningful distinction and I drew the wrong inference.

The correction matters because it changes the thesis, not just the numbers.

The Central Insight: This Is a Data Pipeline

CrowdStrike’s moat is not its detection algorithms. Detection quality is a product advantage, and product advantages erode. Microsoft can hire threat researchers. AI makes baseline detection increasingly commoditized. If the moat were detection, the bear case would be strong and getting stronger.

The actual moat is the data pipeline.

The single-agent sensor on hundreds of millions of endpoints generates telemetry that feeds detection, identity, cloud security, SIEM, and AI-driven response, all from the same data source, in the same data model. Each new module extends this pipeline rather than building a new one. The economics of extension are nearly zero for CrowdStrike. The economics of replication are enormous for a competitor: they would need to redeploy agents across the same environment and then wait years to rebuild behavioral baselines that make the data meaningful.

This is why 97% gross retention held through an event that gave every CISO in the world explicit permission to leave. The switching cost is not “we like the product.” It is “we would lose the forensic history of our environment.” That history does not export. It lives in Falcon.

Here is what makes this moat compound rather than simply persist: every new endpoint Falcon deploys to makes detection better for every other endpoint. Behavioral anomalies are identified across the entire installed base, threats seen in one environment protect all environments. This is not a feature advantage that competitors can hire engineers to match. It is a data advantage that widens as CrowdStrike grows. The more endpoints, the better the detection. The better the detection, the more endpoints. Competitors do not face a product gap. They face a data gap that grows larger every quarter CrowdStrike adds customers.

SIEM makes this architecture visible. Legacy vendors like Splunk built their business around ingesting data from everywhere expensively, charging by volume, effectively taxing customers for the privilege of their own data. CrowdStrike’s argument is that this model is backwards. The most valuable security telemetry already lives inside the Falcon sensor. The security operations center’s data plane does not need to export to a separate system. The separate system is redundant. You already have the data.

At $585 million ARR growing above 75%, with implied Q4 net new ARR exceeding any prior full-year figure, the market is beginning to validate this architecture. Combined platform ARR, Cloud Security, SIEM, Identity, exceeded $1.9 billion, growing above 45% on a $5.25 billion base. Thirty-six percent of the business is growing at twice the company rate. The adjacencies are no longer adjacencies. They are the growth engine, and they are growing faster than the core.

Falcon Flex: The Behavioral Proof

Flex converts enterprise security procurement from episodic budget negotiations into continuous consumption. Customers pre-commit a credit balance, deploy modules by activating them against that balance without triggering new CFO approval cycles, and when the credits run out, return at higher commitments because the modules are now operational and indispensable.

Over 380 customers have renewed early at approximately 50% higher ARR. Nearly 100 have done this multiple times. Flex ARR reached $1.69 billion while non-GAAP subscription gross margin held at 81%, up 100 basis points year-over-year as Flex became a larger share of the base. The bear case that Flex was a discounting mechanism has no data support. Margins expanded as Flex scaled.

What Flex actually does is remove the competitive window. You cannot undercut a deployed module with a cheaper alternative when the deployment decision has already been made and the budget has already been committed. The question is no longer what to buy. It is what to activate next.

The unit economics remain opaque. Management emphasizes behavioral metrics, renewal rates, credit consumption velocity, ARR uplift, over cohort-level detail. This is a choice. Companies disclose what proves their case and omit what complicates it. The margin expansion as Flex scales suggests the economics are favorable; the absence of cohort-level disclosure suggests they are not uniformly favorable across all customer segments. The bet is that aggregate economics dominate the edge cases. That bet looks good today. It requires monitoring.

The metric I want disclosed and isn’t yet: Re-Flex cycle cadence. Whether successive cohorts are burning through credits faster and returning sooner is the earliest leading indicator of flywheel acceleration versus plateau. Ask for it on the Q1 call.

Microsoft: Concession, Not Competition

The standard bear case on Microsoft is structurally correct in one segment and wrong in the one that matters more.

In sub-1,000 endpoint accounts, SMB, lower mid-market, E5 bundling wins. “Good enough plus already paying for it” beats “demonstrably better but requires a separate budget line.” That dynamic is real and probably permanent.

In the enterprise SOC, something different happened this quarter.

CrowdStrike is now purchasable through Microsoft Azure Consumption Commitments. AWS has designated Falcon as the default SIEM in Security Hub. The two largest hyperscalers have both entered structural partnerships with CrowdStrike simultaneously, one giving it access to Azure’s enterprise budget infrastructure, one making it the default security data layer in the world’s largest cloud.

The question worth sitting with: why would Microsoft allow a competitor to ride its consumption commitment framework? The answer is that Microsoft cannot build the telemetry layer. Years of behavioral data across millions of endpoints in specific customer environments is not replicable at any speed or cost that makes strategic sense. By allowing CrowdStrike into the MACC ecosystem, Microsoft is conceding the security data infrastructure layer and capturing the distribution relationship instead. That is not a channel deal. That is a category boundary being drawn.

CrowdStrike is becoming the security layer for Azure, not the alternative to it. The budget infrastructure Microsoft built to consolidate enterprise spend is now a distribution channel for the company Microsoft was supposedly threatening.

The test is Q1 and Q2 FY27 commentary. If CrowdStrike names MACC-attributed mid-market wins, the dynamic is real. If the partnership is never mentioned again, it was a press release.

Variant Perception

The market at $391 has repriced CrowdStrike from momentum compounder to quality at fair value. A record quarter produces essentially no stock reaction because fair value is already embedded in the price. Two things are not embedded.

The normalized margin story. CrowdStrike’s reported FY26 non-GAAP operating margin was 22%. Inside that number: $117.7 million in July 19 incident costs and $44.8 million in strategic plan charges, together about three and a half percentage points of temporary cost that disappears mechanically in FY27. You are not buying a 22% margin business at $391. You are buying a 25-26% margin business carrying costs the income statement is still absorbing. That gap closes in four quarters with no execution required.

The architecture reclassification. ServiceNow spent three years being valued as an ITSM tool before the market reclassified it as an enterprise workflow operating system. The multiple re-rated significantly on that reclassification alone, with no fundamental change in the business. The SIEM number this quarter is the first piece of evidence the market needs for the equivalent reclassification on CrowdStrike, from security product company to security data infrastructure. It is not yet sufficient evidence. One or two more quarters of SIEM confirmation, combined with named hyperscaler-attributed wins, would likely be sufficient.

What the market may be correctly pricing: the SBC structural headwind. Stock-based compensation was $1.097 billion in FY26, 22.8% of revenue, growing faster than revenue, creating a $4.38 per share gap between GAAP and non-GAAP earnings. CrowdStrike is making a deliberate choice to pay for talent in equity rather than cash. That preserves cash flow but dilutes shareholders and creates a widening GAAP loss that institutional screens will increasingly penalize. This is a known trade-off. It is a persistent drag on the multiple independent of fundamental performance.

The regulatory inquiry is the tail risk. Dual-agency DOJ and SEC involvement is categorically different from a routine comment letter. The 10-K filing in March is the first real read on whether the language normalizes or escalates. An ARR methodology disclosure change, even a small one, would be the highest-impact negative catalyst available to this stock right now.

Three Years: Three Honest Scenarios

Bull case: approximately $650, 19% annualized return.

SIEM reaches $1.5 billion ARR through confirmed Splunk and Sentinel displacements rather than greenfield alone. MACC partnership generates named mid-market wins by Q3 FY27. Flex proves margin-neutral or better at 40% of total ARR. Charlotte AI becomes a genuine new pricing tier. The market executes the reclassification, from security software to security data infrastructure, and the multiple re-rates accordingly.

Revenue CAGR FY26-29: 24-25%. FCF margin: 36-38%. Exit multiple: 38-40x FY29 FCF.

Base case: approximately $390-420, essentially zero return.

Platform consolidation continues, SIEM grows but stays a module rather than a disclosed platform pillar until FY28, Flex remains margin-neutral without expanding gross margin, and the multiple holds where it is. The base case is already priced at $391. Buying here in the base scenario produces no return for three years. That is the actual risk, not catastrophic loss but three years of opportunity cost.

Revenue CAGR: 19-20%. FCF margin: 32-34%. Exit multiple: 30-32x FY29 FCF.

Bear case: approximately $230, -17% annualized.

The bear case does not require CrowdStrike to fail at execution. It requires winning the architecture battle and losing the economics.

Here is what that looks like: SIEM becomes the default SOC standard, Microsoft responds by building a comparable telemetry layer into Sentinel over three years, and distributes it to E5 customers at zero marginal cost. CrowdStrike ends up with superior technology in a market that has stopped paying for superiority. The data pipeline is real. The pricing power is not. Meanwhile a regulatory disclosure requires an ARR methodology change, even a 3-4% retroactive revision impairs the growth narrative the entire bull thesis is anchored to, and the multiple compresses from 30x to 22x FCF as the platform premium evaporates.

Revenue CAGR: 14-15%. FCF margin: 28-30%. Exit multiple: 22-25x FY29 FCF. The company survives and generates cash. The stock loses badly.

What To Track

Six questions resolve the thesis over the next 18 months.

• Q1 FY27 net new ARR, May 2026. Above $260 million means FY27 is tracking ahead. Below $230 million means the full-year ARR guide is at risk and Q4 was partially seasonal. This is the single most important data point in the next six months.

• SIEM ARR at Q2 FY27, August 2026. Above $750 million at sustained growth above 60% confirms the second engine. Below $620 million means Q4 was a timed disclosure, not a trend. Named Splunk or Sentinel displacement wins outweigh ten greenfield deployments for thesis purposes. Composition tells you more than the aggregate.

• 10-K regulatory language, March 2026. Compare it word-for-word to the prior filing. Any change in ARR definition, escalation in inquiry description, or new revenue recognition language is the highest-impact near-term negative catalyst.

• MACC attribution in Q1 or Q2 FY27 commentary. Named wins through Azure Marketplace reprice the Microsoft dynamic materially. Silence means press release.

• Non-GAAP subscription gross margin as Flex scales. Above 80% sustained at 35%+ Flex share closes the economics bear case. Below 79% at that scale requires re-examination of the entire Flex thesis.

• Re-Flex cycle cadence, demand this be disclosed. Whether successive cohorts burn through credits faster and return sooner is the earliest leading indicator of flywheel acceleration versus plateau. Not currently disclosed. Ask for it on the Q1 call.

The Updated View

My Q3 view was precise: comp illusion on the NNA headline, no margin of safety at $517. Both correct.

The updated view is equally precise.

At $350, you are paying for a security product company and receiving a data infrastructure company. At $391, you are paying fair price for a security product company and betting the market reclassifies it before the next macro drawdown.

This is a platform option with proof milestones, not a set-and-forget compounder. The milestones, Q1 NNA, SIEM trajectory, 10-K language, arrive within 60 days. Position size should reflect the certainty level, which is meaningful but not complete. The thesis is specific. The falsification conditions are specific. When you know what would make you wrong, you know how much to bet.

Subscribe now

• Shopee’s margins remain low because Sea is deliberately reinvesting growing transaction surplus into infrastructure.

• Cross-business integration: from SPayLater to VIP subscriptions, is creating measurable ecosystem advantages.

• If those advantages persist, Sea’s current valuation assumes a profitability ceiling that likely isn’t real.

“Sea Limited reported fourth-quarter revenue of $6.85 billion, up 38% year-on-year. The company guided for 2026 Shopee GMV growth of approximately 25% with adjusted EBITDA ‘no lower than’ 2025 levels.”, Bloomberg

In November I wrote about Sea Limited’s “Rice Field Advantage”, the argument that years of building logistics infrastructure across rural Southeast Asia and urban Brazil had created an operational moat that asset-light competitors could not replicate. Delivering packages through rice fields and across rivers was a physical reality no algorithm could solve, and Sea had accumulated this capability one village at a time, in a way that could not be purchased or fast-followed.

The question was whether Shopee margins would inflect upward from the 0.6% of GMV level that spooked markets after Q3. They did not.

That non-event is what matters about this quarter, not because it invalidates the thesis, but because it clarifies what kind of company Sea is becoming. The rice field framing was wrong, or rather, incomplete. I described a logistics moat. What Q4 revealed is that the moat is integration density, the number of touchpoints where Sea’s businesses reinforce each other. Logistics was the first and most visible layer. It is not the moat itself. The moat is what happens when you stack logistics, payments, credit, and subscriptions on top of each other in markets where no one else has all four.

The Deliberate Delay

I got the timing wrong. I expected margin inflection to begin in the first half of 2026. Management pushed the timeline to the second half at earliest, possibly 2027. That is a six-to-twelve month delay.

The reason it didn’t arrive is what changes the interpretation. The unit economics improved, meaningfully. Core marketplace revenue grew at nearly twice the rate of GMV. Sea has pricing power it is choosing not to exercise. The take rate keeps climbing. Ad-paying merchants are spending more. The surplus generated per transaction is increasing every quarter.

Management is redirecting all of it into growth.

Chris Feng was remarkably direct on the call. They had guided for 20% Shopee growth at the start of 2025. Over the year, “we realized that there are areas we are able to drive the market to grow, and we also learned that there are different levers that we can pull.” They delivered 27%. Now they want to keep pulling.

The levers are all physical: fulfillment networks, instant delivery fleets, a paid VIP membership, content ecosystem partnerships. Feng described these as having “clear investment cycles”, time-limited, not permanent. He said margins should expand by end of 2026 and that the 2-3% long-term target “is well achievable.”

I believe him because the unit economics confirm it, not because CEO promises are inherently credible. A business whose high-margin revenue grows at twice the rate of its volume is generating margin. It is deploying it elsewhere. The question is whether the deployment is wise, and that depends on what the infrastructure produces once built.

Integration Density

I described the rice field moat as logistics, Sea can deliver where others cannot. That was right but incomplete.

What Q4 showed is that logistics is the foundation for interlocking advantages that create value no individual business could produce alone. The VIP program makes this concrete.

Sea launched a paid membership, free shipping, vouchers, exclusive discounts, that surpassed 7 million subscribers in Q4, more than double the prior quarter. In some markets VIP members contribute over 15% of GMV and spend 30-40% more after joining. Interesting, but not unique. Every e-commerce platform has tried loyalty programs.

The hard part in Southeast Asia is payment failure on renewal. Most consumers don’t have credit cards. Auto-renewal breaks. The renewal rate was 40%, most subscribers churned every month not because they wanted to leave, but because the payment couldn’t process.

Sea routed VIP renewals through SPayLater. The renewal rate in Indonesia went from 40% to 70%.

A standalone e-commerce company cannot replicate this without building its own financial services arm. A standalone fintech cannot replicate it without building its own marketplace. The advantage lives in the connection between two businesses, and it widens as both grow. More subscribers generate more payment data. Better data improves underwriting. Better underwriting enables new products. New products improve retention.

This is integration density. Not the generic “synergy” of investor day slides. A specific, measurable outcome, 30 percentage points of improved renewal, that exists only because two businesses share infrastructure. Every new layer Sea adds creates new connection points with every existing layer. The returns on investment in any single layer accrue across all three businesses simultaneously.

The margin delay is rational if you accept this framing. Sea is not investing in logistics. It is investing in integration density. And the returns compound in ways a single-segment P&L will never capture.

The Monee Question

Monee’s loan book grew 80% to $9.2 billion. NPLs held at 1.1%. Sea shifted from invite-only lending to “all can apply”, opened the doors to everyone, and default rates didn’t move.

The natural question is why. Management points to Shopee’s transaction data feeding AI underwriting models that see things traditional lenders cannot, purchase patterns, return behavior, payment timing, browsing history, processed through transformer-based models trained on long-sequence behavioral data.

Grab has ride-hailing data. GoTo has Tokopedia. Why can’t they build this?

E-commerce transaction data is categorically different from ride-hailing data for credit underwriting. Purchase patterns across thousands of SKUs, return behavior, basket composition, browsing-to-purchase conversion, these are higher-dimensional signals about financial behavior than trip frequency or food delivery orders. Sea has this data at a scale no partnership can replicate, 400 million active buyers, 4 billion quarterly orders, because no partner would share transaction-level data at that granularity.

Permanent? No. Structural for 5-7 years in underdeveloped credit markets? Yes.

What shifts the framing is off-Shopee SPayLater loans, up over 300%, now 15% of the total portfolio. In Malaysia, 30%. Monee is becoming a consumer credit platform with standalone demand, built on a data advantage incubated inside the ecosystem but now being exported beyond it.

In Q3 I described Monee as evidence that the ecosystem worked. Better to say: Monee may be the part of the ecosystem that eventually generates the most value, and the market is still pricing it as a line item.

The Three-Year Question

Brazil and Garena confirm the broader pattern. Brazil is profitable, accelerating into fulfillment, and moving upmarket, the question has shifted from survival to ambition. Garena’s Q4 bookings fell 20% sequentially because Q3 stacked two massive IP events in one quarter; the full-year trend of 37% growth and $1.7 billion EBITDA is what matters for the funding thesis. Neither changed the picture.

The three-year question is simpler and harder: does integration density convert to margin?

If it does, if the infrastructure creates durable advantages that compound and eventually require less incremental investment, then Shopee margins expand as the investment cycles end. Monee keeps compounding on proprietary data. Garena funds the transition. Sea generates $7-8 billion in EBITDA on $40-45 billion in revenue by 2028. That gets you to roughly double the current price. The bull case, Brazil replicating Asia’s margin profile, Monee scaling off-platform to 30-40% of a much larger book, gets meaningfully higher.

If it doesn’t, if competitors find ways around the physical moat, if the investment cycles prove perpetual, if a credit downturn exposes the underwriting as less robust than it appears, then Sea is a fast-growing company that never becomes highly profitable, and the current price is approximately fair.

The base case requires nothing heroic. Shopee margins normalize toward the 2-3% target management has reiterated for two years. Monee keeps doing approximately what it is already doing. That is a bet on continuation, not acceleration.

At 22 times forward earnings on a business growing 30%, markets price Sea as if the margin compression is permanent. The market treats 0.6% as a ceiling. I think it is a waypoint. I held that view three months ago. The waypoint was further away than expected.

The Choice

Markets wanted harvest this quarter. Management looked at what they are building, looked at the levers they discovered over the past year, and decided the fields were not ready. Frustrating for the impatient. Correct for anyone who believes harvest value depends on planting depth.

I am less confident about timing than three months ago. More confident about the payoff itself, because the evidence of what integration density produces, 30 points of improved subscription renewal, stable credit quality at 80% loan growth, a take rate that keeps climbing, got materially stronger.

The open question: is integration density durable, or a temporary lead that erodes as competitors build their own stacks? The next two quarters matter more than any model.

The rice fields are getting bigger. The question is whether they are big enough.

Subscribe now

• The real beat was ARPAC, not revenue. Monetization density jumped from $11 to $15, confirming Nu’s flywheel has shifted from customer growth to profit compounding.

• Provision growth was mechanical, not necessarily deterioration. A 60% increase in unused credit limits front-loaded IFRS 9 losses, while actual NPLs improved. The revenue from that expansion hasn’t shown up yet.

• Nu is trying to change its multiple. A new managerial P&L, platform framing, and U.S. charter progress signal a transition from “high-growth LatAm bank” to “global digital banking platform.”

“Higher costs and provisions overshadow profit.”

— Bloomberg, February 26, 2026

The Quarter the Optics Won

That line explained the stock move, not the quarter.

In my Q3 piece, Building the Moat While Breaking Records, I argued that nuFormer’s reveal marked the shift from UX dominance to decision-quality dominance, and that the 27.7% efficiency ratio reflected deliberate investment, not cost slippage. I closed by writing that the validation quarter was complete and the acceleration quarter had begun.

Q4 proved that right, but in a way the market couldn’t parse cleanly. Revenue beat consensus by 6%. ROE hit a record 33%. Early-stage NPLs improved for the fourth consecutive quarter. ARPAC reached $15. And the stock dropped 5%, because provisions went up.

I think the market traded the optics correctly and the substance backwards. But more importantly, Q4 was the quarter Nu stopped merely reporting results and started arguing for a fundamentally different way of being valued. Management introduced a new Managerial P&L, reframed 2026 as an “inflection year,” and described the company as transitioning from a Latin American leader into a “global digital banking platform.” That is not routine earnings-call language. That is a company trying to change the lens through which investors price it.

Whether it deserves that new lens is the question this quarter is really asking.

Monetization Density Is the Real Story

The headline beat was revenue. The more important beat was monetization density.

For years, the Nu narrative was about customer acquisition, the purple card as rebellion against Brazilian banking’s indignities, the explosive sign-up curves, the crossing of 100 million customers. That chapter is effectively over. Nu added just 4 million customers in Q4, the lowest quarterly add since early 2024.

But look at what is accelerating instead.

That table is the whole story. ARPAC stalled at $11 through 2024, the period when the market was most nervous about the stock. The re-acceleration to $15, a 35% increase on a full-year basis, is the signal that the monetization engine has found a new gear.

Do the simple math: 131 million customers, 83% activity rate, $15 ARPAC. That gets you to roughly $4.9 billion of quarterly revenue, almost exactly what Nu reported. The entire revenue line is now explained by monetization density times installed base. And $15 is still less than 40% of what Brazilian incumbents earn per customer. The gap is not a ceiling. It is the runway.

In The Purple Way, I framed this as the core flywheel: own distribution, deepen engagement, raise ARPAC, let operating leverage compound. Q4 was the quarter that flywheel became visible in a single number.

Of course, ARPAC acceleration from $11 to $15 partly reflects mix shift, more credit products, higher-yielding secured lending, seasonal Q4 effects from Brazil’s thirteenth salary. Whether this pace sustains into Q1, when seasonality works against it, is genuinely uncertain. The number to watch is whether ARPAC holds above $14 in the softer quarter. If it does, the re-acceleration is structural. If it slips back toward $12, the Q4 print was flattered.

The Provision Illusion

Which brings me to what the market actually reacted to, and to a mechanism worth understanding on its own terms before deciding what it means.

When a bank extends a new credit limit to a customer, it does not wait for the customer to use that limit before accounting for the risk. Under IFRS 9, the expected loss on unused credit must be provisioned immediately, the day the limit is granted, before a single transaction occurs. This is front-loaded accounting: the cost appears now, the revenue appears later.

In Q4, Nu’s unused credit limits went from $18 billion to $29 billion, a 60% increase, driven by nuFormer’s deployment in Brazilian credit card underwriting. That $11 billion expansion mechanically required higher provisions. Meanwhile, actual delinquencies moved in the opposite direction: 15-to-90-day NPLs fell 20 basis points to 4.1%, and 90-plus NPLs fell 10 basis points to 6.6%.

On the call, Lago pointed analysts to explanatory note 32 of the financial statements to see the unused limit data. Whether that was deliberate signaling to sophisticated investors or simply a CFO being thorough, the number is worth finding. He then laid out the conversion sequence: limits first, then purchase volume, then interest-bearing balances. “We are starting to see the first step,” he said, referring to a 50-basis-point gain in Brazilian credit card market share, the largest in ten quarters. The NII has not shown up yet. That is two to three quarters away.

My read is that this is a timing mismatch: the market penalized the cost today while the revenue is still ahead. But I want to be honest about the alternative reading, because it is not trivial.

The Risk to the AI Thesis

The alternative is that nuFormer’s limit expansion is too aggressive. A 60% increase in credit availability, even from a sophisticated model, is a large bet. The 3x performance uplift over traditional ML that management cited in Q3 was measured over a limited deployment window; we do not yet know how the model performs through a full credit cycle, or whether the uplift degrades as it encounters edge cases further from its training distribution. The NPL improvement in Q4 was partly seasonal, Brazil’s thirteenth salary helps borrowers catch up on payments every Q4, and management explicitly guided for a seasonal uptick in Q1. If 15-to-90-day NPLs spike meaningfully above 4.5% in Q1, the “AI improves credit” narrative gets harder to sustain.

I still think the provision increase was mechanical, not pathological. But I hold that view with less certainty than the rest of the thesis, because the magnitude of the credit limit expansion is large enough that even a small deterioration in model accuracy would matter.

A New Scoreboard Is Never Neutral

There was something else happening in Q4 that I think matters more than the provision debate. Nu introduced an entirely new reporting framework and restated results back to Q1 2021.

Lago framed it as housekeeping: “This evolution does not change economic reality, it only clarifies it.” That may be true, but companies do not reframe disclosure unless they believe the old framing is obscuring the business they are becoming. Nu also changed its efficiency ratio methodology and began presenting consolidated NPL metrics. The subtext is not subtle: Nu wants to be evaluated as a platform, not as a bank.

From Bank to Proprietary Data Loop

But what kind of platform? This matters for the multiple expansion thesis, and I think it deserves specificity. Nu is not building a Shopify-style infrastructure layer that others build on. It is building something closer to the Amazon model: proprietary aggregation where the company owns the customer relationship, the data loop, and the decisioning layer, while selectively integrating third-party services, NuCel for telecom, NuTravel for bookings, marketplace for commerce, as engagement extensions that feed back into the core underwriting engine. The moat is not that others can plug into Nu. The moat is that every additional service generates behavioral data that makes the credit decisions better, which funds more growth, which generates more data.

That architecture is what justifies a different multiple, if it works. A bank that uses AI is still a bank. A proprietary data loop that uses banking as its first application is a different kind of business. Q4 was the quarter Nu started explicitly making that argument.

Capital as Launchpad

Lago also guided for “four to six quarters” of efficiency pressure from three investment buckets: return-to-office (quantified at 80 to 100 basis points), AI and GPU spending (unquantified), and global expansion (”substantial” and “not capitalized”). After four quarters of accelerating ROE, 27%, 29%, 31%, 33%, I think Lago is setting a floor, not a ceiling. But the guidance serves a practical purpose regardless: it gives management room to invest without every tick in the efficiency ratio triggering a sell-off.

The most important thing he said was not about efficiency. It was this: “Our capital and liquidity positions are not simply a reflection of our past performance. They are, in fact, the foundation of what comes next.” Three billion in unrestricted cash. Two point two billion of excess capital. Available funding at twice the net credit portfolio. That is a company telling you: the investment phase is coming, the capital is in place, and we do not need external financing to execute.

The Zero-Priced Option

In The Purple Way, I described Act 3 as Nu’s emergence as a global financial platform, theoretical at the time. It is no longer theoretical. Nu filed for a U.S. national bank charter in September and received conditional OCC approval in January.

I want to be precise about what “conditional” means. It is not final approval. It means the OCC has determined the application has merit and is imposing conditions, typically around capital, compliance infrastructure, and business plan, that must be satisfied before a full charter is granted. Foreign fintechs have a mixed record navigating this process, and the conditions themselves have not been disclosed. This could take another six to twelve months, or longer.

That said, reaching conditional approval in four months is fast, and it signals regulatory seriousness on both sides. Velez was carefully vague on the call but revealing in his framing: “When you dig in into sub-segments, in certain niches, that, by the way, happen to be the size of Brazil, we actually find opportunity to solve a number of consumer problems.”

Sub-segments the size of Brazil. There are 62 million Hispanic Americans, younger, more digitally native, and more underserved by traditional banking than the broader population. Many have family members in Nu’s existing markets. Nu already has credit data on those family networks. It has nuFormer. It has a brand that carries trust across the Latin American diaspora. And it has a cost structure that makes U.S. neobanks look expensive.

The executive hires, a former Central Bank Governor, a CTO who has scaled to 900 million users, a Chief Design Officer with massive-scale experience, are not consistent with a “de minimis” initiative. But they are also not proof that the U.S. will work. The American market has humbled many sophisticated entrants, including well-funded fintechs with seemingly compelling wedges. I give Nu credit for the approach but not for the outcome. That has to be earned.

The market assigns zero value to U.S. optionality. That is probably wrong, but by how much depends entirely on execution that has not yet occurred.

Scenario Weights: Subtle but Important Shifts

Q4 shifts my probability weights modestly. The bull case, global platform recognition, market re-rating to 25-30x earnings by 2028, moves from 45% to 50%, on the strength of nuFormer’s proven deployment and the U.S. charter progression. Assumptions: revenue CAGR around 30%, ARPAC toward $25-plus, net margins of 18-20%.

The base case, Latin American dominance, 18-22x earnings, moves from 40% to 35%. Assumptions: revenue CAGR around 22%, ARPAC reaches $20-22, Mexico profitable but slower.

The bear case stays at 15%. Assumptions: revenue CAGR around 12%, ARPAC stalls, macro deterioration, multiple compresses to 12-15x.

From roughly $15.50, the probability-weighted outcome implies approximately $32 by 2028. The improvement reflects nuFormer moving from hypothetical to proven, and U.S. optionality moving from aspirational to regulatory.

Five things would change my mind: ARPAC stagnation below $18 for two consecutive quarters. Risk-adjusted NIM compressing below 9% without strategic explanation. Mexico ARPAC failing to reach $18-20 within eighteen months. Efficiency above 26% through 2026 without revenue acceleration. Any U.S. regulatory retreat.

Signal and Noise

The market read Q4 as a noisy bank print. It was. But it was also something else, the quarter the compounding machine revealed itself, briefly, before the noise closed back over it. Whether the reveal sticks depends on whether ARPAC keeps compounding, whether nuFormer’s credit expansion converts to NII without a credit-quality surprise, and whether the platform thesis Nu is now explicitly selling proves to be more than valuation architecture.

The cheapest capital in Latin America is still habitual trust inside an app. And that trust is now compounding with an intelligence layer that traditional banks cannot replicate. I think the thesis strengthened this quarter. I also think it has never been more important to watch the variables that could prove it wrong.

Subscribe now

• Financial strength is undeniable. 68.6% gross margin and accelerating growth suggest pricing power and embedded value.

• The fear has a name: Co-Packaged Optics. The market is discounting a future where external connectivity commoditizes.

• This is no longer a copper thesis. It’s a question of whether managed reliability survives architectural transition.

On the evening of March 2nd, Credo Technology reported the strongest quarter in its history. Revenue hit $407 million, up 52% sequentially and more than 200% year, over, year. Non, GAAP gross margins reached 68.6%. Operating margins crossed 49%. Free cash flow nearly quadrupled. Management guided fiscal 2027 to grow more than 50%, above what the Street expected.

The stock fell 12%.

The market is not saying these numbers aren’t worth much. It is saying these numbers don’t mean what you think they mean. It has categorized Credo as a copper connectivity vendor riding an AI capex wave, and it is pricing the inevitable end of that wave. The market may be wrong about the category. But the error is not irrational. It is structural, rooted in a specific fear about where the data center architecture is heading, and whether Credo’s value survives the journey.

What We Argued Before

In our previous analysis, we made a specific claim: that Credo was misunderstood. The surface, level story, a company selling Active Electrical Cables into AI data centers, obscured a deeper one. Credo’s margins were too high and too durable for a commodity hardware supplier. Its customer relationships were too sticky. Its product architecture, embedding real, time telemetry into every deployment via the Pilot software platform, suggested the beginnings of operational lock, in, not just a better product.

Our view was that Credo looked like a component vendor on the surface but behaved economically like infrastructure. Q3 didn’t settle that question. But it made the evidence harder to dismiss, and the counter, arguments more urgent.

The Fear

To understand why the stock fell on a great quarter, you have to understand the fear. It has a name: Co-Packaged Optics (CPO).

The physics are real. As AI clusters push toward 1.6 terabit ports and 200 gigabits per lane, the effective reach of copper contracts. Not to irrelevance, Credo’s active cables still handle five to seven meters, covering most intra, rack and many rack, to, rack connections, but the directional pressure is undeniable. At some threshold of bandwidth and distance, light replaces electrons.

NVIDIA knows this. On the same day Credo reported, the market was absorbing NVIDIA’s investments in Lumentum, Coherent, and other optical players. When the company whose GPU architectures define AI infrastructure puts capital behind optical suppliers, it signals architectural intent. Not next quarter. But within the planning horizon that determines valuation multiples.

The market’s logic runs cleanly: NVIDIA integrates optics into switch silicon, external transceivers and cables commoditize, Credo’s AECs become a melting ice cube regardless of current demand, compress the multiple now. This logic is coherent. It is also incomplete, in a way that reveals how markets think about technology transitions and where they tend to get the timing wrong.

What Q3 Actually Showed

The quarter’s most important development was not the revenue number, which had been pre, announced in February and carried no surprise. It was the six, month pull , forward of ZeroFlap optical transceivers into production, with TensorWave as the first customer and three hyperscalers in qualification.

This reframes the copper, versus, optics debate. Credo is not defending copper against optical disruption. It is attempting to port its reliability architecture across physical media. ZeroFlap transceivers embed the same Pilot telemetry, continuous link health monitoring, autonomous signal degradation detection, switch, level SDK integration, that made AECs the standard for short, reach connectivity. The value proposition was never “copper is superior to light.” It was “managed reliability is superior to commodity reliability, regardless of the medium.”

There is a tension here worth holding. The six, month acceleration is either evidence of product, market fit arriving faster than expected, or evidence that the optical window is closing faster than management originally planned. It is probably both. Customers are pulling ZeroFlap forward because they genuinely need reliable optical at scale. And Credo is willing to ship ahead of schedule because waiting risks ceding the optical transition to incumbents with deeper resources. Strength and urgency are not mutually exclusive, and the distinction matters less than whether the product actually works at hyperscaler scale. That we will learn in the next two quarters.

CEO Bill Brennan was direct about the competitive narrative on the call. He characterized CPO discourse as a “signal to noise ratio issue” and argued that co, packaged optics would remain limited until delivering “bulletproof reliability.” The dismissive tone may have cost him with investors. But the product roadmap told a more considered story: ZeroFlap for pluggable optical, Active LED Cables for mid, reach distances up to 30 meters, and an OmniConnect gearbox targeting near, package optics with MicroLED, designed explicitly to address “the reliability, serviceability, and availability pitfalls of current CPO solutions.”

The same quarter saw the quiet acquisition of Comera, a company specializing in Layer 2 protocol IP, error correction, and security. Asked whether this represented a move up the networking stack, Brennan’s answer was simply: “Yes, absolutely.” This does not prove Credo is becoming a platform company. But companies that intend to remain cable suppliers do not typically acquire protocol IP on the day of their strongest earnings print.

The Margin Question

The 68.6% gross margin deserves more scrutiny than the market gave it, in both directions.

The bull case is straightforward: in a quarter where revenue grew 52% sequentially, shipped overwhelmingly to hyperscalers who negotiate aggressively on every line item, margins did not compress. They expanded. By 92 basis points. This is not how commodity hardware behaves during a demand surge. It suggests customers are paying for something embedded in the product, reliability, telemetry, operational intelligence, that goes beyond the physical cable.

Management guided Q4 at 64–66%, roughly 300 basis points below Q3 actuals. The market read this as compression. The historical pattern suggests it is more likely conservative forecasting, management guided Q3 at 64–66% and delivered 68.6%, and the same pattern has repeated for four consecutive quarters.

But there is a version of the bear case that is more interesting than “management is finally right about its own margins.” As Credo’s revenue mix shifts toward optical, ZeroFlap transceivers, eventually ALCs, the blended gross margin may genuinely settle closer to 64–66%. Optical products carry higher component costs, early yield ramp friction, and potentially introductory pricing to win qualifications. If that is what the guide reflects, then 68.6% was not a platform baseline but a copper peak, and the long, term margin is what management has been saying all along: 63–65%. This would not break the thesis, but it would change what the thesis is worth. The difference between a 68% margin business and a 64% margin business, compounded over several years, is meaningful.

The Hardest Question

The debate around Credo has evolved. It is no longer copper versus optics. The more interesting framing is commodity connectivity versus reliability, managed connectivity. Credo is betting that hyperscalers will pay a structural premium for managed reliability, observable, predictable, autonomously maintained link health, across any physical medium.

But there is a harder question underneath this one, and it is the question the thesis ultimately depends on: does reliability remain a separable layer that Credo can own?

The historical pattern in hyperscaler infrastructure is not encouraging for vendors. Google built Andromeda to replace Cisco. Amazon built Nitro to replace Intel networking. The playbook is consistent: adopt the best external solution, learn from it, then build or commoditize internally. If Pilot’s telemetry value can be replicated at the switch OS level, through SONiC extensions, NVIDIA’s management stack, or a hyperscaler’s own monitoring tools, then Credo’s premium erodes regardless of whether it ships copper or optical.

The counterargument is that physical, layer telemetry is structurally harder to replicate than switch, level software. Pilot operates inside the cable’s DSP, reading eye diagrams, pre, cursor equalization data, and signal integrity measurements that do not exist at the switch level. To build an internal alternative, a hyperscaler would need to design its own SerDes, develop its own firmware, build its own cable assembly and qualification process, and accumulate years of deployment data across millions of links. That is not writing monitoring software. That is building a semiconductor company. The barrier is higher than what Google faced replacing Cisco.

Higher, however, is not insurmountable. And the 68.6% gross margin, while powerful evidence that the reliability premium is being paid today, does not guarantee that the payment continues to be routed through Credo as the architecture evolves. This uncertainty is real. The thesis must be honest about it.

Updating the View

Our prior thesis was that Credo was probably transitioning from hardware to infrastructure. After Q3, we believe it is likely, with a more specific understanding of what “likely” requires.

Three of four key debates moved in favor of the bull case. Customer concentration continued improving: the largest customer fell to 39% of revenue, down from 86% a year ago. A fifth hyperscaler was confirmed. ZeroFlap optics are opening an entirely new neocloud customer channel. Growth sustainability was confirmed above expectations, with management guiding fiscal 2027 above 50%. And the optical pivot is no longer theoretical, production units are shipping.

Margin durability remains open. Not because Q3 was weak, but because the product mix is changing and the steady, state margin of a blended copper, optical portfolio is genuinely unknown.

The conceptual upgrade in our thinking: the variant perception on Credo is no longer “AECs will keep winning.” It is that Credo’s reliability advantage may travel further than copper. If ZeroFlap optics gain traction and Pilot telemetry becomes embedded in hyperscaler operational workflows, the total addressable market expands with each architectural transition rather than contracting. If reliability gets commoditized or absorbed into the vertical stack, Credo remains a very good but ultimately cyclical hardware company. That is the distribution of outcomes the stock is pricing, and at 24 times forward earnings with 50% growth, it is pricing closer to the pessimistic end.

Three, Year Framework

The right way to think about Credo is not a single outcome but a range tied to whether its reliability advantage proves portable.

Bull case (35% probability): $350–400. Revenue compounds at 45–50% through fiscal 2029, reaching approximately $4.5 billion. ZeroFlap optics and ALCs become material revenue lines. Gross margins sustain in the mid, to, high 60s as Pilot, embedded products command structural premiums. The market re, rates Credo as infrastructure rather than hardware. This case requires not just execution but sustained AI capex, CPO delays, no macro recession, and, critically, Pilot proving durable against internal hyperscaler alternatives. Many things must go right. The reward if they do is substantial.

Base case (50% probability): $180–210. Revenue compounds at 35–38%, reaching $3–3.5 billion. AECs remain dominant, optics becomes additive but not transformative. Gross margins settle at 64–66%, management’s own target, possibly reflecting an honest blended margin rather than conservatism. Pilot is strategically valuable but never separately monetized. The market values Credo as a premium semiconductor company.

Bear case (15% probability): $70–85. Revenue growth decelerates to 20–22% as AI capex normalizes and CPO timelines compress. Gross margins drift toward 58–62% under hyperscaler pricing pressure. Pilot never achieves visible lock, in. The market treats Credo as a cyclical component supplier. This requires multiple simultaneous failures but cannot be dismissed.

What to Track

Six signposts will determine whether the thesis is evolving correctly.

• ZeroFlap optical revenue. Optics must become real in the income statement, not just the press release. Exiting fiscal 2027 above 15% of revenue would be meaningful confirmation.

• Gross margin through the mix transition. Can Credo sustain 64%+ as optical grows? Another quarter of beating a conservative guide would ease the concern. Falling to guide would suggest the blended margin reality is settling in.

• Customer breadth. Does the fifth hyperscaler scale? Does the neocloud channel expand beyond TensorWave?

• Pilot evidence. Any concrete disclosure, attach rates, SDK adoption, operational workflow references, would begin to close the separability question.

• Hyperscaler capex tone. April and May earnings from Microsoft, Amazon, Meta, and Google will signal whether AI infrastructure spending is sustaining or peaking.

• CPO progress. Broadcom and TSMC commentary on silicon photonics yields. Current yields remain below production thresholds, but the gap is narrowing faster than copper bulls expected. Delays extend Credo’s runway. Acceleration compresses it.

Where We Stand

Our view is stronger on the company and more specific about what the thesis requires. Q3 made it harder to argue that Credo is just a well, timed hardware vendor. But it also clarified what still needs proving: not whether Credo can win in copper, that is settled, but whether its reliability advantage remains ownable as the architecture shifts from copper to optical and as hyperscalers do what hyperscalers always do, which is try to commoditize their suppliers.

The stock at $99 already assumes the market’s answer to that question is no. If the answer turns out to be even partially yes, the re, pricing will be significant. The next twelve months will determine which side of that bet is right.

Subscribe now

• Zscaler delivered its strongest quarter in two years (26% growth, Rule of 62), but shares fell ~10% as investors fixated on organic guide math.

• The real story: Z-Flex lock-in and consumption pricing are quietly reshaping the growth algorithm beneath headline deceleration.

• At ~7x revenue, the stock reflects a maturing SaaS multiple, not a security infrastructure platform compounding backlog and traffic.

“It’s not clear why the stock is getting hit on the report other than people were ‘hoping for more.’ The results are ‘fundamentally fine’ but likely ‘not good enough.’”

— Vital Knowledge, February 26, 2026

I have been writing positively about Zscaler’s architectural advantages for several quarters now, the “Hotel California” lock-in of Z-Flex, the positioning as the enforcement layer for agentic AI traffic, the RPO backlog building far ahead of recognized revenue. And for several quarters, the market has violently disagreed with me. The stock has walked down a staircase from $337 to $152, a 55% drawdown, while the business improved on nearly every metric that matters.

In public markets, being early is indistinguishable from being wrong. So it is worth asking directly: am I just wrong? To answer that, I need the right framework. And the right framework isn’t another security company.

Subscribe now

We’ve Seen This Movie Before

In January 2016, ServiceNow traded at $62, roughly 5x forward revenue, down 40% from its 2014 highs. Revenue growth was decelerating from 50%+ to the low-30s. The narrative was that cloud software was “maturing.” Amazon and Microsoft were bundling competitive features into their own stacks. Investors argued that growth SaaS companies were structurally overvalued and that multiples needed to permanently reset. The sector was in a penalty box.

ServiceNow’s business, meanwhile, was doing something the stock didn’t reflect. Net retention rates were expanding. Enterprise deal sizes were growing. The company was quietly becoming the operating system for IT workflows, not because the product was irreplaceable on day one, but because after three years of embedding into procurement, HR, and security workflows, ripping it out became a multi-quarter, multi-team project nobody would authorize. Switching costs were compounding invisibly while the stock went sideways.

By 2019, ServiceNow traded at $280. It quintupled from its penalty box low. Growth didn’t re-accelerate to 50%, it stabilized in the mid-20s. The multiple re-rated because the market belatedly recognized that durable 25% growth with expanding margins and infrastructure-level switching costs was worth a premium, not a discount. The business hadn’t changed. The market’s framework for evaluating it had.

I bring this up because the parallel to Zscaler today is striking. AI is expanding the enterprise attack surface the same way cloud expanded the enterprise footprint a decade ago. The fear that “AI replaces security” is equivalent to the fear that “cloud replaces enterprise software”, directionally wrong, but correct in the sense that the architecture changes. Zscaler is betting that its inline switchboard architecture is the right one for the AI era, just as it was the right one for the cloud era. The stock is priced as if growth is structurally capped and architectural risk is rising.

There is a real risk embedded in this analogy, and I want to be explicit about it. The structural question, raised by several thoughtful analysts, is whether the enforcement layer migrates away from the cloud proxy and toward the endpoint, browser, or identity layer. If Microsoft Entra and Edge become the enforcement point for most enterprise traffic, Zscaler’s cloud proxy could become redundant middleware rather than essential infrastructure. Zscaler’s SquareX acquisition and Entra partnership are defensive moves against exactly this risk. It is non-zero, and it represents the upper bound on conviction. But to judge whether Zscaler is actually losing that architectural war, or simply suffering through a noisy model transition, we need to look at what they just reported.

Why SaaS Is in the Penalty Box

First, let’s acknowledge what’s actually driving the pain, because most of it has nothing to do with Zscaler.

In 2020–2021, a 25%-growth SaaS company was a rare asset in a zero-rate world. Capital paid 15–20x revenue because there was nowhere else to find durable, high-margin growth. In 2026, that same profile competes for capital against NVIDIA, power infrastructure, AI compute, and data center buildouts, categories with more dramatic near-term revenue acceleration. SaaS isn’t broken. It’s no longer scarce.

Three forces are compressing the cybersecurity cohort specifically. First, AI narrative whiplash, every new AI capability (Anthropic’s security tools on February 23, Microsoft Copilot features) triggers a reflexive selloff, conflating “AI changes security architecture” (true) with “AI replaces security spending” (false). Zscaler fell 11% on February 23 alone, three days before earnings, on an announcement that had nothing to do with inline enforcement. Second, capital rotation. These are portfolio rebalancing decisions, not fundamental calls. Third, the “good enough” bundling fear, Microsoft is packaging security into E5 licenses, Palo Alto is pitching SASE consolidation. The narrative, not yet the reality, is that baseline security becomes table stakes, commoditizing the low end of what Zscaler sells.

This is the environment into which Zscaler reported Q2. Every metric beat. Every guide raised. The stock dropped 10%. The penalty box held.

The Quarter: Beats, Raises, and a Red Canary

Here are the numbers, stripped of narrative:

That is a clean beat across every line. Revenue growth re-accelerated from 21% in Q4 FY25 to 26% in both Q1 and Q2, a genuinely rare inflection for a $3B+ ARR company. So why the 10% haircut?

Because Wall Street did the math on the guidance raise and didn’t like what it found.

Management raised full-year FY26 revenue guidance by ~$24 million at the midpoint. Solid. But buried in the commentary, CFO Kevin Rubin noted they were raising the expected revenue contribution from Red Canary, their MDR acquisition, by $35 million, from $90 million to $125 million. If total guidance goes up by $24 million but the acquisition goes up by $35 million, the organic revenue guide arguably contracted. They didn’t even pass through the $17 million Q2 beat to the organic full-year outlook.

Rubin provided air cover, noting that organic net new ARR growth guidance was actually raised from 6.7% to 9.5% for FY26, a real improvement. He also flagged that Red Canary’s post-acquisition churn has been “elevated.” Both statements are true. But by the time this nuance landed, 45 minutes into the call, the stock was already down 8%. The fast-money accounts had done the subtraction, hit sell, and moved on.

This is the credibility paradox of having the most conservative CFO in enterprise software. Rubin has beaten his own revenue guidance for 29 consecutive quarters, averaging ~$20 million above the high end. The beat is structurally priced in before the number hits. All that remains is the guide, and the guide drew a deceleration path: 26% → 23% in Q3 → ~21% implied in Q4. Whether that’s real deceleration or Rubin’s 29th consecutive sandbagged setup is the entire investment debate.

The Variant Perception: Pricing the Wrong Company

Here is what I think the market is getting wrong, stated as clearly as I can:

Consensus sees a decelerating SaaS company growing in the low-20s heading toward high-teens, worth 6–8x revenue. The correct framing is a business mid-transition from per-seat SaaS to consumption-based security infrastructure, with its fastest-growing revenue stream, 100%+ YoY, invisible in every consensus model I’ve reviewed. The implication is that the market is valuing the legacy pricing model while the new model compounds underneath.

Three pieces of evidence from Q2:

The lock-in is reaching critical mass. Z-Flex, Zscaler’s flexible platform licensing program, generated $290 million in total contract value in Q2, up 66% sequentially, roughly 4.5x the level three quarters ago, and now over 30% of all RPO bookings. Average deal: 8-figure TCV. Average term: 4 years. Chaudhry shared that customers on average triple their initial purchase within four years. A Fortune 500 retailer landed as a brand-new customer with 11 modules in a single 5-year commitment. This is the Hotel California at scale, once 16 modules are woven into your security architecture, ripping them out is an 18-month project no CISO will authorize. And it explains the RPO-revenue gap: Z-Flex TCV hits backlog immediately but rolls into recognized ARR over 6–12 months as modules activate. The market watches the speedometer. The engine is building torque.

The business model is shifting underneath the P&L. Non-seat-based, consumption-priced solutions, Zero Trust Branch (per device), Zero Trust Cloud (per traffic volume), AI Security (per token), delivered over 25% of new ACV in Q2, with related ARR growing 100%+. This changes the growth algorithm fundamentally. In a seat-based model, revenue scales with headcount. In a consumption model, revenue scales with traffic. Chaudhry was direct on the call: AI agents don’t have seats, but they generate massive traffic, and every agent-to-application communication routes through the Zero Trust Exchange. They’re processing millions of MCP requests monthly, up from zero two quarters ago. If even a fraction of enterprise agentic traffic routes through the Exchange by default, the consensus revenue models, which assume steady deceleration to ~19% by FY28, are pricing the wrong company.

Management’s disclosure choices tell you where the confidence is. I track what companies introduce, drop, and maintain across earnings reports. In Q2, Rubin volunteered for the first time that core ZIA/ZPA is growing “mid-teens” on its ~$2 billion base, pre-empting a bear concern about core erosion that he knew was suppressing the multiple. He introduced non-seat revenue as a metric because it’s accelerating. But he dropped the specific AI Security ARR number that was $400 million in Q1, restructuring the narrative into two sub-pillars instead. Companies spotlight metrics that are beating expectations and restructure metrics that are merely tracking. AI Security is likely in the $430–460M range: solid, but not accelerating fast enough for a headline. The metric management wants you watching is Z-Flex TCV. They’re right to.

The Three-Year Math

Frameworks are useful. Numbers are better. Here are three scenarios with explicit assumptions:

Probability-weighted expected value: ~$213, or roughly 40% upside from here.

The asymmetry matters. The bear case requires active fundamental deterioration, growth fading to high-teens and multiple compression, not just “more of the same, slower.” The base case requires nothing heroic: consensus execution plus a modest re-rating from 6.4x to 8x as $876 million in FY26 free cash flow becomes undeniable. The bull case needs the consumption model to scale and the market to re-rate Zscaler from “decelerating SaaS” to “security infrastructure.” The upside/downside skew is roughly 2.3x, you risk $27 of downside for ~$61 of expected upside.

One data point for context: comparable 20%+ growers with 25%+ FCF margins, trade at 10–15x NTM revenue. Zscaler at 7x. For that discount to be justified, you need to believe the business is structurally inferior to its peers. Nothing in Q2 supports that conclusion.

What Confirms the Thesis, and What Kills It

Theses are only useful if they’re falsifiable. Here’s what I’m tracking and what would change my mind.

Confirmation signals over the next two quarters:

Z-Flex TCV crossing $350 million in Q3, which would put it at 35–40% of total bookings, the point at which revenue visibility starts resembling infrastructure durability rather than SaaS. Non-seat ACV mix holding at 25% or rising toward 30%, confirming the consumption transition isn’t a one-quarter artifact. The RPO-revenue gap persisting at 5+ points, meaning the backlog is still building ahead of the P&L. Organic net new ARR delivering on the 9.5% full-year guide, or better, beating it, which would confirm Rubin’s conservatism. And large deal counts sustaining the 2x year-over-year pace in the Americas, confirming enterprise consolidation onto the Zscaler platform.

Invalidation signals, what kills the thesis:

Revenue growth drifting below 20% without a clear seasonal explanation. RPO growth decelerating to match revenue, which would mean the backlog is being consumed, not built. Z-Flex penetration stalling at 30% of bookings for two or more consecutive quarters. Non-seat ACV mix falling back below 20%. Gross margins compressing below 79% from hardware cost inflation that management flagged proactively this quarter. Or core ZIA/ZPA growth dropping from mid-teens to single digits, the one signal that would indicate genuine competitive displacement.

I want to be precise about this: if three or more of those invalidation signals trigger simultaneously, the 2016 analogy breaks. That would mean the penalty box isn’t temporary multiple compression, it’s the market correctly pricing a business losing its architectural moat. At that point, 6.4x isn’t cheap. It’s fair.

Early, Not Wrong

In 2016, ServiceNow went sideways for 18 months while the business compounded underneath. The re-rate came not from a single catalyst quarter, but from the market gradually recognizing that switching costs had made the business more durable than the multiple implied. By the time the stock reflected the reality, the easy money had already been made by the investors who bought into the penalty box.

Zscaler’s Z-Flex is building the same compounding lock-in in real time, $650 million in cumulative commitments, customers tripling their spend, consumption pricing expanding the addressable market beyond headcount. The Q2 print didn’t break the company out of the penalty box. No single quarter will. Penalty boxes break on regime shifts, either the company forces the market to re-rate with undeniable acceleration, or capital flows rotate back toward software as the next leg of the AI trade.

The penalty box is real. The pain is real. The timing is unknowable. But the pattern is identifiable: the business has strengthened faster than the stock has appreciated, but not yet fast enough to force a regime change. Q2 moves this closer to “early” than “wrong.”

In 2016, that distinction was worth a 5x return for those with the patience to hold through the discomfort. The price of admission then, as now, was the willingness to own a thesis the market wasn’t ready to validate.

Subscribe now