• Bloom is not an AI pivot; it is a 25-year-old edge-power thesis finally meeting the right customer problem. AI data centers do not merely need cheaper electricity: they need power in months, not years, and Bloom’s onsite, modular fuel-cell architecture is unusually well matched to that job.

• The business inflection is real, but the revenue pipe is messy. Q1 2026 revenue grew 130% YoY and guidance was raised sharply, yet roughly half of Q1 revenue and nearly three-quarters of Q4 2025 revenue came from related-party structures, making conversion quality the key analytical issue.

• The stock now needs the Boulton Ending. Demand is visible; the next proof is operating megawatts, cleaner non-related-party customer wins, service margins holding, cash conversion improving, and Fremont scaling toward 5 GW without quality or margin breaks.

In March 1776, James Boswell visited the Soho Manufactory outside Birmingham. His host was Matthew Boulton, the manufacturer who had entered partnership with James Watt. Boulton’s introduction became one of the cleanest sentences in industrial history: “I sell here, Sir, what all the world desires to have, POWER.”

The quote is famous; the business arrangement was more interesting. Boulton and Watt did not simply sell engines. They charged customers based on the coal their engine saved versus the older Newcomen machine. The price was not anchored to the cost of iron, labor, or assembly. It was anchored to the value of avoided waste.

The second detail matters even more. Watt’s engine was not originally built for factories. It was built to pump water out of mines. The factory use case arrived later, once manufacturers realized that steam could free production from the tyranny of falling water. A mill no longer needed to sit by a river. It could sit near workers, suppliers, and customers. Power moved to where production wanted to be.

That is the useful frame for Bloom Energy. The company is not interesting because it makes fuel cells. It is interesting because AI data centers are discovering that the grid has become the river. And Bloom’s claim is that its machines let the factory leave the river again.

The Rocket Scientist’s Long Wait

Bloom’s founder, KR Sridhar, likes to trace the company’s origin to Mars. That story can sound theatrical, but it is analytically useful. Mars has no grid, no repair crew, no local utility, and no tolerance for failure. A machine designed for that environment begins with different assumptions from a machine designed for ordinary interconnection.

In a 2026 founder interview, Sridhar said his 2001 Kleiner Perkins pitch deck already showed a Bloom box powering a data center, using waste heat for cooling, and connected to nothing else. That is the founder’s cleanest version of the story: the AI data-center use case was there from the beginning; the market simply took twenty-five years to arrive.

The less clean version is just as important. Bloom did not spend those twenty-five years marching in a straight line. It spent them surviving. The company began as Ion America, became Bloom Energy in 2006, came out of stealth with enormous fanfare, and then endured the usual hard-tech middle: high costs, complicated financing, persistent losses, service issues, customer concentration, and the constant question of whether a technically impressive machine could become a financially attractive business.

Sridhar’s Andy Grove story captures the company’s scar tissue. The first few Bloom boxes, hand-built by engineers, worked. The next batches, built from written instructions, failed badly in the field. Grove told Sridhar to stop studying binders and start talking to the people on the factory floor. The lesson was not “manufacturing matters,” a cliché; it was that manufacturing knowledge lives in repeated practice, tacit skill, and feedback from failure.

That matters because Bloom’s edge, if it has one, is not just ceramic chemistry. It is twenty years of learning how to make the chemistry in volume, keep it running, monitor it remotely, service it, finance it, and convince conservative customers that it will not fail.

The Wilderness, Honestly Told

Founder stories become dangerous when they erase the detours. Bloom had many.

In 2021 and 2022, the story was broad: hydrogen, electrolyzers, biogas, marine, carbon capture, Korea, resilience, sustainability, and cost predictability. The South Korean relationship with SK ecoplant was central. Hydrogen was described as a major future market. Marine looked like a possible new vertical. The company had a long list of ways the same core technology might matter.

Some of those stories faded. Marine largely disappeared from the earnings-call narrative. Electrolyzers remained strategically interesting but not central to near-term revenue. Korea remained important, but policy changes and amended purchase timing made it less of a straight-line growth engine. Service economics, for years a drag on margins, had to be repaired.

The company’s historical accounting scar also belongs here. Bloom’s 2019 filing disclosed that prior financials should no longer be relied upon because certain Managed Services revenue had been recognized upfront when it should have been recognized over the contract duration. The correction reduced cumulative revenue through September 2019 by $192.1 million and revealed a material weakness in internal control over financial reporting.

That does not make today’s story invalid. It makes one thing clear: Bloom’s financial plumbing has never been simple. Any investment case that skips that fact is incomplete.

The New River

The old grid was built for a different demand curve. Large power plants, transmission lines, local distribution, interconnection studies, regulated returns, rate cases, and multi-year construction schedules made sense in a world where electricity demand grew slowly and predictably.

AI is a different customer. A large AI data center is a factory whose input is electricity and whose output is compute. The customer does not want an academic answer to the question of whether the grid can eventually serve the load. The customer wants to know whether power will be available before the revenue opportunity moves on.

This is what changed for Bloom. The old question was: “Can Bloom produce a cleaner or more resilient electron at an acceptable cost?” The new question is: “Can Bloom bring power online before the grid, turbines, permitting, and interconnection queues can?”

Bloom’s Q1 2026 results showed why the change matters. Revenue was $751.1 million, up 130.4% year over year, driven by product revenue growth of 208.4%. Gross margin was 30.0% on a GAAP basis and 31.5% on a non-GAAP basis. Operating income was $72.2 million on a GAAP basis and $129.7 million on a non-GAAP basis. Cash flow from operations was $73.6 million.

The company then raised full-year 2026 guidance to $3.4–$3.8 billion of revenue, roughly 34% non-GAAP gross margin, $600–$750 million of non-GAAP operating income, and $1.85–$2.25 of non-GAAP EPS.

This is not just a narrative inflection anymore. It is hitting the P&L.

The Machine for This Age

Grant that the demand is real: why this machine?

The answer starts with time. Bloom’s systems are modular, can be placed onsite, and do not require the same years-long route as new transmission or conventional power equipment. They use fuel, mostly natural gas today, and convert it into electricity through an electrochemical process rather than combustion. That gives Bloom a local-emissions, water-use, noise, and siting argument versus many combustion-based alternatives.

The second answer is reliability design. Bloom’s systems are built from many smaller power modules. That matters because an AI data center is also built from redundancy. A single turbine is a large point of failure; a field of modules can be maintained in pieces. That does not eliminate all risk, but it aligns with how data centers already think.

The third answer is load behavior. AI loads are not always smooth. GPU demand can ramp and pulse. Bloom argues that its solid-state systems can respond faster than rotating machinery and can support newer data-center electrical designs. In Q4 2025, management said every Bloom server shipped from that point would be 800V DC ready, framing the machine as better aligned with high-density AI rack design.

The fourth answer is learning. Product margin remains robust even as volume grows. In Q1 2026, product revenue was $653.3 million, product non-GAAP gross margin was 35.3%, and service non-GAAP gross margin improved to 18.0% from 4.8% a year earlier.

That is the financial fingerprint investors should care about: not just growth, but growth without immediate margin collapse.

Oracle, AEP, Brookfield: Three Kinds of Proof

The proof now comes in three layers.

Oracle is technical proof. The expanded Oracle agreement covers up to 2.8 GW, with Project Jupiter accounting for up to 2.45 GW of Bloom-only power. The Q1 2026 call described Project Jupiter as replacing previously planned gas turbines and backup diesel generators with a Bloom-only power block.

That is a huge validation. It is also concentration risk. If one campus represents most of the headline Oracle ceiling, then permitting, gas availability, site construction, and customer schedule at that location matter enormously.

AEP is channel proof. Bloom’s 2025 10-K says AEP is expected to procure up to 1 GW of Bloom solid oxide fuel cells under a strategic supply agreement, beginning with an initial 100 MW order, and that Bloom began working with AEP in 2025 as both channel and financing partner.

This matters because a utility buying Bloom to serve large loads is structurally different from an enterprise customer buying Bloom to bypass the utility. The incumbent is hiring the substitute because it cannot serve the new load fast enough.

Brookfield is financing proof. The Brookfield arrangement matters not because it is revenue, it is not, but because it can reduce the capital bottleneck for customers that want power availability without owning the equipment themselves. Bloom’s filings show this relationship is already material. The strategic value is real. So is the accounting complexity.

The Loop That Could Matter

The self-reinforcing mechanism is straightforward:

This is slower and more fragile than software. It requires equipment, permits, gas access, construction, service, and cash collection. But in physical markets, trust compounds. The first large installation is hard. The fifth is easier. The fiftieth is a reference case.

Service is part of this loop. The installed base becomes valuable only if service stays profitable as the fleet grows and ages. Bloom has made progress. Q1 service revenue was $61.9 million, service GAAP gross margin was 13.3%, and service non-GAAP gross margin was 18.0%.

But this is not a software-like royalty. It is a maintenance obligation. It includes field labor, replacement modules, repair and overhaul, performance guarantees, and fleet degradation risk. The service book becomes an asset only if reliability improves faster than the fleet ages.

That is why the next phase is not about demand alone. It is about whether the operating proof compounds faster than the physical complexity.

The Pipe Matters

This is where the Bloom story becomes more interesting than the usual AI-power stock pitch.

The demand is real. The reported growth is real. But the path from end demand to reported revenue is not simple.

In Q1 2026, Bloom reported $373.3 million of related-party revenue on $751.1 million of total revenue. In Q4 2025, related-party revenue was $574.2 million on $777.7 million of total revenue. That means roughly half of Q1 revenue and nearly three-quarters of Q4 revenue flowed through related-party structures.

That is not a footnote. It is one of the central analytical facts in the story.

The balance sheet adds nuance. Related-party accounts receivable fell from $151.9 million at year-end 2025 to $0.6 million at March 31, 2026, which is a positive collection signal. But related-party contract assets rose from $3.0 million to $74.1 million, and the company disclosed related-party investments in Brookfield joint ventures and deferred profit from transactions with those unconsolidated affiliates.

The cash-flow statement also shows why investors should avoid treating Q1 as mature cash generation. Operating cash flow was positive at $73.6 million, but inventories consumed $88.6 million, contract assets consumed $64.7 million, prepaid and other current assets consumed $54.2 million, while deferred revenue and customer deposits added $89.5 million.

Again, none of this means the demand is fake. It means the customer proof needs to get cleaner over time.

The right sentence is simple:

The cleaner the customer proof becomes, the more valuable Bloom becomes.

If growth increasingly comes through direct or clearly attributable end-customer demand, with operating assets and cash collection, the market will have more reason to pay a premium. If growth remains dominated by related-party vehicles, unconsolidated affiliates, deferred profit, warrants, and working-capital movements, investors should discount the headline numbers.

What the Market Still Misses

The obvious view is that Bloom is now an AI power winner. That view is no longer contrarian. The stock has already moved from disbelief to expectation.

What the market may still underappreciate is that Bloom’s best asset is not any single order. It is the possibility that a few successful large deployments turn the machine from unusual to acceptable. In conservative physical markets, acceptability is the moat. Once customers, utilities, financiers, and local regulators believe a system works, the next sale becomes easier.

The second underappreciated point is that AI changes the metric. Bloom is not competing only on the cost of electricity. It is competing on the value of time. A delayed AI facility does not merely pay more for power; it may lose customers, compute revenue, and strategic position.

The third underappreciated point is the DC angle. If AI racks increasingly standardize around high-voltage DC designs, Bloom’s 800V DC readiness could matter more than traditional comparisons against turbines or engines imply.

But there is also a less pleasant truth the market may be underestimating: related-party revenue is not the same thing as clean end-user adoption. Brookfield may be an accelerant, but the more Brookfield-related entities become the buyer of record, the more investors need to separate financing proof from customer proof.

The business is more real than old skeptics believed. The stock is more demanding than new believers admit.

Three Possible Endings

The scenarios should be anchored to capacity, not only demand. Management said on the Q3 2025 call that doubling capacity to 2 GW by December 2026 would support about four times 2025 revenue. Since 2025 revenue was about $2.0 billion, that implies roughly $8 billion of blended revenue support at 2 GW.

The 2025 10-K also says the Fremont facility can accommodate up to approximately 5 GW of annual production capacity, with each additional 1 GW requiring roughly six to nine months and $100–150 million of capex.

That makes the scenario table more disciplined:

The Newcomen Ending is not failure. It is repricing. Bloom grows, but customers use it as bridge power, related-party revenue remains elevated, margins disappoint, and the market values the company more like energy equipment than scarce AI power capacity.

The Toll Road Ending is solid execution. Bloom converts the current logos, margins improve, service remains profitable, and Fremont scales beyond 2 GW. But growth is lumpy, customer concentration remains visible, and the multiple settles lower as the story becomes more industrial.

The Boulton Ending is the premium case. Bloom fills much of Fremont’s 5 GW potential, wins more large customers without material equity sweeteners, service margins remain high, and related-party revenue falls as a percentage of total. In that world, Bloom is not merely selling machines. It is charging for time saved.

A supercharged version above $25 billion of revenue probably requires either a second major manufacturing footprint or unusually high revenue per GW. That should not be the main case. It is the option.

The Scoreboard

The thesis is trackable. Use the same funnel for every large project:

Revenue announced is not revenue earned. Revenue earned is not necessarily cash. Cash is not necessarily clean if it is pulled forward by deposits, payables, or related-party structures.

The first tell is Oracle/Jupiter conversion: how much of the 2.45 GW moves from announcement to operating megawatts.

The second tell is the next large customer. If the next hyperscaler-scale deal comes without a warrant or equity sweetener, Bloom has more pricing power than skeptics think. If equity incentives become normal, early validation is being partly purchased.

The third tell is related-party revenue share. Q1 2026 was roughly half related-party. That number should fall over time if the customer proof is broadening.

The fourth tell is contract assets and deferred profit. If these grow faster than revenue, the quality question gets louder.

The fifth tell is product margin. It needs to stay in the mid-30s or better as AI projects grow.

The sixth tell is service margin. High-teens to 20% is supportive. A decline would suggest the installed base is becoming more obligation than asset.

The seventh tell is installation margin. Q1 2026 installation non-GAAP margin was negative 29.4%. If large AI projects increase installation complexity, this line matters more than the headline narrative suggests.

The eighth tell is capacity. The story requires 2 GW, then 3 GW, then 5 GW, not just as nameplate capacity, but as real output with acceptable quality.

The ninth tell is language. If management retreats from focused language about onsite power and time-to-power back into a broad catalogue of hydrogen, marine, biogas, carbon capture, and every possible application, that is not diversification. It is regression.

The Prediction

By mid-2027, Bloom will likely be in one of two places.

In the first, a second major AI customer has signed a large, clean, multi-site deal; related-party revenue has fallen as a percentage of total; Fremont expansion is visibly progressing toward 5 GW; service margins remain high-teens or better; and cash conversion improves. In that world, the market continues treating Bloom as one of the few public companies with direct leverage to the physical bottleneck in AI.

In the second, Oracle remains the defining logo, Brookfield-related sales stay a large share of revenue, customer warrants or financing structures remain necessary, and project timing becomes the recurring excuse. In that world, Bloom may still be a good company, but the stock will stop being valued like the scarce answer to a scarce problem.

Boulton charged for coal saved because there was no other engine. Bloom’s equivalent is time. The company will deserve its premium only when it can charge for the time it saves without giving too much of that value back through warrants, financing structures, margin concessions, or accounting complexity.

Bloom spent twenty-five years waiting for the world to need power where the grid was not ready. AI has made that world visible. Now Bloom has to prove that the machine can move the factory, and that the cash follows.

Subscribe now

• Our original Great Energy Mismatch thesis was right, but it has matured. The market no longer needs convincing that AI needs power; it now wants proof that scarcity can become signed contracts, durable cash flow, and per-share compounding.

• Q1 showed Vistra’s identity is changing. East delivered $801M of adjusted EBITDA versus $586M from Texas, while retail weakness from mild weather did not stop record Q1 EBITDA. The old ERCOT-weather lens is becoming incomplete.

• The next leg depends on conversion, not narrative. Cogentrix, Meta, Helix, gas bridge power, new PPAs, hedge pricing, and buybacks are the signposts. The base case is attractive, but the stock needs signed duration to move from “AI power winner” to “AI reliability platform.”

A year ago, our argument on Vistra was not really about Vistra.

It was about the grid.

The AI boom, we argued, was creating a mismatch between two clocks. On one clock were hyperscalers, building data centers on digital timelines: quarters, not decades. On the other clock was the power system, where generation, transmission, permitting, and interconnection are still measured in years. The scarce commodity was not simply electricity. It was time-to-power — reliable, dispatchable, located power delivered fast enough to matter.

That thesis has aged well. Maybe too well.

The market no longer needs to be convinced that AI needs power. It no longer needs to be told that nuclear is scarce, that gas is useful, that transmission is slow, or that data centers are forcing a rethink of the utility model. Those ideas have moved from variant perception to consensus narrative. Vistra is no longer an undiscovered beneficiary of the Great Energy Mismatch. It is now a stock the market already knows, already debates, and in many cases already likes.

That changes the question.

The old question was: will AI make power scarce?

The new question is: can Vistra convert that scarcity into contracted, visible, per-share cash flow before investors lose patience with the timeline?

That is the lens through which Q1 2026 should be read.

From Scarcity to Conversion

Our July framework was right about the direction of the world. AI demand was not merely adding load; it was exposing the inability of the grid to respond at digital speed. The four mismatches we highlighted — speed, scale, reliability, and location — are still the correct way to think about the problem. Hyperscalers do not just need electrons; they need dependable power in the right place, on the right timeline, with the right reliability profile.

But our framework was incomplete in one important respect. Scarcity creates optionality. It does not, by itself, compound.

What compounds is conversion: scarcity into contracts, contracts into predictable free cash flow, free cash flow into buybacks and reinvestment, and buybacks into per-share growth. That mechanism is sequential. Each step requires execution. The market is right not to pay fully for the last step simply because the first step is true.

There are also things we underweighted. We grouped Vistra and Constellation too closely as beneficiaries of the same scarcity trade. We treated gas mostly as a volatility hedge, when it is now becoming a time-to-power product. And we did not fully anticipate how quickly Vistra’s earnings center of gravity would move away from the old ERCOT-only mental model.

Q1 matters because that shift is no longer theoretical.

The Inversion

At first glance, the quarter looked straightforward. Vistra delivered $1.494 billion of ongoing operations adjusted EBITDA, up from $1.240 billion a year ago. It reaffirmed 2026 adjusted EBITDA guidance of $6.8–7.6 billion and adjusted free cash flow before growth guidance of $3.925–4.725 billion. The company also noted that the guidance range and the 2027 EBITDA opportunity exclude any potential benefits from Cogentrix and the Meta PPAs.

That is a good quarter. It is not, by itself, a thesis-changing quarter.

The more important detail was inside the segment table. Retail adjusted EBITDA fell from $184 million to $68 million, largely because Texas had an unusually mild quarter. Texas adjusted EBITDA rose from $490 million to $586 million. But East jumped from $514 million to $801 million.

For years, the simplest way to think about Vistra was through ERCOT: Texas heat, forward curves, retail margins, summer volatility. That lens is not wrong. It is just becoming incomplete. Q1 did not show that Vistra has stopped being a merchant power company. It showed that a smaller share of the value now comes from simply being long Texas weather.

This is also where Vistra and Constellation begin to diverge. Constellation is the cleaner nuclear scarcity trade. It has more direct torque to the re-rating of nuclear assets and long-duration clean baseload contracts. Vistra is messier, but that messiness is part of the point: nuclear duration, gas speed, retail integration, hedging discipline, and buyback-driven compounding all sit inside the same company.

Constellation may be the purer story. Vistra may have more ways to win.

The East-over-Texas quarter does not complete that transformation. But it does make the old label harder to defend. The market still trades Vistra partly as an ERCOT-sensitive merchant generator. The income statement is beginning to say something else.

The Product Is Time

The most useful line from our original Vistra article was not about nuclear or gas. It was that the scarce commodity in the AI era is time-to-power.

That remains the right idea, but Q1 sharpened it.

Nuclear is the premium product. It is clean, scarce, 24/7, and increasingly contractable over long durations. That is why the Meta PPAs matter and why investors naturally gravitate toward nuclear as the center of the AI-power story.

But gas solves the more immediate customer problem: how to start.

On the Q1 call, management repeatedly returned to bridge power, co-location, distributed generation, and customer flexibility. This was not an accident. Customers want grid connections, but they also want to energize projects before the full grid solution is ready. Jim Burke noted that the conversation has evolved toward bridge power, and that more of those conversations have been leaning toward gas.

That is a subtle but important update. Gas is not just a commodity exposure. In the AI infrastructure stack, gas is increasingly the liquidity layer, the resource that bridges the timing mismatch between hyperscaler urgency and grid reality.

That is also why the Helix Digital Infrastructure announcement matters. KKR launched Helix with more than $10 billion of committed capital, alongside Kuwait Investment Authority, NVIDIA, and Vistra; Vistra is the preferred power partner. The point is not that Helix is already a new PPA. It is that large pools of capital and AI infrastructure partners are looking for an integrated solution to data centers, power, and connectivity. They are not looking for a commodity megawatt supplier. They are looking for a partner that can help solve timing, reliability, and scale together.

That is the strategic opening for Vistra. Not power alone. Power certainty.

The Market Is Not Wrong

The temptation after a quarter like this is to say the market does not get it. That would be too easy.

The market does get it. That is why the stock already re-rated over the last two years. That is why investor conversations around Vistra now begin with AI load, PJM, ERCOT, co-location, hyperscaler PPAs, and dispatchable scarcity. The market’s problem is not that it has missed the thesis. The market’s problem is that the thesis has moved from concept to execution.

There was no new major PPA in the quarter. Guidance was reaffirmed, not raised. Cogentrix has not yet closed. Meta is not yet in the guide. PJM rules remain complex. ERCOT forwards have been soft. Batteries have dampened volatility in some periods. Interconnection queues still include too much fiction and not enough committed load.

All that matters.

Management’s answer is that physical development takes time. That is consistent with our own original thesis. The irony is that the same timing mismatch that creates Vistra’s opportunity also delays its monetization. Customers want speed-to-power, but market design, tariff rules, interconnection processes, and approval cycles are still catching up.

The evidence suggests the pipeline is real. Management described customer engagement as extremely strong, and Helix supports the idea that Vistra is being pulled into the AI infrastructure stack. But the market is right to distinguish pipeline from signed duration.

This is a better stock for patient capital than for event traders.

The Compounding Mechanism

The buyback is not the thesis, but it is the mechanism that makes the thesis per-share relevant.

Since November 2021, Vistra has executed roughly $6.3 billion of share repurchases, reduced shares outstanding by about 30%, and still had about $1.5 billion of authorization remaining as of May 1. On the Q1 call, management noted that the program had retired about 169 million shares at an average price of roughly $37.

That math matters. A company that buys back stock at $37 and then sees its earnings base structurally reset has not merely “returned capital.” It has used the old market’s skepticism to compound the new market’s earnings base.

This is why the distinction between scarcity and conversion is important. If Vistra simply rides a power-price cycle, the buyback is nice but not transformative. If Vistra converts scarcity into more durable cash flows, the buyback becomes part of a flywheel: higher demand, higher utilization, more contracted revenue, more free cash flow, lower share count, higher per-share value.

That is the real bull case. Not AI demand alone, but AI demand converted into a per-share compounding machine.

Three Vistras in 2029

The stock is not obviously cheap on current-year numbers. Based on Bloomberg consensus shared, Vistra trades at roughly 10x FY2026 EV/EBITDA, falling to around 9x FY2027 and 8.5x FY2028. That is not deep value. It is a bet that current estimates understate the durability of the earnings base and the upside from contracts not yet in guidance.

The base case is not heroic. It assumes the visible pieces gradually enter the model and that management continues doing what it has been doing: hedging intelligently, buying back stock, keeping leverage consistent with investment-grade ratings, and deploying capital only where returns justify it.

The bull case requires more than AI excitement. It requires signed duration.

The bear case does not require AI demand to disappear. It only requires monetization to disappoint.

What We Watch

The first signpost is the Cogentrix close and the subsequent guide update. The current guide explicitly excludes Cogentrix and Meta benefits, so the market will want a clean bridge from today’s $6.8–7.6 billion EBITDA range to a higher 2027 earnings base.

The second is the next large-load contract. Beaver Valley, additional Comanche capacity, gas-backed bridge power, or a new-build arrangement would all matter, but the market needs to see one of them move from discussion to paper.

The third is the East/Texas mix. If East remains comparable to, or larger than, Texas, the ERCOT-weather label becomes less useful.

The fourth is 2028 hedge pricing. Vistra is heavily hedged for 2026 and 2027, but 2028 still leaves meaningful exposure. If load growth begins to show up in forward curves, the market will reprice that optionality.

The fifth is Helix. If it becomes a source of real deal flow, it will validate Vistra’s role as a full-stack AI power partner. If it remains branding, the market will discount it.

Scarcity Is Not Enough

Vistra’s quarter did not settle the debate. It clarified it.

The old question was whether AI would make power scarce. The new question is whether Vistra can turn scarce power into durable, contracted, per-share cash flow. Q1 gave us evidence that the earnings base is already changing: East over Texas, generation over retail, hedged cash flow over weather noise. But it also reminded us that the next stage will not be won by narrative. It will be won by contracts, guidance revisions, and disciplined capital deployment.

That is a less exciting story than “AI needs power.”

It is also the story that matters now.

Subscribe now

• Q1 vindicated Meituan’s network, not its old margin structure. The subsidy war did not break the platform: users stayed, losses narrowed sharply, and Core Local Commerce began to repair. But the old monopoly-like economics are unlikely to return unchanged.

• The war did not end; it moved. Food-delivery subsidies are being regulated down, but competitors are now attacking other layers of Meituan’s local-commerce stack: Alibaba in grocery supply, JD in fulfilment, Douyin in discovery, and AI agents at the consumer interface.

• The investment question is no longer whether Meituan survives, but what it earns as the local-commerce operating system. If Meituan keeps control of intent, trust, supply and fulfilment, the stock looks mispriced. If those layers fragment, Meituan becomes infrastructure, still large, but worth much less.

In September, we wrote that Meituan was the digital bellhop.

The point was not food delivery. It was urban complexity. The old hotel bellhop was valuable because he knew which doctor made house calls, which restaurant stayed open late, which shortcut avoided the rain. Meituan’s version of that knowledge was reviews, merchants, riders, routes, membership, payments and habit. Our argument was simple: subsidies are temporary; the local operating system is permanent.

In March, after Q4, we updated the thesis. The moat had survived, but the old margin structure probably had not. Users stayed. Engagement held. The platform did not lose relevance. But the economics of running it had changed. The right question was no longer “when does the subsidy war end?” It was “what does the peace treaty look like?”

After Q1, the next update is clearer.

The peace treaty did not arrive. The war moved.

That is not a reversal of the thesis. It is the thesis becoming more precise. We have not changed our mind about what Meituan is. We have changed our mind about what it can earn, and where the next attack will come from.

Q1 vindicated the network thesis. It did not vindicate the old margin thesis.

What the Bellhop Bought

The most important thing Meituan did this quarter was not only in the income statement. It changed how it reports revenue.

Starting in Q1, Meituan separately disclosed “product sales.” This is gross-basis revenue, mainly from grocery retail businesses such as Xiaoxiang Supermarket and Kuailv. The number was RMB21 billion, or 23% of revenue, growing 46.6% year over year.

Meanwhile, delivery services revenue declined 2.9%. Merchant services, the higher-margin platform line, grew only 3.5%.

Companies do not reorganize disclosure for no reason. They do it when they want investors to see the business the way management sees it. What Meituan is saying is uncomfortable but important: the fastest-growing part of the company is not a pure platform line. It is retail, supply chain and inventory.

The market sees this and says: margin dilution.

That is correct, but incomplete.

Wang Xing made the strategic argument on the call. Consumers do not care whether supply is 1P or 3P. They care whether the product is available, good quality, fairly priced and reliably delivered. That is the argument for why an operating system sometimes needs hardware.

A platform can stay asset-light when supply is abundant, standardized and reliable. Fresh grocery is none of those things. Availability matters. Cold chain matters. Private label matters. Returns matter. Trust matters. If the platform’s promise is “everything now,” Meituan cannot rely only on someone else’s shelf.

That is why Xiaoxiang matters. That is why Dingdong matters. And that is why Alibaba’s reported bid for Pupu matters.

Grocery may lower reported gross margin, but it increases strategic control. It gives Meituan the pantry.

That trade is not automatically good. Grocery has humbled many confident technology companies. It is operationally heavy, low-margin and unforgiving. But calling it accidental dilution misses the point. Meituan is accepting lower accounting purity in exchange for more control of the daily-consumption layer.

The question is whether that control compounds or merely consumes cash.

Kiosks in the Lobby

While Meituan was remaking itself from the inside, competitors were attacking from the outside, but not exactly where our original framework expected.

The food-delivery subsidy war appears to be moving toward regulation. SAMR’s draft rules target practices such as forcing merchants into subsidies, selling below cost, and using financial strength for unfair competition. That matters because it gives all players political cover to step back from economically destructive behavior.

This is good for Meituan. It makes Q1’s sequential repair more credible.

But competitors are not leaving local commerce. They are changing the venue.

Alibaba is trying to build or buy grocery density. JD is extending instant delivery into group buying. Taobao Flash Sale is piloting local group buying. Gaode is pushing local ranking. Douyin has Doushengxing, a dedicated group-buying app that reportedly reached meaningful daily usage and a high coupon verification rate within months.

This challenges one part of our September framework.

We distinguished between intent and impulse. Meituan was intent: users opened the app because they needed lunch, groceries or a service. Douyin was impulse: users discovered a deal while scrolling. Intent had better economics because demand already existed.

Doushengxing is ByteDance trying to build an intent product. No videos, no livestreams, no entertainment. Just local deals.

That is not trivial. It is the right attack.

But it still helps to return to the bellhop.

A hotel kiosk can solve one problem extremely well: check-in. It is fast, available and efficient. For the guest who knows exactly what he wants, the kiosk may beat the bellhop.

But the guest who arrives in an unfamiliar city, has dietary restrictions, wants to avoid tourist traps, needs a reliable dentist, or has a delayed flight still wants the bellhop.

Doushengxing is a kiosk. It can distribute coupons efficiently to deal-seeking users. That will pressure Meituan in the price-sensitive in-store segment. Some of Meituan’s historical in-store margin was earned because there was no serious alternative. That margin will not fully return.

But Meituan’s in-store business is more than a coupon book. It is reviews, rankings, merchant verification, reservation, queuing, prepaid-service assurance, medical-device checks, artisan profiles, loyalty integration and merchant tools. Those are not features you replicate by launching a coupon app.

The question is not whether kiosks work. They do.

The question is how much of local commerce is check-in, and how much is judgment.

If Douyin owns discovery, Meituan becomes infrastructure. If Meituan keeps trust, reviews and transaction closure, Douyin becomes another demand channel. That distinction will decide the margin pool.

What Q1 Actually Confirmed

The numbers matter because they tell us whether the operating inflection is real or aspirational.

It is real, but not complete.

Revenue rose 5.6% year over year to RMB91 billion. That is not impressive by itself. The company still lost RMB6.8 billion, compared with a RMB10.1 billion profit a year ago. Core Local Commerce, the segment that matters most, moved from RMB13.5 billion profit last year to RMB2.0 billion loss.

That is the damage.

But sequentially, the repair was significant. Total operating loss narrowed from RMB16.1 billion in Q4 to RMB6.5 billion in Q1. Selling and marketing fell from 34.4% of revenue to 25.2%. Core Local Commerce operating margin improved from negative 15.5% to negative 3.2%. New Initiatives loss narrowed from RMB4.6 billion to RMB2.1 billion.

This is why Q1 matters. Meituan reduced the intensity of the subsidy response and the system did not collapse.

Management’s language was also revealing. It was not chasing all volume. It emphasized high-AOV orders, core users, better user structure, resilient GTV, and supply quality. It even warned that order volume growth may turn negative year over year in the second half because the company is prioritizing healthier mix.

That is not a company trying to win a vanity metric. It is a company trying to repair unit economics without breaking habit.

So the quarter answered the old question: can Meituan survive the subsidy war?

Yes.

But the post-Q1 developments ask a better question: can it still control the local-commerce loop when competitors attack each layer separately?

The Operating System Test

The market sees a battered delivery company recovering from a bad year.

I think that framing is too narrow.

Meituan is a local-commerce operating system. Delivery is the backbone, but it is not the whole product. The product is the relationship: the user who starts with Meituan, the merchant who depends on Meituan, the rider who fulfils the order, the review that reduces uncertainty, the membership that increases habit, the grocery supply that guarantees availability, and the AI assistant that may become the new interface.

The operating system works because each layer reinforces the next.

More users bring more merchants. More merchants create more choice. More choice creates better reviews and rankings. Better reviews increase trust. More transactions improve fulfilment density. Better fulfilment increases frequency. More frequency improves merchant advertising ROI. Better merchant ROI improves supply. Better supply brings more users.

That is the loop.

Competitors are not trying to beat the whole loop at once. They are attacking pieces of it.

Alibaba attacks supply and commerce. JD attacks fulfilment credibility. Douyin attacks attention and discovery. Gaode attacks local search. AI agents may attack the front door.

Meituan wins if integration beats specialization.

It loses if the loop fragments.

Variant Perception

Consensus sees a repair story: food-delivery subsidies fade, losses narrow, margins recover.

The correct framing is a reclassification story. Meituan is becoming a lower-margin but more deeply embedded local-commerce infrastructure company. The old margin percentage may not come back. But lower margins on a larger, more controlled base can still create attractive absolute earnings power.

The market is also collapsing two timelines into one.

Food delivery repair is happening now. In-store pressure is emerging but has not yet fully appeared in the P&L. Grocery dilution is visible, but its strategic value will take longer to prove. AI is noisy today, but the interface question may matter later.

At today’s price, the market is pricing Meituan as if all the bad timelines arrive together and the good timeline is uncertain.

Q1 suggests the recovery timeline is real. The competitive developments suggest the risk timeline is also real.

That is why the stock is interesting, not easy.

Three Futures

Using roughly HK$67–68 as the reference price, I would frame the next three years this way:

The bear case is not that Meituan disappears. It is that the best economics leak. Douyin captures more in-store discovery. Alibaba and JD make instant retail more competitive. Grocery becomes necessary but low-return. Meituan remains large, but more like infrastructure than operating system.

The base case is that food delivery rationalizes, Core Local Commerce returns to profitability, and in-store margins settle lower but remain attractive. Xiaoxiang and Dingdong improve supply control without becoming uncontrolled cash drains. KeeTa narrows losses but does not drive the valuation. Meituan earns less than old bulls remember, but more than the current price implies.

The bull case requires reclassification. Food delivery becomes rational. In-store withstands Douyin. Grocery density strengthens instant retail. Merchant advertising recovers. Membership deepens habit. Xiaotuan and agent integrations protect the front door. The market stops valuing Meituan as delivery and starts valuing it as local-commerce infrastructure.

What Would Change My Mind

First, Core Local Commerce failing to reach breakeven in Q2 or Q3 would mean the food-delivery recovery is slower than Q1 implied.

Second, in-store commentary deteriorating sharply would mean the kiosks are taking more than the deal-seeking segment.

Third, product sales growing while New Initiatives losses widen would suggest grocery is consuming control rather than creating it.

Fourth, S&M returning toward 30% of revenue would mean the subsidy war has reappeared through another door.

Fifth, operating cash flow staying meaningfully negative into the second half would weaken the argument that this is disciplined strategic investment.

The Building and the Bellhop

In September, the bellhop’s problem looked simple: two wealthy guests were throwing money around the lobby to prove they did not need him. We said the money would eventually run out. Directionally, that was right. But it was too simple.

In March, the problem became subtler: the bellhop survived, but at lower wages. We said the question was no longer whether he survived, but what he earned.

Now the building has kiosks in the lobby, a rival management company pitching the landlord, and a grocery store in the basement the bellhop just bought.

His advantage is no longer monopoly. It is judgment.

The kiosk tells you which restaurant is cheapest. The bellhop tells you which one is good. The kiosk processes your coupon. The bellhop remembers that you do not eat shellfish. The kiosk can complete a transaction. The bellhop manages a relationship.

The market is watching the kiosks multiply and concluding the bellhop is obsolete.

I think the market is confusing transactions with relationships.

Transactions can be automated. Relationships compound.

Meituan’s fifteen years of reviews, merchants, riders, memberships, fulfilment infrastructure and operating data are not just transactions. They are the operating memory of local commerce in urban China.

At HK$67–68, the market is paying for a damaged delivery company.

The Q1 evidence suggests it may be buying a repaired local-commerce operating system.

That difference is the opportunity.

The next war will decide how much of it Meituan keeps.

Subscribe now

• The Tencent thesis has not reversed; it has climbed the stack. We started with under-monetized distribution, then AI-driven yield, and now the question is whether Weixin becomes the operating layer where AI agents actually act.

• 1Q26 gave evidence, not victory. Core Tencent remains strong enough to fund the buildout, ex-new-AI operating profit grew 17%; while Hy3, WorkBuddy, AIM+, Video Accounts, and Mini Programs-as-skills show the tollbooth moving from metaphor to architecture.

• The market is rationally skeptical, but the price may be too harsh. At around HK$420, Tencent is being valued like old Tencent under China-internet pressure, while investors get a low-cost option on new Tencent, if agents route through Weixin rather than bypass it.

The stock is at HK$420. It was HK$662 when we wrote The Patience Premium. It was around HK$500 when we wrote that the AI dividend was being reinvested. Everything we have said about the business has become more specific. None of it has been rewarded by the stock.

That requires an honest reckoning before we earn the right to say anything about 1Q26.

The original Tencent thesis was about restraint. Tencent had spent years building assets it deliberately refused to over-monetize: Video Accounts, Weixin Search, Mini Programs, Mini Shops, payments, games. The point was not that Tencent lacked revenue levers. It was that Tencent had more levers than the market valued, because management had chosen not to pull them too early.

The second thesis was about yield. AI was not primarily valuable to Tencent because it could build the biggest foundation model. It was valuable because Tencent already owned the distribution surface. Better recommendations, better targeting, better content creation, better search and better live operations could raise yield on users Tencent already had.

The third thesis forced a correction. The AI dividend was real, but it was not simply being harvested. Tencent was reinvesting it. We were right that AI could improve the core. We were wrong that Tencent would avoid the infrastructure cycle. Management chose to spend.

That is not a reversal. It is an evolution. The strategic direction has been consistent: Tencent’s value sits in distribution, not just in products; in controlled surfaces, not just in content; in patient monetization, not extraction. What changed was the mechanism. First levers, then yield, now execution.

The question after 1Q26 is therefore not whether Tencent still has a patience premium. It does. It is not whether AI is improving the business. It is. The question is more ambitious and more dangerous:

Can Tencent turn Weixin from a super-app for human attention into an operating layer for AI agents — without turning AI into a permanent tax on the core business?

The Question Changed

A chatbot answers a question. An agent does something.

That distinction matters. Doing something requires more than intelligence. It requires identity, permissions, service access, payment rails, communication and trust. Tencent has a model, Hunyuan. But that is not the heart of the thesis.

Tencent has Weixin.

Weixin has the user relationship. Mini Programs provide service access. Weixin Pay processes transactions. WeCom and QQ extend the communication layer into work and social contexts. Video Accounts captures attention. Weixin Search captures intent. Tencent Cloud provides infrastructure. Games, music, video and literature provide IP and engagement. Hunyuan increasingly provides intelligence.

This is why the model race is the wrong frame. Tencent does not need every Chinese consumer to open Yuanbao before Doubao, Qwen or Kimi. It needs agents to act inside Tencent’s surfaces when users delegate tasks.

The product is not the chatbot. The product is Weixin with agents inside it.

1Q26 mattered because this idea moved from metaphor toward architecture.

Tencent said users can control AI agents through Weixin, WeCom, QQ, Yuanbao, QQ Browser and third-party applications. More importantly, it said that over time, AI agents will be able to access the Mini Programs ecosystem using Mini Program code as AI skills.

That sentence is the article.

Mini Programs were originally lightweight apps inside Weixin. In an agentic world, they become callable tools. The same service layer that let merchants, restaurants, government services, games, retailers and content providers live inside Weixin can become the tool layer agents call when they need to complete a task.

If a user says, “book a table, use my coupon, tell my friend, and pay the deposit,” the model is only one part of the job. The agent needs identity, merchant access, transaction capability and a communication loop. Weixin has all four.

That is the tollbooth.

Two Companies, One Filing

The most important number in Tencent’s quarter was not revenue. Revenue grew 9% to RMB196.5 billion. Gross profit grew 11%. Non-IFRS operating profit grew 9%. Non-IFRS net profit grew 11%. Fine numbers, not extraordinary numbers.

The important number was RMB84.4 billion.

That was non-IFRS operating profit excluding new AI products, up 17% year-on-year. Consolidated non-IFRS operating profit was RMB75.6 billion. The difference — roughly RMB8.8 billion — is the cost of AI Tencent: Hunyuan, Yuanbao, CodeBuddy, WorkBuddy, QClaw and the associated product investments.

Tencent is now effectively showing investors two companies in one filing.

The first is Core Tencent: games, ads, payments, Weixin, content, cloud and enterprise services. This business is still compounding operating profit at a mid-teens rate and, excluding new AI products, generated a 43% non-IFRS operating margin.

The second is AI Tencent: young, expensive, not yet fully monetized, and consuming some of the profit the core business generates.

This is the cleanest evidence that the thesis has not broken. The core is not being hollowed out. It is funding the experiment.

That does not mean the experiment is automatically good. A great core business can fund bad projects for years. The question is whether AI becomes a compounding layer or a permanent tax. But the first condition is met: Tencent can afford to build.

The market sees the same facts and reaches a different conclusion. It sees capex, depreciation, slower buybacks and no explicit AI ARR. It sees a China internet stock in a hostile regime. It sees management calling the stock dislocated while still choosing compute over a more aggressive repurchase.

The market is not stupid. It is refusing to capitalize the tollbooth until the tolls are visible.

The Storm and the Tollbooth

This is the part we underweighted.

The prior pieces focused on Tencent’s supply side: what Tencent could build, what levers it could pull, what AI could improve. The demand side — what happens when China consumption is weak, Hong Kong tech de-rates, and investors sell the whole internet complex — was treated as a risk, not the regime.

That was too light.

Tencent is not falling because investors carefully modeled WorkBuddy and rejected it. It is falling because the regime is hostile. China internet is in a penalty box. AI capex is being treated as a burden before revenue is visible. Capital is not rewarding Hong Kong-listed platform optionality. The market is asking for proof, not architecture.

The uncomfortable truth is that both sides can be right. Tencent can be building something valuable, and the stock can stay cheap until the value is legible.

But there is an inversion worth taking seriously: the same regime crushing the stock may make the tollbooth more necessary.

When consumers are cautious, merchants need more efficient distribution. When ad budgets tighten, better targeting matters more. When enterprises scrutinize spend, productivity agents that save time and cost become more attractive. When platforms compete for scarce dollars, closed-loop attribution becomes more valuable.

That does not automatically make Tencent immune. Weak consumption can still hurt payments, ads, games and merchant appetite. But it does make the strategic problem Tencent is solving more urgent: how to make every unit of attention, intent and transaction more productive.

Advertising is the first proof.

Marketing Services grew 20% year-on-year to RMB38.2 billion. AIM+, Tencent’s automated campaign management solution, powered roughly 30% of total marketing services spend. Video Accounts time spent grew over 20%. Weixin Search query volume grew over 25%.

This is AI monetization already showing up in the old business.

Better models improve recommendations. Better recommendations increase time spent. More time spent creates inventory. Better targeting improves advertiser ROI. Higher ROI supports pricing. Higher monetization funds more compute. More compute improves the models again.

That is the AI dividend. The difference now is that the dividend is funding the tollbooth.

The Quarter Was Better Than the Headline

The headline revenue miss was real, but it was not the story.

Domestic Games revenue grew only 6%, while gross receipts grew at a teens rate. The gap was timing: the later Spring Festival shifted revenue recognition into future quarters. Deferred revenue rose sharply, which means some of the economic activity already happened but has not yet appeared in reported revenue.

More importantly, the games business looked structurally strong. Honour of Kings, Peacekeeper Elite and Delta Force reached lifetime highs in gross receipts. Peacekeeper Elite hit a lifetime-high peak DAU of 90 million. Roco Kingdom launched with more than 13 million average DAU in its first month. League of Legends saw a resurgence after process and content changes at Riot.

This is not just a good games quarter. It is evidence that Tencent’s games are behaving less like hits and more like platforms.

A hit launches, peaks, fades and needs replacement. A platform accumulates users, adds modes, expands content, deepens social loops and monetizes through live operations. AI helps because it increases content velocity: 3D assets, animation, rendering, intelligent guides, NPC behavior, personalization and faster refresh cycles.

The reported quarter looked slower than the economic quarter.

What Would Prove the Tollbooth

The thesis is stronger than before, but the burden of proof is also higher.

Tencent has not yet disclosed agent transaction volume, agent task completion, WorkBuddy ARR, Mini Program AI-skill usage, or a clear payback period for AI capex. Retention is useful. Deployment across internal products is useful. OpenRouter token usage is useful. None of it is the same as toll revenue.

That is why the market is skeptical.

The single biggest risk is the OS-layer bypass. If Apple, Huawei, Android or another device-level agent becomes the default instruction layer, Weixin may remain important for execution but lose control of intent. In that world, Tencent is still a powerful infrastructure layer, but not necessarily the agent operating layer.

The second risk is capex permanence. If AI spending keeps rising while cloud, ads, productivity and games do not show enough incremental monetization, the AI layer becomes a tax on Core Tencent rather than a multiplier of it.

The third risk is the regime. China internet de-ratings can last longer than investors expect. Cheap can stay cheap when the factor is out of favor.

So the dashboard matters.

Ex-new-AI operating profit growth above 15% means the core can still fund the transition. Below 10% means the engine is weakening. Business Services growth above 25–30% in the second half would validate management’s claim that cloud AI monetization was supply constrained. Video Accounts ad load moving toward 6–7% while time spent still grows double digits would prove AI-powered monetization. WorkBuddy and CodeBuddy need to move from retention to revenue: paid seats, ARR, enterprise customers, task volume or token consumption. Quarterly free cash flow above RMB50 billion means the buildout is fundable. Below RMB40 billion for multiple quarters means the market’s concern is right. Any disclosure around Mini Programs as AI skills — timeline, developer tools, usage, task volume, closed-loop transactions — would move the thesis from concept to proof.

Three Futures

The bull case is The Tollbooth Works. Revenue compounds at 12–14% over three years. Gross margin reaches 59–60%. Non-IFRS operating margin settles around 40–42% even after AI investment. Business Services sustains 25%+ growth as Tencent Cloud monetizes AI demand. Marketing Services grows high-teens to low-20s as Video Accounts ad load rises without damaging engagement. WorkBuddy and CodeBuddy move from retention to revenue disclosures. Mini Programs become agent skills in practice, not just concept. The market re-rates Tencent to 18–20x FY29 earnings. That points to roughly HK$900–1,050. Probability: 25%.

The base case is The Machine Funds the Road. Revenue compounds at 8–10%. Gross margin rises gradually to 57–58.5%. Non-IFRS operating margin remains around 37–39%. Games remain resilient. Ads continue to outgrow the industry. AI capex stays elevated, but free cash flow recovers after FY26. The agent layer becomes real but not yet dominant. The market applies 14–16x FY29 earnings. That points to roughly HK$650–750. Probability: 50%.

The bear case is The Road Is Bypassed. Revenue compounds at 4–6%. Gross margin falls toward 54–55%. Non-IFRS operating margin compresses to 31–34%. AI capex stays high without visible monetization. Business Services fails to accelerate. Video Accounts monetization stalls or hurts engagement. OS-level agents capture intent before Weixin agents form habits. Tencent remains a good old internet company, not the action layer for the new one. The market values it at 9–11x FY29 earnings. That points to roughly HK$330–400. Probability: 25%.

The scenario table says something important. The bull case is no longer simply “Tencent monetizes Video Accounts.” That was the Patience Premium. The bull case now requires Tencent to own a new layer.

The burden of proof is higher. The entry price is lower.

The Stock and the Thesis

So, have we changed our view too often?

I do not think so. The strategic direction has been consistent. The evidence changed, and the mechanism evolved.

We were right that Tencent had under-monetized distribution. We were right that AI could raise yield on that distribution. We were wrong that Tencent could avoid the infrastructure cycle. We were right to update the thesis from harvesting to building. 1Q26 now gives evidence that the buildout is real, but not yet proof that it pays.

That is the honest answer.

Tencent’s 1Q26 did not prove it owns China’s agent economy. It proved something narrower but still important: the core machine is strong, the AI spend is visible, the product evidence is emerging, and the market is not paying much for the possibility that the tollbooth works.

The stock is trading the regime. The thesis is trading the tollbooth.

One of those is temporary. The other is still unproven. That is what makes the position uncomfortable. It is also what makes it interesting.

Subscribe now

• The key shift is buyer behavior: memory sellers once tried and failed to fix prices; now customers are voluntarily signing five-year take-or-pay contracts with deposits because AI memory shortages can strand far larger infrastructure investments.

• HBM and data-center products are weakening the old commodity framework: frontier memory is becoming less swappable as qualification, packaging, thermals, base dies, controllers, and firmware matter more than JEDEC-style interchangeability.

• The race is against supply: Micron’s upside depends on moving enough revenue into contracted, differentiated products before CXMT and incumbent greenfield fabs bring the industry back toward ordinary memory-cycle economics.

In 2002, executives from Samsung, Hynix, Infineon, and Micron met in hotel bars to fix DRAM prices. They had no choice, or so they believed. Memory was the purest commodity in technology, every chip interchangeable with every rival’s, every margin competed to zero, and the only way to hold a price was illegally. They got caught. The United States levied criminal fines exceeding $700 million. Executives went to prison. And prices fell anyway, because a cartel cannot hold together in a commodity where any member can defect by adding a wafer line.

That episode is the most useful fact about the memory industry, and not for the reason people usually cite. The cartel is typically offered as evidence of bad behavior. It is evidence of structural helplessness. These companies tried the most extreme measure available to escape the commodity curse, and it did not work. The standard was too open, the product too swappable, the entrants too numerous. No amount of coordination could override the structure.

What happened twenty-four years later is the opposite in every respect, not illegal coordination among sellers, but voluntary commitment from buyers. Micron reported $41.5 billion in quarterly revenue, 84.9% gross margins, and $18.3 billion in free cash flow, and then announced that sixteen of its customers had voluntarily signed five-year contracts with price floors, volume commitments, and $22 billion in cash deposits to guarantee access to a product that used to be available from anyone, at any time, at whatever the spot market offered.

The sellers used to beg for discipline. Now the buyers beg for supply. That inversion is the story.

The Quarter Was Evidence, Not the Thesis

The numbers deserve acknowledgment, not extended discussion. They are the evidence. They are not the thesis.

A traditional memory analyst sees this table and immediately asks: where is peak? DRAM ASPs rose roughly 60% sequentially on low-single-digit volume growth. NAND ASPs rose roughly 85% on mid-single-digit volumes. This was a price story, not a volume story. In old memory, that is exactly when you sell.

That instinct is not wrong. It is simply answering the wrong question. The question is not whether Q3 was great. It was. The question is whether Q3 belongs to the old memory cycle or to a new earnings architecture. The difference is the contracts, and the reason the contracts exist is that something deeper has changed about the product itself.

The Death of Swappability

For forty years, memory’s commodity-ness was enforced by JEDEC, a standards body whose entire purpose was to guarantee interchangeability. A DDR module from Micron had to work in the same slot as one from Samsung or SK Hynix. When a component is perfectly swappable, buyers compete vendors entirely on price. That was the commodity curse, and it was the system working as designed.

What is happening at the AI frontier is that the most valuable memory is becoming less swappable. HBM nominally adheres to a standard, but the practical reality of HBM4 is that the base die, the logic layer at the bottom of the stack, is fabricated on an advanced foundry node and increasingly customized for specific accelerators. A stack qualified for NVIDIA’s platform is not the same product as one built for AMD. The thermal profile, packaging, power envelope, and qualification are all different. You cannot unplug one and plug in the other.

This matters because it changes the economics. The more memory becomes tied to accelerator design and customer qualification, the less it behaves like a spot commodity. The same logic extends to data-center SSDs: Micron’s data-center SSD revenue exceeded $5 billion and more than doubled sequentially. That is not proof NAND has become a franchise, but it is evidence that the most valuable NAND demand is also moving up the stack, toward controllers, firmware, and enterprise qualification.

AI has not repealed the memory cycle. It has changed which parts of the market matter most. And the parts that matter most are, for the first time, not perfectly swappable.

Sixteen Agreements

If swappability is weakening at the frontier, the SCAs are Micron’s attempt to address the other half of the problem: cyclicality.

Sixteen signed agreements. Five-year terms, take-or-pay, non-cancellable. Covering roughly 20% of DRAM volume and a third of NAND, with a target of half or more of total revenue. Fourteen of the sixteen carry cumulative minimum-price revenue of approximately $100 billion, supported by $22 billion of customer deposits and financial commitments. Management says the price floor enables margins “well above our peak quarterly margins in any past cycle.”

The coverage is incomplete. Twenty percent of DRAM volume is not transformation; it is the beginning of one. But the direction is unmistakable, and the most important evidence for the direction is not the contracts themselves, it is why the customers signed them.

Take-or-pay contracts with billions in deposits are not procurement decisions. They are board-level strategic commitments. A hyperscaler with GPUs, power infrastructure, and cloud demand lined up loses far more from a memory shortage than a deposit costs. The $22 billion is insurance against stranded infrastructure, a revealed preference about how these companies expect AI spending to evolve through 2030. When the largest technology companies in the world voluntarily lock into five-year purchase commitments at current pricing, that is a stronger demand signal than any analyst TAM forecast, because it has money behind it.

A contract is only useful if both sides have a reason to honor it. Micron wants earnings visibility. Customers want supply certainty. The agreements exist because scarcity has become strategic for both parties.

China Is The Clock

Here is where the story gets complicated, and where the market’s instinct is probably directionally right.

SemiAnalysis published a detailed analysis of CXMT the day before Micron’s earnings. The numbers were striking: Q1 2026 revenue of $7.3 billion with roughly 700% year-over-year growth, operating margins of approximately 70%, wafer capacity approaching Micron’s ~385,000 wafers per month, and ASPs within 5–10% of the big three on commodity DRAM. CXMT is adding roughly 85,000 wafers per month this year, the largest single-player ramp in the industry. HP is reportedly in discussions about sourcing CXMT chips for Asia-bound PCs.

I wrote about the strategic logic the same day. The big three’s discipline created a price umbrella: pricing so elevated that a fourth player can shelter underneath it profitably, funding its own R&D and capacity expansion from the rents the incumbents enabled. Maximum rent in hardware does not just attract competition, it finances competition. That is the unintended consequence of 85% margins.

CXMT does not need to win HBM to hurt Micron. It only needs to pressure enough commodity DRAM for investors to refuse a franchise multiple on the blended business. The frontier is protected by a knowledge moat that took decades to build. But if 30–40% of Micron’s revenue faces real competition from a state-subsidized entrant, the market will cap the blended multiple regardless.

This is where the SCA timing reveals its real purpose. Micron disclosed sixteen agreements at the exact moment the fourth player is reaching competitive scale. The SCAs concede, implicitly, that commodity DRAM will face Chinese competition, and they ensure that when it does, half of revenue is contractually insulated. It is the legal solution to the problem that a criminal cartel could not solve in 2002, not because the law changed, but because scarcity flipped the leverage.

Notice what CFO Mark Murphy deflected on the call. When pressed for a hard FY27 capex number, he refused, conceding only that it would be above the mid-$40 billion range, then immediately pivoted to emphasize that more than half the increase would be construction, spending that takes years to produce bits. Management is acutely aware the market fears oversupply. The construction framing is their way of saying: this capex will not flood the market in FY27. Whether it floods the market in FY28 is the question they are not yet answering.

The Right Framework

The valuation debate is not about whether Micron’s earnings are going up. That has already happened. The debate is about what kind of earnings they are, and how that determines the multiple.

Old Micron was the first row. If 85% margins are purely cyclical, then peak EPS of $175–200 at 4–6x gives you one range. If those same earnings are partially contracted, partially differentiated, and partially protected by floor pricing, then 8–11x on that earnings power gives you a very different one. The gap between those two frameworks is the entire stock.

The honest answer is that Micron is migrating from the first row toward the third, and the speed of that migration is what matters. The more revenue under SCAs, the more mix in HBM and data-center products, the more the trough rises and the more the multiple deserves to expand. The more revenue remains in spot-priced commodity DRAM, the more the old framework applies.

This is why “cyclical franchise” is the right category. Not an infrastructure monopoly, China is real, capex is real, memory supply will eventually respond. But also not the old book-value memory stock, if a meaningful and growing portion of earnings is contracted, prepaid, and tied to products customers cannot easily swap.

The Race

Micron is running two races against two clocks.

The contract race: from 20% DRAM coverage toward 50%+ of revenue before supply normalizes. Each SCA signed removes supply from the spot market and pushes remaining customers toward their own commitments. The mechanism is self-reinforcing, but at 20% it is a thesis, not a fact.

The product race: HBM, data-center SSDs, and qualified AI memory growing as a share of the business so the commodity floor becomes less relevant. Data center revenue already exceeds $25 billion per quarter. The frontier is growing faster than the base.

The clocks: CXMT adding capacity and gaining share in commodity DRAM. Greenfield fabs from all three incumbents beginning to produce bits in calendar 2028. The supply response is coming. It always does.

In 2002, memory executives broke the law to hold a price, and it did not work. In 2026, their customers are voluntarily locking in prices, depositing billions, and signing contracts they cannot exit. The structure that made the cartel necessary, perfect swappability, open standards, commodity economics, is fraying at the frontier, even as it persists at the floor.

Micron’s opportunity is to contract enough of the cycle, and differentiate enough of the product, before China and its own new fabs make memory look ordinary again. That is the race. This quarter’s evidence says Micron is ahead. The clock says the race is not over.

Subscribe now

• Adobe’s core issue is not lack of AI traction, but weak AI monetization: AI-first ARR, Firefly, GenStudio, enterprise adoption, and freemium usage are all growing strongly, yet organic ARR quality weakened because Adobe is sacrificing near-term paid conversion to rebuild its funnel around freemium.

• AI has broken Adobe’s old moat while creating a new possible one: the traditional skill-based Creative Cloud moat is eroding as AI makes creation easier, but the same explosion in content strengthens Adobe’s opportunity in enterprise content operations, governance, workflow orchestration, personalization, and measurement.

• The stock depends on whether the enterprise “orchestration loop” becomes visible and measurable: bulls need proof through organic ARR stabilization, freemium conversion data, sustained >20% enterprise CXO growth, margin resilience, and better segment transparency; until then, the market will keep valuing Adobe as a disrupted creative-tools business rather than a high-growth enterprise platform.

In August 2025 we wrote about Adobe at $353 and laid out three scenarios. Our most probable outcome, Disruption Realized, had a floor of $220. Ten months later, the stock is $206.

Most of the product and enterprise signposts we said to track have improved. AI-first ARR quadrupled from roughly $125 million to over $500 million. GenStudio ARR grew more than 25%. AEP and its native apps grew subscription revenue above 30%. Enterprise customers with more than $10 million in ARR grew over 20%. Firefly ARR is approaching $300 million. Traffic to Adobe.com grew 40-50%. Creative freemium monthly active users nearly doubled. In our September “Pixels to Pipelines” piece, we argued Adobe was becoming enterprise content infrastructure. The enterprise metrics have since exceeded the trajectory we needed to see.

The monetization signpost did not improve. Organic ARR quality weakened. The market’s most important measure, whether Adobe can convert AI traction into durable, direct-to-paid recurring revenue, moved in the wrong direction.

We were right about the fundamentals and wrong about the stock, because we did not anticipate that Adobe would voluntarily sacrifice near-term ARR to rebuild its acquisition funnel, and that the market would read that sacrifice as confirmation of structural weakness rather than evidence of strategic adaptation.

This is not a quarterly miss. It is a framework error. We were measuring what Adobe does. The market was measuring what Adobe controls. Those turned out to be different things. The distinction between the health of a business and the durability of its moat is the entire Adobe story right now, and it may be the most important structural lesson in technology investing: your numbers can improve while your structural power diminishes, and the market will price the power, not the numbers.

The Algorithm That Broke

For forty years, Adobe ran one of the most elegant compounding algorithms in software. Creative intent led to skill investment. Skill investment created switching costs, learning Photoshop took years, and those years became a professional credential that made switching irrational. Switching costs created pricing power. Pricing power funded innovation. Innovation rewarded further skill investment. The cycle repeated for decades. It justified a premium multiple because each turn of the wheel made the next turn more certain.

AI broke the first link. When a user can describe what they want in a sentence and receive usable output in seconds, the skill investment that was Adobe’s deepest moat becomes optional. Not worthless, a retoucher finishing a luxury campaign still needs Photoshop’s precision. But optional for a rapidly expanding share of creative tasks. And when the skill barrier falls, everything downstream, the switching costs, the pricing power, the compounding certainty, weakens with it.

This is why the stock re-rated from 50x earnings to 8x while revenue grew 51% and EPS nearly doubled. The market is not saying Adobe’s current earnings are wrong. It is saying the mechanism that made future earnings predictable has been structurally impaired.

Critically, this is not a demand problem. More people want Adobe’s capabilities than at any point in the company’s history. Adobe.com traffic grew 40-50% in Q2. Creative freemium MAU went from 50 million to 90 million. Acrobat MAU hit 850 million. But users are arriving through a different door, intent-based search queries, LLM conversations, task-completion prompts, and that door does not have a cash register at the entrance.

Q2 confirmed this. Management said explicitly that traditional direct-to-paid journeys “may not always fulfill visitor intent,” that Adobe would rebalance traffic toward friction-free freemium experiences, and that this shift would cost near-term ARR. They deferred planned Creative Cloud price increases. They guided FY26 ARR growth to 10.2%, but that target now includes $480 million of acquired Semrush ARR, meaning organic ARR growth was cut roughly 200 basis points versus prior expectations.

The stock fell 16% in a week. Twenty-two analysts cut their price targets. Three downgraded. The market heard what it feared most: Adobe’s old monetization machine no longer maps cleanly onto how customers discover and adopt software in the AI era.

The Abundance Paradox

This is where most analyses stop, at the disruption narrative. But the disruption narrative, while partially correct, misses the structural paradox at the center of Adobe’s situation.

When creative output was scarce, the tool that produced it captured the value. Adobe owned the tool. When creative output becomes abundant, and AI is making it abundant at an extraordinary rate, value migrates away from creation and toward everything that happens after: editing, refinement, brand governance, rights management, workflow orchestration, personalization, distribution, and measurement. A marketing team that once produced fifty assets a month can now generate five thousand variants. The bottleneck is no longer making the content. It is controlling it.

This is the abundance paradox: the same AI force that weakens Adobe as a creation-tool provider strengthens Adobe as a content-operations platform. Both effects are real. Both are accelerating. The question, the only question that matters for the stock, is which one is larger.

Q2 provided evidence for both simultaneously.

The commoditization loop is visible: management deferred pricing it had planned to take, rebuilt the acquisition funnel around freemium, and acknowledged rising customer acquisition costs. The uncomfortable arithmetic, even allowing for imprecision in how ARR categories are defined, is that AI-first ARR is growing inside a pool of total net-new ARR that is getting smaller, not larger. The new engine is building. It is not yet building fast enough to offset the pressure on the traditional engine. That is the single hardest data point for bulls to dismiss.

The orchestration flywheel is also visible: AEP and native apps grew subscription revenue above 30%. GenStudio ARR grew more than 25%. Customer Experience Orchestration AI-first ARR grew fourfold. Over 80% of enterprise customers adopted agentic capabilities. Fifteen hundred enterprises entered trials for agentic web products. CX Enterprise Coworker reached general availability with 150 enterprises in early adoption. Every major agency holding company standardized on Adobe. Forward-deployed engineering services grew 60% quarter-over-quarter.

The market is pricing the first loop because it is visible and frightening. It is largely ignoring the second because Adobe consolidated its reporting segments into a single unit, at the worst possible time, making it impossible for analysts to value the enterprise platform independently.

Two Businesses, One Stock

The abundance paradox creates two structurally distinct businesses inside one stock.

The first is creative tools and consumer productivity, Creative Cloud subscriptions, Firefly standalone, Express, Acrobat individual plans, and the freemium funnel. This business is not dying. Revenue grew double digits. Professional users still need precision editing, advanced workflows, and production-grade output. But the moat has weakened. Growth will come from volume and consumption rather than pricing power. It deserves a mature software multiple.

The second is enterprise content operations, AEP, GenStudio, AEM, agentic web products, Brand Intelligence, CX Enterprise, Firefly Enterprise, and Semrush. This business has switching costs that AI cannot erode because they are not skill-based. They are built on customer data, workflow configurations, brand rules, agency integrations, compliance frameworks, and years of accumulated intelligence. Every workflow that runs through the platform makes it harder to leave.

The market is right to worry about the first business. It may be wrong to apply that worry to all of Adobe.

The IBM comparison that haunts Adobe investors applies to the first business but not the second. IBM’s core market was shrinking, fewer mainframe workloads each year. Adobe’s creative market is expanding, more content than ever. And IBM never had a 30%-growth enterprise platform embedded in its financials. The correct frame is not “Adobe becomes IBM.” It is that Adobe contains an IBM-like mature franchise and something resembling a high-growth enterprise platform, and the market is pricing the entire company as if only the former exists.

The Freemium Bet and the Bridge to FY28

The freemium pivot is where the argument gets hardest, because it is simultaneously the right structural response and the most damaging near-term signal.

Management’s logic is straightforward. Users arriving through intent-based search want to complete a task, not buy a subscription. Routing them into the product for free, building engagement, and placing paywalls after habit forms should produce higher lifetime value than forcing them into a purchase funnel they will abandon. Early signals support this: freemium-to-paid converts show higher engagement and credit consumption than direct-to-paid cohorts. Firefly ARR grew roughly 50% quarter-over-quarter through apps and credit packs.

But management has disclosed no conversion rate, no LTV/CAC data, and no cohort analysis. They have said the payback “plays out over 2027.” Asking the market to fund a multi-quarter bet on faith, during a CEO and CFO vacancy, inside a hostile sector narrative, with organic ARR decelerating, is an extraordinary ask. The stock price reflects the uncertainty premium of that ask.

The Acrobat Reader analogy management drew is instructive but imperfect. Reader succeeded because PDF had network effects, every shared document created another user. Firefly-generated images are format-agnostic PNGs. The freemium funnel must rely on habit and product quality alone, which is a slower and less certain monetization path.

For the aggregate ARR trajectory to inflect, AI-first ARR needs to add more annually than the traditional book loses. At current trajectories, that crossover is roughly two years away. The market is pricing Adobe for the gap between now and then.

Three Futures

We see three outcomes over the next three years.

The orchestration loop wins. Freemium conversion works. Enterprise CXO sustains above 20% growth. Organic ARR stabilizes by FY28. A new CEO provides segment transparency, and the market sees the enterprise platform clearly. Revenue compounds at 10-11% through FY29 to roughly $35 billion. Non-GAAP margins hold at 45-46%. EPS reaches approximately $36. The market re-rates to 15x forward earnings as the enterprise platform earns its own premium. Stock price in three years: approximately $540.

Contested repositioning. Freemium converts but slowly. Enterprise remains strong but not dominant enough alone to re-rate the stock. Creative Cloud pricing power partially recovers when line optimizations are reintroduced. Revenue compounds at 8-9% to roughly $33 billion. Margins hold at 43-45%. EPS reaches approximately $32. The multiple stays at 10-12x. Stock price in three years: approximately $350.

Permanent de-rating. Freemium MAU grows but conversion stalls. Traditional Creative Cloud erosion accelerates. AI-native tools absorb the mid-market. Revenue compounds at only 5-6% to roughly $30 billion. Margins drift to 41-43%. EPS reaches approximately $27. The multiple compresses to 7-9x. Stock price in three years: approximately $220. This is not extinction. It is the real IBM risk, a profitable company whose category-defining role has moved elsewhere.

The asymmetry is better than the sentiment suggests, but the catalyst path is long.

What Changes Our Mind

Six signposts will determine which future arrives.

Organic net-new ARR excluding Semrush: inflecting positive by Q1 FY27 confirms the thesis; declining three consecutive quarters impairs it. Freemium conversion disclosure with quantitative rigor: the single most important missing datapoint. Enterprise CXO growth sustaining above 20%: the orchestration flywheel’s proof. A CEO with product and AI credibility named before Q4 FY26: the narrative catalyst. Non-GAAP margins holding above 44%: confirms AI costs are manageable. Creative Cloud line optimizations reintroduced in FY27 without churn stress: evidence that professional pricing power remains intact.

We will revisit the thesis after the CEO appointment and publish a full update with Q3 results.

Our prior framework measured the right things and missed the structural shift underneath. Adobe’s business is not broken. Its old compounding algorithm is. The market is no longer asking whether Adobe has AI traction. It is asking whether the orchestration loop, workflow, governance, distribution, measurement, can replace the skill loop that drove forty years of compounding.

That is the only question that matters now. At $206, the market’s answer is a clear no. The enterprise evidence, on balance, suggests the answer is more nuanced than that, but Adobe has not yet given the market the data it needs to see it. Until the orchestration loop becomes visible, measurable, and separately valued, the abundance paradox will remain unresolved, and the stock will remain hostage to a narrative about the business it used to be rather than the platform it is becoming.

Subscribe now

• Our prior Rubrik thesis was that the market was making a category error: Rubrik was not merely backup; it was becoming state infrastructure.

• Q1 confirmed the recovery platform is real and self-funding, while Investor Day sharpened the next ambition: recovery itself becomes agentic.

• The debate has moved again. Identity now has numbers. Agentic recovery has architecture. Agent Cloud still needs revenue disclosure.

When we first wrote about Rubrik, the question was whether this was infrastructure, illusion, or intelligent evolution. That was the right question at the time. Rubrik had the growth, the customer expansion, and the beginnings of free cash flow, but it was still easy to dismiss the story as an expensive backup company wearing a cybersecurity costume.

Then the company changed the question. Q2 showed that the cyber resilience engine could fund AI ambition. Q3 was the janitor-becomes-bodyguard quarter: the same system that once cleaned up after failure was being sold as the way to survive when attackers were already inside. Q4 sharpened the frame again. The market thought backup; Rubrik looked more like state infrastructure, a time-indexed, immutable map of enterprise data, identity, SaaS applications, and configurations.

Q1 did not overturn that thesis. It advanced it but also made it more demanding.

The question is no longer whether Rubrik can escape backup. It largely has. The question now is whether owning the enterprise’s last-known-good state becomes more valuable when failure moves from human speed to AI speed.

That is the fundamental question for this quarter:

Does Rubrik own the enterprise’s last-known-good state strongly enough that AI agents turn recovery from insurance into infrastructure?

My answer: the recovery platform is proven, identity is becoming measurable, and agentic recovery now has a credible architecture. But Agent Cloud is still not yet a disclosed revenue engine. That distinction matters.

The Question Changed Because Failure Got Faster

Backup is what recovery looked like when failure was slow.
A server failed. A file was deleted. A database needed restoring. The task was operational, dull, and usually invisible. The best backup product was the one you forgot existed until something broke.

Cyber resilience changed the category. The threat was no longer accidental loss; it was malicious compromise. Ransomware did not merely delete files. It encrypted data, corrupted systems, compromised identities, and targeted the recovery environment itself. In that world, the scarce capability was not “do I have a copy?” It was “do I know what clean looked like before compromise, and can I return the business to that state?”

AI agents change the category again.

Agents do not merely assist humans. They increasingly act inside enterprise workflows. They assume identities, access sensitive data, write code, modify systems, approve actions, trigger refunds, delete files, and interact across SaaS, cloud, and internal applications. The optimistic version is 10x or 100x productivity. The darker version is failure at machine speed.

This is why Rubrik’s Investor Day mattered. Management was not simply saying, “we have an AI product.” The claim was more radical: human-speed recovery cannot respond to AI-speed compromise. If agents are going to act faster than humans can supervise, then the recovery system also must become more autonomous. Rubrik is trying to make Rubrik itself an agent.

That is the conceptual leap.

Backup restored data. Cyber resilience restored the business after an attack. Agentic cyber resilience is supposed to restore the business when both attackers and operators are machines.

The asset is no longer just the wall. It is the undo button, and, increasingly, the agent that knows when and how to press it.

From Map to Agent

Rubrik’s original asset was the map.

The company protected data, but the more important by-product was context: where data lived, who had access, how it changed, what was sensitive, what was infected, and what clean looked like before the event. Investor Day gave this a more precise technical language: self-describing data. Rubrik does not simply store copies. It captures data and metadata together, over time, across applications, identities, and environments.

That time-series map is what made the Q4 “state infrastructure” thesis work. If an attacker compromises Active Directory, endpoint detection can tell you something happened. It cannot reconstruct the directory as it existed before the attacker created backdoors, escalated privileges, and changed access rights. Rubrik’s claim is that it can, because it has the historical state.

But a map alone is not enough.

The next layer is pre-computation. Rubrik’s Preemptive Recovery Engine does work in peacetime so that recovery is possible in wartime. It identifies clean recovery points, maps dependencies, scopes compromised identities, assesses sensitive data impact, and builds recovery plans before the crisis arrives. This is important because the hardest part of recovery is not the mechanical restore. It is knowing what to restore, in what order, and from which clean point.

The final layer is agentic orchestration. Rubrik AI and Rubrik Agent Cloud are the action layer: monitoring, recommending, planning, guarding, and, where allowed, executing recovery with humans supervising the highest-judgment steps.

That is the real architecture:

Rubrik’s old ambition was to know what the enterprise used to be. Its new ambition is to put the enterprise back together before humans can finish the post-mortem.

That is a big claim. The question is how much of it is already real.

The Flywheel We Can Measure

Q1 was financially clean.

Revenue grew 39% to $387 million. Subscription ARR grew 32% to $1.57 billion. Revenue excluding material rights grew 43%. Non-GAAP gross margin reached roughly 83%. Free cash flow was $74 million, or 19% of revenue. Subscription ARR contribution margin improved to 13.2%.

Those are good numbers, but the point is not that Rubrik beat estimates. High-quality software companies beat estimates all the time. The point is what the numbers say about the mechanism.

The recovery graph is producing software margins and cash.

A backup vendor should not have 83% gross margins if it is being paid mainly for capacity. A fragile growth company should not be producing nearly 20% free cash flow margin while still growing ARR above 30%. A one-product company should not be steadily moving customers across data, cloud, SaaS, security, and identity surfaces.

The flywheel is not a classical network effect. One Rubrik customer does not make the product directly better for another customer. The effect is inside the enterprise. Every new workload, SaaS application, identity provider, and agentic action connected to Rubrik makes the state map more complete. A more complete map makes recovery more useful. More useful recovery increases the customer’s willingness to adopt the next module. The value is not users attracting users; it is state attracting more state.

Investor Day made this concrete. Management described faster product ramps over time: cloud took years to reach scale, M365 ramped faster, data security faster still, and identity reached $50 million of ARR in just over a year. That is what a multi-product platform is supposed to look like: each new product launches into an installed context, not a cold market.

This is not yet Salesforce. It is not yet ServiceNow. But it is no longer platform theater.

Identity Has Numbers; Agentic Recovery Has Architecture

Identity is the bridge from data state to enterprise state.

That matters because modern attacks are increasingly identity-led. If you restore files but leave corrupted access rights, rogue admin accounts, or compromised identity systems intact, you have not recovered the business. You have restored the crime scene.

This is why the identity business matters disproportionately. It is not merely an add-on product. It extends Rubrik’s map from what the enterprise owns to who can touch it. That is the difference between data recovery and business-state recovery.

Q1 gave identity a real scoreboard: more than $50 million of subscription ARR and strong sequential growth. That makes identity the first measurable proof that Rubrik can extend beyond the original backup surface. It also makes the prior thesis less speculative. The map is widening.

Agentic recovery is at a different stage.

Rubrik Agent Cloud, SAGE, Agent Rewind, Gemini integrations, Microsoft Defender integration, Anthropic / Glasswing, Rubrik AI, autonomous Minimum Viable Business recovery, these are strategically coherent. They fit the architecture. They answer a real question: how do enterprises deploy agents safely when those agents can act on sensitive data with real authority?

But Agent Cloud does not yet have separate ARR disclosure. We have architecture, examples, product language, and early customer evidence. We do not yet have a financial line item.

That is not a criticism. It is discipline.

Identity has crossed from narrative to metric. Agentic recovery has crossed from slideware to architecture. Agent Cloud has not yet crossed from architecture to disclosed revenue.

Why the Stock Did Not Care More

This is why the stock reaction was not irrational.

Rubrik beat the visible numbers. It did not obviously clear the invisible ones.

For a company with unanimous sell-side enthusiasm, the market is not asking whether Rubrik is good. It already knows Rubrik is good. The marginal buyer is asking whether the flywheel is accelerating fast enough to justify the category expansion.

Q1 net new ARR was about $103 million, a record Q1, but below the roughly $115 million Q4 record. That is not a fundamental problem; seasonality matters, and management expects the year to be back-half weighted. But the stock was not trading a normal consensus beat. It was trading acceleration.

The guide raise was good, but not a dramatic reset. The AI architecture became more credible, but not yet financially measurable. The business confirmed quality, but the stock needed evidence of a new curve.

That is the variant perception.

The Street sees beat-and-raise.
The company sees agentic cyber resilience.
The market sees a high-quality compounder that still needs to prove the AI layer can become financial, not just strategic.

Rubrik cleared the published bar. It did not yet clearly clear the acceleration bar.

Three Futures

The stock is not simple because the business is no longer simple. The bear case is not that Rubrik is bad. It is that Rubrik remains merely excellent software. The bull case requires a category shift from state infrastructure to agentic recovery infrastructure.

The most important risk is not Veeam or Commvault. Those are old-category competitors. The cleaner strategic risk is Microsoft, Google, or AWS making “good enough” native recovery and agent governance part of the cloud platform. Rubrik’s answer is vendor neutrality and cross-environment context. That answer is compelling. It is also why the clouds are both partners and long-term threats.

What We Need to See

The next few quarters should be judged on a smaller set of questions.

First, net new ARR needs to validate the back-half weighting. Reacceleration toward or above the Q4 record would keep the flywheel thesis intact.

Second, identity ARR needs to compound from $50 million toward the next obvious milestone. If identity becomes a $100 million-plus business quickly, the second-engine argument becomes hard to dismiss.

Third, Agent Cloud needs disclosure. ARR, paid customers, attach rate, contribution to net new ARR, any of these would move the thesis from architecture to numbers.

Fourth, multi-product adoption and Rubrik Flex need to show that the platform is becoming easier to buy, not just broader to sell.

Fifth, free cash flow margins need to remain strong while the company invests. The core engine must fund the next act.

Finally, watch the hyperscalers. Partnerships are positive. Native “good enough” cross-workload recovery or agent rollback would be the structural warning sign.

Rubrik has proven it is not simply backup. It has increasingly proven it is state infrastructure. Q1 and Investor Day made the next ambition clearer: to become the agentic recovery layer for the AI enterprise.

That thesis is more plausible today than it was last quarter.

It is still not fully proven.

Subscribe now

• Q4 strengthened the company but made the stock less forgiving: Credo proved it is more than an AI cable supplier, with FY26 revenue more than tripling, Q4 revenue up 157%, gross margin at 68.3%, and FY27 revenue guided above 80% growth but the market now understands the reliability-layer thesis, so the stock needs proof, not discovery.

• The key debate has moved from copper to optics: AECs already validated Credo’s reliability advantage in AI clusters; the new question is whether that advantage can travel into optical DSPs, silicon photonics PICs, and ZeroFlap Optics, where management expects more than $600M of FY27 revenue.

• Gross margin is now the referendum on the whole thesis: high-60s margins would suggest Credo is selling an ownable system-level reliability layer; mid-60s would make it a premium component leader; low-60s would imply optics, modularity, or customer bargaining power is diluting the moat.

There is a particular discomfort in watching the market agree with you.

For the past year, our Credo argument has taken different forms, but the underlying claim has been the same: the market was miscategorizing Credo as a cable company riding an AI capex wave, when the more interesting possibility was that it was building something closer to infrastructure, a reliability layer, sold through hardware, that hyperscalers were beginning to standardize on.

Each quarter made that argument harder to dismiss. The margins were too high. The customer adoption was too broad. The language around telemetry, diagnostics, and ZeroFlap was too deliberate. The company looked like a component vendor but behaved economically like something stickier.

Q4 FY26 was the quarter where the market stopped ignoring that possibility.

Credo reported revenue of $437 million, up 157% year over year. Full-year revenue more than tripled to $1.3 billion. Non-GAAP gross margin was 68.3%. Management guided fiscal 2027 revenue growth above 80%. And the stock fell.

That sounds contradictory. It is not. The first few times Credo fell on good numbers; the opportunity was that the market misunderstood the company. This time is different. The stock did not fall because investors still thought Credo was merely a cable supplier. It fell because they now understand enough of the reliability thesis to demand proof of the next claim.

Closing a mispricing is a one-time event. We have spent much of it.

What Q4 Actually Settled

The question coming into this quarter was not whether AECs work. That is settled. Credo’s active electrical cables became the wedge because AI clusters do not merely need bandwidth; they need stable links that help preserve GPU utilization and reduce cluster instability.

The harder question, which we framed last quarter, was whether that reliability advantage could travel from copper into optics. Copper has a real role at short distances, but the architecture is moving. Higher bandwidth, 1.6T ports, 200G lanes, scale-up networks, and eventually CPO or NPO all point toward more optical content. If Credo’s advantage was copper-specific, the thesis had an expiration date.

Q4 moved that debate forward. Management now expects more than $600 million of FY27 revenue from the optical portfolio, with optical DSPs, silicon photonics PICs, and ZeroFlap Optics each contributing more than $100 million. That is no longer a side project. It is the plan of record.

But we should be precise. Commercial portability is now in the model. Economic portability is still the test.

Customers appear willing to adopt Credo’s optical portfolio. That does not yet prove that optical carries the same margin structure, workflow lock-in, or platform value as AECs. The question has changed from “can reliability travel?” to “can reliability remain ownable once it arrives?”

That is the new fundamental question.

Integration Wins This Round

I think the answer, for now, is yes.

Credo’s strategy is integration. Own the SerDes. Own the system. Own the firmware. Own the telemetry. Own more of the optical stack. Deliver reliability as a system-level outcome, not a component specification.

The modular counterattack is equally clear. Merchant DSPs, reference designs, partner ecosystems, cable assemblers, and switch-level software all exist to make reliability good enough while pushing economics back toward component margins. Marvell’s Golden Cable initiative is the cleanest expression of that strategy: broaden the AEC ecosystem, provide validated architectures and firmware, and let partners scale around merchant silicon.

This is not just Credo versus Marvell. It is integration versus modularity.

The history of technology says modularity eventually wins when the problem becomes understood well enough to standardize. That is the Cisco cautionary tale. Integrated systems dominate when performance is not good enough; modular ecosystems attack once the interfaces settle.

The question is where AI-network reliability sits on that curve. I think it is still early. Clusters are getting larger, not simpler. The cost of link instability is rising, not falling. Customers are still optimizing for bring-up time, uptime, power efficiency, and utilization, not merely lowest-cost interconnect.

That is the kind of environment where integration should win.

DustPhotonics fits this view. Credo did not simply buy optical revenue. It bought deeper visibility into the optical link. If silicon photonics PICs let Credo combine DSP, optics, telemetry, diagnostics, and system-level optimization more tightly, then DustPhotonics is not a departure from the original thesis. It is the attempt to make the original thesis survive light.

That does not mean Credo wins forever. Integrated advantage is often strongest just before the modular wave begins to matter. The view is narrower: Credo wins this round, because AI-network reliability is not yet good enough to modularize. The day it becomes good enough, the thesis changes.

The Metric We Still Cannot See

There is one place where we need to be more disciplined than before: PILOT.

Our earlier argument leaned on PILOT as a possible data flywheel. Every deployment could become a sensor. Every link could generate failure-mode information. Every customer workflow built around Credo’s telemetry could make switching harder.

The logic still makes sense. But the disclosure has not caught up. Management continues to talk about telemetry, diagnostics, monitoring, SDK integration, and autonomous mitigation. What it does not provide is attach rate, pricing, separate software economics, customer workflow evidence, or proof that PILOT has become an operational dependency.

That absence matters.

It does not break the thesis. Physical-layer telemetry embedded in the product is still harder to replicate than dashboard software sitting above the network. But we should not put too much weight on a data moat we cannot measure. The visible moat today is not standalone software. It is architecture: integrated reliability delivered through hardware, firmware, optics, telemetry, and customer-specific execution.

That is still valuable. It is just less clean than the “software-like platform inside a cable” version of the story.

Gross Margin Is the Referendum

This is why gross margin is the most important number now.

Not because gross margin is an accounting line. Because it is where integration versus modularity becomes visible.

If Credo is truly selling reliability as a scarce system-level outcome, then high margins should survive the optical transition. If optics is merely a larger but more competitive market, margins should drift lower as component costs, customer bargaining power, dual sourcing, and modular ecosystems show up.

The framework is simple:

A 68% gross-margin Credo and a 64% gross-margin Credo can both be good businesses. They are not the same business. The first suggests an ownable reliability layer. The second suggests a high-quality semiconductor supplier with a strong product cycle.

That difference matters because the stock is no longer cheap on misunderstanding.

Three Futures

The three-year outcome is less about forecasting every quarter than deciding which architecture wins.

The bull case is now explicitly above consensus. That is important. The Street already models a large FY29 business. The upside is not merely that revenue reaches scale; it is that scale comes with high-60s margins and platform-like durability.

The base case is also not bad. A company doing more than $4 billion of revenue with mid-60s gross margins would still be exceptional. It just would not deserve the same multiple as a reliability control point.

The bear case is where the discipline comes in. If the optical ramp slips or arrives with lower margins, the downside is not just an estimate cut. It is a category reset.

What We Track

The signposts are now clear.

First, FY27 optical revenue. Above $600 million confirms the handoff; below $500 million raises timing concerns.

Second, the Q3 and Q4 FY27 step-up. The first half is not the story. The second-half optical ramp is.

Third, gross margin. High-60s supports the platform case. Mid-60s lowers the multiple. Low-60s impairs the thesis.

Fourth, PILOT evidence. We need more than language: attach rates, customer workflow references, SDK adoption, or proof of operational dependency.

Fifth, customer breadth. Top-two concentration must keep falling, and neoclouds need to become more than narrative support.

Sixth, CPO/NPO timing. Delays extend Credo’s runway. Acceleration without Credo participation compresses it.

Where We Stand

Our view on Credo the company is stronger. Q4 makes the reliability thesis more concrete, not less. The company is no longer merely an AEC winner. It is attempting to extend managed reliability across copper, optical, and eventually near-package connectivity.

Our view on the stock is more disciplined. Credo is no longer deeply misunderstood. The market has begun to understand the story, and that means investors are no longer being paid simply to be early. They are being asked to be right.

That is the uncomfortable but healthy update.

We were early in asking the right question. Q4 gives us more confidence that the question matters. But the answer is still ahead.

Credo has moved the proof from copper to light. Now the margin line will tell us whether the toll road followed.

Subscribe now

• Zscaler’s Q3 did not break the business, but it broke the re-rating case: revenue and ARR still grew 25%, margins hit a record, and platform metrics like Z-Flex, Zero Trust Everywhere, AI Protect, and non-seat usage continued scaling, but the FY27 growth outlook reset the stock from “hidden compounder” to “prove-it story.”

• The key thesis shift is from interception to anticipation: Zscaler remains highly valuable where it is strongest, inline traffic inspection, access enforcement, app invisibility, and lateral-movement reduction, but AI may move the highest-value security decisions upstream into code, identity, endpoint, browser, cloud configuration, and agent governance.

• The stock now depends on whether Zscaler can move upstream, not just monetize the wire: the bear case is maturity, the base case is a durable high-teens compounder, and the bull case requires proof that Symmetry, AI Protect, non-seat usage, and Zero Trust Everywhere can turn Zscaler into an AI-era security control plane rather than a utility-like SASE platform.

“Zscaler shares tumble as much as 28%... after the security software company gave a fourth-quarter revenue forecast that was weaker than expected.”

That was the market’s version of the story. A clean enough quarter, an imperfect guide, a violent stock reaction.
I do not think that is quite right. Zscaler did not fall because Q3 was bad. Revenue grew 25%. ARR grew 25%. Operating margin reached an all-time high. Z-Flex kept scaling. Zero Trust Everywhere added customers. AI Protect crossed $100 million of bookings. This was not a broken print.
The harder issue is that Q3 made my prior framework incomplete.

For the last few quarters, I have argued that Zscaler was being priced as decelerating SaaS while becoming security infrastructure. The logic was reasonable. Z-Flex was turning point-product purchases into platform commitments. Non-seat metered usage suggested a shift from headcount-based SaaS to traffic-based infrastructure. Zero Trust Everywhere expanded the platform from users to branches and workloads. AI agents seemed to make Zscaler’s inline architecture more valuable because agents do not buy seats; they create traffic, identities, access requests, and policy decisions.

That thesis was not built on bad facts. The facts are still there.

What was wrong was the leap from those facts to inevitability. I asked whether Zscaler’s architecture was valuable. It is. I should have asked whether the most valuable security decisions were still happening at the point where Zscaler is strongest.

That is the question Q3 forced.

From Interception to Anticipation

Cybersecurity has historically been an interception business.

Firewalls intercepted packets. Antivirus intercepted files. EDR intercepted processes. Zscaler intercepted connections. The stack was organized around catching threats at the moment of action.

Zscaler’s architecture was an elegant answer to that world. Instead of exposing applications to the internet and defending the perimeter around them, Zscaler hid applications behind the Zero Trust Exchange, connected users to specific applications rather than networks, inspected traffic inline, and reduced lateral movement. In a world where the decisive moment was the connection, being in the path of the connection was a powerful place to sit.

AI changes the geography.

The obvious effect of frontier models like Mythos is speed: vulnerabilities can be discovered faster, exploits can be reasoned through faster, and attackers can move faster. That is the point management emphasized, and it is real. Jay Chaudhry’s answer is that enterprises cannot patch fast enough, so the better defense is to make applications invisible and limit lateral movement.

That is still right.

But there is another implication. If frontier models help defenders find vulnerabilities in their own codebases, configurations, endpoints, identities, and deployment environments before attackers do, then the marginal value in security begins to move upstream. The decision is no longer only made at the moment of connection. It is made earlier: in the codebase, the identity graph, the endpoint, the browser, the cloud configuration, and the deployment pipeline.

This does not make Zscaler irrelevant. It does make the bull case harder.

In the interception era, the inline exchange looked like the natural control plane. In the anticipation era, the control plane may be wherever the enterprise can understand and reduce risk before traffic exists.

That is a different game.

Glasswing Validates the Problem, Not the Winner

Project Glasswing matters, but not in the simplistic way I first wanted it to matter.

Zscaler has access to Claude Mythos Preview through Glasswing and is using it to harden its own systems and improve security for customers. That is good. It validates the urgency of the problem Zscaler has been describing.

But access to Mythos is not, by itself, the edge.

Microsoft can attach frontier-model security to Entra, Defender, Windows, Edge, Azure, GitHub, and Microsoft 365. CrowdStrike can attach it to endpoint telemetry, identity protection, managed detection and response, and incident response workflows. Palo Alto can attach it to firewalls, SASE, Prisma Cloud, Cortex, XSIAM, and now CyberArk’s identity layer. Zscaler can attach it to the Zero Trust Exchange.

That is a strong control point. It is not obviously the default one.

The right lesson from Glasswing is not that Zscaler lost. The right lesson is that AI-security urgency is broad, and the companies best positioned to monetize it will be the ones that convert frontier-model intelligence into closed-loop workflow: find the issue, understand the blast radius, enforce the policy, remediate the weakness, and prove the risk was reduced.

Zscaler has part of that loop. It sees traffic. It enforces policy. It hides applications. It reduces lateral movement. With Symmetry Systems, it is adding an access graph: which identities are touching which applications, models, and data sources. That is the right move.

But it is also a tell. Zscaler is trying to move upstream from the wire because the value is moving upstream too.

What Q3 Still Proved

It would be too easy, and too wrong, to turn this into a clean dismissal.

Zscaler’s platform evidence improved in Q3.

Zero Trust Everywhere customers rose to more than 700, up from over 550 in Q2. Non-seat-based metered usage was just over 30% of new ACV, with ARR tied to those offerings growing more than 100% year over year. Z-Flex generated just over $480 million of TCV in the quarter, up more than 60% sequentially, with more than $1 billion over the last twelve months and an average four-year term. Data Security crossed $500 million of ARR, growing more than 30%. Zero Trust Branch ARR roughly tripled year over year. AI Protect bookings crossed $100 million over the past twelve months.

These are not filler numbers. They are the best evidence that Zscaler is not merely a mature seat-based SASE company.

The old bull case said Zscaler could escape the headcount model. A user-seat business grows with employees. A traffic-and-policy business grows with applications, workloads, branches, data, agents, prompts, responses, and machine-to-machine interactions. Q3 provided more evidence that this shift is happening.

The problem is that it is not yet large enough to control the total growth algorithm.

That is where the thesis needs to be updated, not abandoned.

The Model Caught Up With the Structure

If the marginal security dollar is moving upstream, then a network-layer company with a maturing core should look exactly like Zscaler just looked: strategically relevant, commercially active, operationally strong, but financially normalizing.

That is what Q3 showed.

Revenue and ARR still grew 25%, but Q4 revenue guidance implies roughly 22% growth. Full-year FY26 ARR guidance implies net-new ARR growth excluding Red Canary of roughly 9.5%. Free cash flow margin guidance was cut to 22.8–23.3%, largely because CapEx is moving higher. Two sales leaders departed at the end of Q3. Most importantly, management’s early view for FY27 is only 16–17% ARR and revenue growth.

The guide was not the diagnosis. It was the diagnosis showing up in the model.

A 25% ARR grower with visible platform expansion can still ask investors to underwrite a premium compounder. A 16–17% forward grower with higher CapEx and sales-transition uncertainty is something else. It may still be high quality. It may still be strategically important. But it no longer gets paid for hidden torque unless that torque reaches reported growth.

That is the difference between Q1/Q2 and Q3.

In Q1 and Q2, I could argue that the market was staring at the speedometer while the engine-built torque. Q3 says the road changed. The engine may still be there, but the hill is steeper than I thought.

Retiring the ServiceNow Analogy

The ServiceNow analogy should be retired.

I used it because the pattern looked similar: decelerating growth, multiple compression, switching costs compounding underneath, and a market too impatient to recognize an infrastructure asset in formation. ServiceNow eventually re-rated because it became the workflow operating system for enterprise IT. Being the operating system is the most valuable position in a technology stack.

Zscaler is not obviously becoming the operating system of cybersecurity.

It may be becoming something more like the cardiovascular system: essential, durable, difficult to remove, economically attractive, but not necessarily where the highest-value decisions originate. The brain may be moving upstream, toward identity, endpoint, code, browser, cloud configuration, and agent governance.

That distinction matters for the multiple.

The updated interpretation is therefore not that Zscaler is broken. It is that Zscaler is more durable than bears think and less re-ratable than bulls hoped.

Variant Perception, Revised

The old variant perception was: The market sees decelerating SaaS. We see security infrastructure.

The updated variant is narrower: The market is right to remove the premium multiple but may be wrong to assume the platform transition has failed.

That is less exciting. It is also more honest.

At roughly $125, most of the obvious de-rating has already happened. The stock is not priced like a category-defining 25% compounder anymore. It is priced more like a high-teens software company with good margins, a strong installed base, and a credible but unproven path into AI-era security.

For underwriting discipline, the three-year setup now looks different.

In the bear case, Zscaler is a mature SASE-plus platform growing revenue at 12–14% through FY29, with free cash flow margins around 22–23% and a 3.5–4.0x sales multiple. That supports a stock around $105–125. This is not a disaster case; it is a maturity case.

In the base case, Zscaler is a reset compounder. Revenue compounds at 16–17%, free cash flow margin settles around 25–26%, and the market pays 5.0–5.5x sales for a durable but no longer exceptional security platform. That gets you roughly $165–190.

In the bull case, Zscaler proves it can move upstream. Non-seat usage scales, Symmetry’s access graph matters, AI-agent traffic monetizes, and Zero Trust Everywhere becomes a default architecture rather than a bundle. Revenue compounds at 20–22%, free cash flow margin reaches 28–30%, and the multiple expands to 7–8x sales. That supports $245–285.

The important change is not the arithmetic. It is the migration of assumptions.

What used to be my base case is now the bull case. The base case is a good company at a fair price. The bear case is maturity. The bull case requires Zscaler to prove it can participate in anticipation, not just interception.

What Would Change This View

The thesis is still alive, but the tests need to be tighter.

The first test is growth. FY27 ARR and revenue growth need to move back above 18–19%, and organic net-new ARR needs to recover toward the mid-teens. If Zscaler stays at 16–17% total growth and around 10% organic net-new ARR growth, the market is right to treat it as mature software.

The second test is the consumption transition. Non-seat ACV needs to stay above 30% and move toward 35%; Z-Flex needs to sustain quarterly TCV above $450–500 million; Zero Trust Everywhere needs to keep adding customers at a strong pace. That is the proof that Zscaler is escaping the human-seat model.

The third test is upstream relevance. Symmetry and AI Protect need to become more than product slides. The signpost is whether Zscaler can move from “we inspect traffic” to “we map identity-to-data risk and enforce policy before the connection matters.”

Finally, free cash flow margin needs to stabilize above 24–25% despite higher infrastructure spending. If growth slows and cash quality falls, the stock does not deserve a re-rating.

Upstream of the Wire

Zscaler’s strategic thesis did not break in Q3. Its grace period did.

The company still matters. Interception is not going away. Most cybersecurity work for the next decade will still involve stopping attacks in motion, limiting lateral movement, inspecting traffic, and enforcing access. Zscaler remains one of the best-positioned companies for that world.

But the marginal dollar may be moving upstream, to code, identity, endpoint, browser, cloud configuration, and agent governance. That is where anticipation happens. And multiples are set by the marginal dollar.

This is the correction.

I was right that Zscaler’s inline architecture is valuable. I was too quick to assume it would become the highest-value control plane in the AI era. Q3 did not prove that assumption false, but it made the burden of proof much heavier.

Zscaler is not in the penalty box waiting for the market to remember what it is. It is in a structural trial over where security value accrues next.

Being early was indistinguishable from being wrong because the re-rating I was underwriting was built on the wrong altitude. The architecture that looked like infrastructure in the interception era may look more like utility in the anticipation era.

I should have listened sooner. The business is durable enough to keep owning, but the re-rating case now must be earned.

Subscribe now

• Q4 validated Oracle’s AI “bridge,” but not for free: customer-funded / customer-supplied GPU contracts are now real enough to support the $638B RPO story, but Oracle still needs heavy financing, including more debt and equity issuance.

• The core debate is identity: Oracle is part sticky enterprise-data compounder, part capital-intensive AI infrastructure builder. The stock depends on whether infrastructure spending deepens the high-margin data moat or simply turns Oracle into a leveraged GPU landlord.

• Watch cash-flow mechanics, not headline beats: the key tells are BYOH share of new RPO, RPO conversion, net CapEx discipline, ATM dilution, multi-cloud DB disclosure, and whether interest expense compounds faster than operating income.

Oracle beat on revenue, EPS, and operating income, and posted a backlog, $638 billion of remaining performance obligations, that would have been unimaginable for a legacy software company a few years ago. The stock fell anyway.

The easy explanation is “good quarter, scary CapEx.” True, but not enough. The better read is that this is the next answer in a sequence. Over the past year Oracle has retired its risks one by one: first whether the AI demand was real, then whether it could deliver against it, now whether it can finance the delivery without giving away too much of the upside. The answer after Q4 is yes, with asterisks. Financeability is possible, not yet cheap, not yet self-funded, not yet fully equity friendly. Those asterisks are the entire debate.

What We Have Been Updating

We owe a word on our own arc first. In Q2 we argued Oracle had found the future but might strain its balance sheet building it, the Capital Chasm. In Q3 the company gave the first believable answer: large AI contracts could use customer-funded or customer-supplied GPUs, with Oracle providing the operating layer on top. That was the Monetization Bridge, and it changed the question from whether Oracle could sign AI demand to whether it could monetize it without funding every mile of the buildout alone.

Q4 validates that mechanism and raises the bar. We were right about the bridge; we underestimated the weight. A fair reader should ask whether this is refinement or rationalization, and the honest answer is that each quarter has not moved the goalposts so much as revealed the next layer beneath them. Demand was real. Delivery improved. BYOH scaled. But a better funding model is not a self-funding business, and we should also admit we have drifted steadily more bullish across these four quarters. Hold that drift in view; it is why the conclusion must be handed to you rather than asserted.

The Data Business and the Infrastructure Bill

Oracle becomes more valuable as AI grows if proprietary enterprise data becomes the scarce input for useful inference, and much of that data already sits in Oracle databases, applications, and workflows. Every new agent, integration, and multi-cloud deployment raises the cost of leaving. This is not a classic network effect; it is a compounding switching-cost moat built on data gravity and institutional trust. The strongest evidence this quarter was not OCI growth but multi-cloud database growing 404%, Oracle’s database running where the customer already is, monetizing the data layer without owning every layer beneath it. That is the capital-light flywheel.

The discomfort is that Oracle has chosen to protect that flywheel with a capital-heavy buildout. There are two companies inside Oracle, with opposite economics. One deserves the premium, database, multi-cloud database, applications, agents, security, mission-critical workflows: sticky, high-margin, comparatively capital-light. The other consumes the capital, data centers, GPUs, power, cooling, networking: enormous demand, but closer to infrastructure than software, and requiring the spend before the cash arrives. Oracle has fused them by choice. That is the strategic bet: the infrastructure business is the wedge that deepens the data business, win the AI deal, anchor the database, application, and security relationship around it. This is why “GPU landlord” is too simple a slur; it misses the power, networking, identity, operations, and database adjacency wrapped around the chips. But the market is right to press the harder question: if the build is the moat around the data business, how much must Oracle spend before shareholders enjoy the moat? The logic is sound. The financing window is the whole game.

A Better Backlog, not a Free One

The most important confirmation in Q4 is that Q3’s bridge was not a one-quarter artifact. Most of the RPO increase in both quarters came from contracts where customers prepaid for or supplied the GPUs, and the cumulative prepaid-and-customer-supplied pool now stands at $75 billion. We still don’t have the full forward mix, but that is enough to validate the bridge: customer-funded hardware is no longer a theoretical answer to the chasm.

Read what it means, not just what it measures. When Oracle shifts GPU capital to the customer, it concedes, through contract structure, where concessions are most honest, that owning the silicon was never the differentiated thing. Anyone with a balance sheet can buy GPUs. What Oracle sells is the operating layer: networking, security, orchestration, database, the software that turns accelerators into a functioning cloud. Bears focus on lower margin per contract; the better question is return on Oracle’s own capital. If customers fund the most expensive component and Oracle earns attractive economics on the operating layer, BYOH raises the return on Oracle’s capital even as the income statement looks less software-like. That inverts the landlord fear we ourselves carried: the risk is not that BYOH makes Oracle too capital-heavy, but that the market may still be pricing a capital-light operator as if it were buying every GPU itself.

Q4 also drew the limit. BYOH changes the return on Oracle’s capital; it does not make Oracle capital-light. If the model solved the problem outright, Oracle wouldn’t still need another large financing year, $43 billion of debt and $5 billion of equity raised in FY26, roughly $40 billion more planned for FY27 including a $20 billion at-the-market equity program. The bridge is real. It is not free. And this is how to read the new “net cash outlay for CapEx” metric: useful, because it separates what Oracle must fund from what customers prepay, but free cash flow remains the shareholder’s lens. Net CapEx explains funding; free cash flow explains ownership. That Oracle needs the former does not make the latter irrelevant, and the latter was negative $23.7 billion.

So RPO has changed meaning. It is collateral, obligation, and temptation at once, collateral because it supports financing, obligation because Oracle must deliver against it, temptation because the larger it grows, the more seductive it becomes to build ahead of cash conversion. That is the Capital Chasm in a more sophisticated form.

Variant Perception

Consensus sees an AI infrastructure winner with a funding problem. Not wrong; incomplete. The funding problem is real, and the market is right to put Oracle in the penalty box. The variant question is what that box is. Customers own capacity. Creditors own claims. New equity owns dilution. Shareholders own what is left. The whole debate is whether that residual is correctly priced as the leftover of a leveraged infrastructure developer, or wrongly discounted because the market cannot see the data-gravity compounder underneath, the one self-financing its own moat. If Oracle is renting GPUs, the multiple should compress. If it is using infrastructure as the wedge to deepen its grip on enterprise data, the CapEx is not spending, it is the price of the moat. Both readings are plausible on today’s evidence. That is why Q4 mattered: it did not settle the debate. It made it investable.

Three Futures

The honest scenarios aren’t bull, base, and bear, that’s too mechanical. Oracle’s future turns on which identity wins, and each below is gated by a mechanism, not a revenue line.

The point isn’t precision; it’s that each world is gated by a different mechanism, the share of BYOH, the direction of margin, the pace of conversion, the execution of the ATM, the race between interest and operating income. The top tier is mostly a re-rating story, not a revenue story; know that when you weigh it. Watch the mechanism, not the revenue.

What We Think Now

After Q3 we thought Oracle had found a believable way across the chasm. After Q4 we still think so, but the bridge must carry more weight than we appreciated.

The bear has evolved and deserves its strongest form. It is no longer “BYOH may not scale”; Q4 showed it is scaling. The sharper risk: BYOH self-finances new growth, but the roughly $130 billion already on the balance sheet self-finances nothing. Interest expense was $1.44 billion this quarter, up 47%, and on the path toward ~$178 billion of net debt it climbs toward eight-to-ten billion a year, a structural tax that compounds whether the new model works. The real contest is between two identities: a new operator model getting lighter, and an old funding structure getting heavier. Which compounds faster is the whole question.

So, we’ll be opinionated about the parts and silent overall. We believe the bridge is real, the data-gravity flywheel is real, and the landlord fear partly inverts. We will not tell you which identity wins, not from cowardice, but because the one risk that could break this is the one we’re now least equipped to judge, having spent a year talking ourselves toward the other side.

The metrics that adjudicate it aren’t generic beats: BYOH as a share of sequential RPO adds (is the model spreading?); twelve-month RPO recognition above the low-teens (is it converting?); net CapEx near the $70 billion guide, not drifting past $80 billion (is the spend controlled?); equity issued under the ATM (how much do holders surrender?); multi-cloud DB disclosure (is the light flywheel material?); interest expense against operating income (is there time?).

Oracle is no longer simply a software company, nor simply an AI infrastructure company. It is a premium enterprise-data business fused to a capital-intensive buildout designed to defend that data moat. If the fusion works, Q4 will read as the moment the market fixated on the price of admission and missed what the admission bought. If it fails, it will read as the moment Oracle handed investors the future and the bill on the same page. Either way, the answer won’t be in the income statement, where the beats live and everyone is looking. It will surface, quarter by quarter, in the cash statement, in whether the self-financing model spreads faster than the interest compounds.

That is the price of the moat.

Subscribe now

• AI is turning cybersecurity into a runtime-control problem, not just a perimeter problem.

• Palo Alto’s quarter strengthened the control-plane thesis, especially across XSIAM, AIRS, identity, observability and firewalls.

• The stock is strategically more interesting, but not obviously cheap until FY27 proves organic acceleration and cleaner cash-flow delivery.

Palo Alto Networks had a strong quarter. Revenue grew 31% to $3.0 billion. Next-generation security ARR grew 60% to $8.1 billion. Remaining performance obligations grew 36% to $18.4 billion. Adjusted free cash flow was $910 million. Guidance went up.

The stock went down.

That is not the story. It is the symptom.

This is our third article on Palo Alto. The first argued that cybersecurity had moved from gate to graph. The enterprise no longer had one perimeter to defend. It had a distributed graph of users, applications, cloud workloads, identities, data flows, and AI agents. Palo Alto’s answer was to assemble enforcement authority across that graph: firewalls, SASE, browser, XSIAM, cloud, identity, and now observability.

The ambition was right. The concern was execution.

CrowdStrike’s architecture was cleaner: one agent, one data lake, rapid module expansion. Palo Alto’s architecture was broader but messier. It was trying to coordinate many enforcement surfaces into one operating system for security. Our concern was not that Palo Alto had the wrong vision. It was that assembly is harder than architecture.

After Q2, we said Palo Alto had won the operational argument but lost the financial one. The stock fell despite strong results because investors saw the cost of assembly: dilution, margin pressure, integration risk, and organic growth that was good but not decisive.

Q3 changes the discussion.

It does not eliminate the assembly risk. In some ways, it makes the risk harder to measure. But it gives the strategic thesis its strongest validation yet. The reason is simple: the world moved faster.

The Twenty-Five Minute Problem

Our prior argument was that fragmentation becomes dangerous when attacks move faster than humans can investigate. At human speed, the latency between best-of-breed tools is manageable. One tool detects, another verifies, a third revokes access, and a human analyst coordinates the response. It is inefficient, but it can work.

At machine speed, the delay becomes the vulnerability.

That is why the Mythos discussion matters. Palo Alto’s Unit 42 simulated a ransomware campaign from initial entry to data exfiltration in twenty-five minutes. The typical enterprise still needs days to identify a breach. That gap is not a workflow problem. It is an architecture problem.

If attackers move in minutes, dashboards are too slow. Query-based tools are too slow. Manual correlation is too slow. The winning architecture must see what is happening in real time, understand it in context, and act before the attack completes.

That is Palo Alto’s strongest claim: not that it sells more security products, but that it sits close enough to enterprise activity to become the runtime enforcement layer.

The loop is straightforward: Enforce → observe → detect → automate → enforce.

More enforcement surfaces generate more telemetry. More telemetry improves detection. Better detection enables faster automated response. Faster response improves outcomes. Better outcomes drive more platformization. More platformization creates more enforcement surfaces.

That is the control-plane thesis. This quarter moved that thesis from interesting to urgent.

From Gate to Graph to Runtime Control

The graph still matters. AI agents move across identity, cloud, network, browser, apps and data. They do not respect product boundaries. They do not look like traditional malware. Often, they do not “break in” at all. They log in, with credentials, permissions, API keys and approved access.

That changes the security question. The old question was: Is this packet malicious? The new question is: Is this agent, using these credentials, accessing this database, through this path, currently, for this purpose, behaving normally?

That question cannot be answered by a single point product. It requires identity context, network context, endpoint behavior, cloud workload visibility, application telemetry and policy enforcement. More importantly, it must be answered while the action is happening.

That is runtime control.

This is the framework for understanding Palo Alto now. The company is trying to build four layers of runtime control:

This is a more helpful way to read the quarter than the income statement.

The key question is not whether each product grew. The key question is whether the layers are beginning to reinforce each other. If they are, Palo Alto is becoming a control plane. If they are not, it is becoming a very large security suite.

That distinction matters enormously.

Evidence That Runtime Control Is Becoming Real

Q3 gave real evidence that the runtime-control thesis is working.

XSIAM crossed $600 million of ARR, doubled year over year, and now has more than 700 customers. More importantly, management says the majority of deployed customers are responding to threats in under ten minutes. That is the metric that matters. If attacks can complete in twenty-five minutes, response time is not a product statistic. It is the product.

Prisma AIRS reached more than 300 customers. This was not supposed to happen this fast. Two quarters ago, it was a small emerging product. Now management says it has visibility toward $100 million of ARR. A global consulting firm signed a record Prisma AIRS deal to secure AI applications and agents consuming more than two trillion tokens per month. That is not a feature. That is a budget category forming.

CyberArk, the largest integration test, grew around 27% and appears to be ahead of plan on profitability convergence. That matters because identity is the policy layer for the AI enterprise. AI agents do not only attack systems. They log in. Governing agents requires identity to become part of the security control plane.

SASE reached $1.6 billion of ARR, growing 40%. Browser security scaled. Software firewalls kept growing. The distributed enterprise still needs access control, and AI makes the access problem more complicated, not less.

Chronosphere surpassed $300 million of ARR. That matters because AI infrastructure produces massive telemetry. Observability is not traditional cybersecurity, but in an AI-native world it becomes another sensor for understanding system behavior.

These are not isolated product wins. They map onto the same architecture: enforcement, understanding, policy and observation.

But the most interesting evidence came from the oldest part of the company.

The Firewall Comes Back as a Sensor

The old article was about the gate dissolving. Q3 showed something more subtle.

The gate did not return. The gate became a sensor.

For years, Palo Alto’s firewall business was treated as the mature part of the company: profitable, sticky, important, but no longer the strategic growth engine. The multiple came from SASE, cloud, XSIAM and AI security. Hardware was the past.

Q3 complicates that view.

Product revenue grew sharply. Management called it the strongest hardware quarter in a decade. Next-generation firewall bookings grew nearly 40%. The demand was not explained as a normal refresh cycle. Palo Alto pointed to AI data centers, frontier labs, sovereign infrastructure providers and enterprise AI networking.

That is a strategic reclassification.

AI agents create traffic. They call APIs, access tools, authenticate into systems, interact with databases, and communicate with other agents. The more agentic the workflow, the more machine-to-machine traffic it creates.

That traffic has to be inspected somewhere. The closer the inspection point sits to the live flow, the more valuable it becomes. In that world, the firewall is not exciting because it is a box. It is valuable because it is inline.

The firewall comes back as a sensor.

This is the underappreciated positive from the quarter. AI does not only create a new product category like Prisma AIRS. It may also make Palo Alto’s existing enforcement footprint more valuable.

That is the strongest argument for becoming more constructive on the company.

The View Through Fog

Then comes the complication.

Starting FY2027, Palo Alto will no longer break out organic versus acquisition-driven performance. It will move toward total-company guidance and segment reporting across Network Security, Cortex and Identity.

Management has a fair argument. Once CyberArk and Chronosphere are integrated into Palo Alto’s go-to-market, the organic/inorganic distinction gets blurry. If a CyberArk relationship leads to a Palo Alto firewall deal, what is that? If Chronosphere telemetry improves XSIAM detection, is that acquired or organic? Operationally, management wants investors to see one company.

The problem is timing.

This was the first quarter where investors could clearly see the split. Reported revenue grew 31%; organic revenue grew 14%. Reported NGS ARR grew 60%; organic NGS ARR grew 28%. Reported RPO grew 36%; organic RPO grew 22%. The headline beat was strong, but much of the upside came from CyberArk and Chronosphere exceeding expectations.

Those are all good numbers. They are not the same story.

The company is asking investors to believe that assembly is becoming architecture just as it removes one of the cleanest ways to measure whether the core is accelerating. That does not mean management is hiding weakness. But it does mean investors should apply a visibility discount until the new segment disclosures prove they are useful.

Chronosphere has a similar duality. A leading AI lab contributing more than $200 million of ARR is powerful validation. It also means concentration. If observability ARR is a little over $300 million, one customer represents a very large portion of the business. That is AI-native proof, but not yet diversified platform proof.

Cortex Cloud is another open loop. Management acknowledged that the Prisma Cloud to Cortex Cloud migration is still a drag. That matters because cloud security is where the platform data flywheel should be proving itself, and where competition from Wiz and CrowdStrike is most relevant.

This is why the quarter is so interesting.

The thesis is stronger. The evidence is better. The visibility is about to get worse.

The EPS Problem

There is also a simple valuation reason the stock struggled.

CyberArk adds revenue, identity capability and strategic breadth. It also adds dilution. Bloomberg consensus shows FY27 revenue growing around 21%, but adjusted EPS growing only around 9%. That is not the profile investors expect when paying a premium P/E multiple.

The reason is denominator math. The acquisition increases the share count. Revenue gets the benefit immediately; per-share earnings take longer to catch up. Synergies help, but they flow through over time.

This is why free cash flow is the better valuation anchor. Palo Alto’s adjusted free cash flow remains strong, and the path to roughly 40% adjusted FCF margins by FY2028 remains credible if integration synergies materialize.

EPS explains the selloff. Free cash flow explains why the stock still has a floor.

How Our View Changes

Our prior view was that Palo Alto had the right ambition but a difficult architecture problem. It was trying to assemble a control plane across multiple security surfaces while competitors like CrowdStrike had cleaner native architectures.

Q3 makes us more sympathetic to Palo Alto’s approach.

AI changes the security problem. It is not enough to have a beautiful endpoint architecture if the risk moves across identity, network, cloud, browser, AI applications and infrastructure telemetry at machine speed. Breadth may matter more than we previously appreciated because the security question itself is becoming more contextual.

The question is no longer simply: can assembly compete with architecture?

It is now: can assembly become architecture before AI makes real-time control mandatory?

That is a better question for Palo Alto. It does not eliminate the risk. It raises the potential reward.

Our updated view is therefore more constructive strategically, but still disciplined financially. Palo Alto has moved closer to proving that enforcement plus understanding can beat fragmentation. It has not yet proven that assembly can compound like architecture.

Three Loops

The stock is around $260. That already assumes more than “good cybersecurity company.” It assumes Palo Alto becomes something closer to the control plane for AI-era security.

The next three years depend on which loop investors get.

We think the Open Loop is still the most likely outcome. That means the stock is not obviously cheap at $260. It is fairly valued for existing holders who believe the architecture is moving in the right direction, but not yet compelling for new capital without a better entry or clearer FY27 proof.

The Closed Loop requires real platform evidence: AIRS following the XSIAM curve, CyberArk sustaining growth while margins converge, Chronosphere diversifying, cloud migration resolving, organic revenue moving toward the high teens, and free cash flow margins pushing above 40%.

The Broken Loop does not require Palo Alto to become a bad company. It only requires the company to look like a complex security suite growing low-to-mid teens organically while trading at a control-plane multiple.

What Determines the Loop

The next twelve months matter more than the quarter just reported.

The first test is FY27 guidance. A guide that implies revenue above $14 billion with margin expansion would keep the Closed Loop possible. A guide closer to $13.5 billion would suggest that organic acceleration is not yet visible.

The second test is the new segment disclosure. If Network Security, Cortex and Identity are disclosed with enough detail to understand growth and margins, the loss of organic/inorganic disclosure may be manageable. If the segments are too broad to be useful, the market will apply a trust discount.

The third test is Prisma AIRS. Reaching $100 million of ARR is the first checkpoint. Following XSIAM toward several hundred million would be much more important.

The fourth test is platformization renewals. The company cites roughly 120% net retention and single-digit churn among platformized customers. The key is whether that reflects structural switching costs or promotional economics.

The fifth test is cash flow consistency. A real control-plane business should produce increasingly predictable free cash flow, not only strong Q4 collections.

The Bottom Line

In August 2025, we described Palo Alto as a company betting everything on enforcement authority. In February 2026, we said the bet was working operationally but getting more expensive financially. In June 2026, the world moved closer to Palo Alto’s architecture than we expected.

AI compressed time. That makes fragmentation more costly. It also makes Palo Alto’s broad enforcement footprint more valuable.

That is why we are more constructive on the strategic thesis.

But the proof is still ahead. Organic growth is good, not decisive. EPS growth is diluted. Chronosphere is concentrated. Cortex Cloud is still a drag. And the company is about to remove the cleanest window into organic acceleration.

The thesis is stronger. The evidence is better. The visibility is worse.

That combination rewards patience, not blind trust.

Subscribe now

• Snowflake’s Q1 FY2027 shifted the AI debate from “disintermediation risk” to “consumption accelerator,” with product revenue up 34%, NRR at 126%, and guidance raised by more than the quarterly beat.

• CoCo is the key architectural change: not just an AI product, but a new interface layer that compresses migrations, expands who can build on Snowflake, drives platform consumption, and improves Snowflake’s own operating leverage.

• The thesis is stronger but less asymmetric at $240: Snowflake now looks fairly priced as a premium AI-enabled data platform, with further upside depending on whether CoCo/CoWork adoption, NRR, gross margins, and AI monetization prove durable over the next two quarters.

In December 2025, after Snowflake reported its third quarter, I wrote that the most important number in the print wasn’t product revenue growth of 29%, it was the 800-basis point gap between RPO growth and revenue growth. Enterprises were committing faster than they could deploy. “The ‘slowing growth’ narrative,” I argued, “has about two quarters left to live.”

The stock dropped 8%. By April 2026, it had fallen to $118, down 56% from year-end. The software sector was in what trading desks called the “SaaSpocalypse.” Forward P/E multiples fell below the S&P 500 for the first time in history. Capital rotated from cloud into AI infrastructure and industrials. The market’s thesis was structural and specific: agentic AI would disintermediate data platforms entirely. If agents could query and act on enterprise data directly, why would anyone need a centralized warehouse with governance, permissions, and audit trails?

The uncomfortable part was that our architectural thesis made sense before the numbers confirmed it. Data sharing networks were scaling. The Anthropic partnership embedded models inside the governance perimeter, an architectural statement that models should operate inside Snowflake’s security boundary, not beside it, and harder for competitors to replicate than simply offering access to the same model. Nine-figure deals kept arriving, including one worth $400 million. But net revenue retention sat flat at 124–125% for three quarters, hard to reconcile with a platform acceleration narrative. The initial FY27 guide of 27% growth felt like managed deceleration. A thesis built on architecture without consumption proof is just a story the market has no obligation to believe.

What held us anchored were commitment signals the stock price couldn’t see: the RPO gap, the unprecedented deal sizes, and a governance architecture that became more valuable as AI raised the stakes of data access and auditability. But anchored is not the same as certain.

Then Snowflake reported Q1 FY2027, and the consumption evidence arrived.

The Fundamental Question

Product revenue of $1.334 billion, up 34% YoY, accelerating from 30% last quarter and 26% a year ago. Net revenue retention at 126%. Full-year guidance raised by $180 million to $5.84 billion. The stock moved 36% in a session.

The numbers are good. They are not the point. The point is what they reveal about a question that has been building for eighteen months: does AI consumption compound on top of data platform consumption, or substitute for it?

The chatbot era of AI was a retrieval layer that sat above the data platform. A user asked a question, a model answered, done. Minimal compute, minimal platform consumption, minimal strategic significance for Snowflake. Agentic AI is structurally different. When an agent builds a pipeline, deploys a workflow, executes a migration, or launches a monitoring process, those actions run through the data platform, compute-intensive, sustained, and generating new data and governance requirements that feed back into the platform. This distinction explains how the same underlying technology could first appear as a threat and then become a tailwind.

The confirming evidence is strong. Growth re-accelerated by 800 basis points in a single quarter. The $180 million guidance raise exceeded the Q1 beat by $114 million, management embedded a higher forward consumption curve, not just a flow-through. CFO Brian Robins said Cortex Code was “the largest driver to the increase in our forecast.” Use cases deployed on the platform increased 114% YoY. Net new customers of 616 were the most in company history. Snowflake usually waits for multiple quarters of consumption evidence before embedding new products into guidance; CoCo was embedded after one quarter because the signal was strong enough.

The ambiguous evidence requires attention. RPO of $9.21 billion grew 38% but missed consensus and decelerated from 42%. RPO was the leading indicator in my Q3 piece, the clue that commitments were outrunning consumption. Now that revenue has accelerated, the clue is less central but not irrelevant. A softening leading indicator alongside an improving lagging indicator is a mixed signal, not a clean one. Management says bookings are increasingly Q4-weighted, which is plausible but also convenient.

The concerning evidence centres on product gross margin: 75.1%, down 50 basis points. Management acknowledged AI products carry lower margins and said they are offsetting the drag with a $6 billion AWS bandwidth deal. Robins’s phrase, “committed to find efficiency to maintain 75%”, implies active management, not structural protection. The risk is that investors celebrate AI-driven growth before understanding whether it is high-quality growth. If AI crosses 20% of the product mix within the next eighteen months, which the adoption trajectory suggests is plausible, the offset math gets harder.

The honest conclusion: confirming evidence is materially stronger than concerning evidence. But it is one quarter deep.

CoCo Is Not the Product; CoCo Is the Interface

In August 2025, I described a pattern I called “interface collapse”, the elimination of complexity layers requiring specialised knowledge. Oracle won the database wars in the 1980s by making the DBA unnecessary. Snowflake’s first collapse did the same to infrastructure engineers. CoCo is the second collapse: it turns expert data work into natural-language workflows that a much larger population can execute. Migrations that took two years now take one to two quarters. A business analyst who needed an engineering team to build a pipeline can describe what they want and CoCo creates it.

The Investor Day in June revealed an architectural commitment that goes beyond CoCo-as-product. Christian Kleinerman said every new capability must answer: how is the experience simpler with CoCo? Some features will be CoCo-accessible before they get traditional interfaces. Snowflake is building CoCo-first, meaning the proprietary context advantage compounds with every product release, because each release is designed to be CoCo-native. Management calls the accumulated patterns “memory concepts”: cross-customer learning that improves the agent with usage. This is a data loop, not merely a product feature. The more customers use CoCo within Snowflake, the better CoCo gets at Snowflake-specific tasks, and the harder it becomes for generic coding agents to compete on that terrain.

But the most revealing evidence of CoCo’s structural impact was not in the customer metrics. It was in Snowflake’s own headcount. The company added 190 employees in Q1, 173 from the Observe acquisition. Organic hiring was 17. Seventeen people, at a company growing revenue 34% with 9,250 employees.

The internal data explains why: support case resolution 25% faster, engineering time per ticket down 40%, developer productivity doubled, over 100 workflows automated across finance, marketing, sales, and HR. In the Q4 piece, I flagged Robins saying “AI has really changed the framework for investing in growth, it’s no longer tied to headcount.” At the time, that was a forward claim. Q1 made it measurable.

This is where CoCo transcends the “AI product” framing. It operates across four layers of the business simultaneously: direct AI revenue from token consumption; core platform acceleration as CoCo-built workloads consume compute; sales cycle compression as implementations happen faster, with use cases won per account executive up 86% YoY; and internal operating leverage as Snowflake produces more with fewer people. No single product in Snowflake’s history has touched all four layers of the P&L at once. The operating leverage is not a coincidence alongside CoCo, it is a consequence of the same mechanism driving customer revenue.

The counterargument is real. CoCo has been GA for one quarter. Cost governance could throttle adoption, Ramaswamy acknowledged it is “absolutely an issue.” CoWork, the more ambitious business-user agent surface, needs validation through large departmental rollouts before it can be considered proven. The loop is plausible. It is not yet proven durable.

The Variant Perception

The market moved from “AI destroys Snowflake” to “AI helps Snowflake.” Both framings are correct in direction and insufficient in magnitude.

What remains mispriced is the operating leverage, and this is the variant that matters most. Consensus projects roughly 18% non-GAAP operating margin for FY29. The trajectory is already accelerating, 6% in FY25, 10.5% in FY26, 13.5% guided for FY27, and the mechanism is self-reinforcing: better internal AI tools produce more output per person, funding investment in better tools. If the 17-hire dynamic persists across multiple quarters, terminal margins are 28–32%, not 18–22%. Two consecutive quarters of near-zero organic hiring while growth accelerates is a pattern, not a coincidence. No analyst model captures this because the mechanism, AI replacing headcount growth at scale inside a software company, has no precedent in enterprise software. The evidence is early. But it is no longer speculative.

The second mispriced element is underpenetration. Top 25 customers average $34 million in annual spend. G2000 customers average $2.4 million, a 14x gap. CoCo compresses implementation timelines. CoWork, if it achieves departmental scale, expands the user base from data specialists to every knowledge worker with a business question. The 46 customers crossing the $1 million threshold in Q1, versus 26 a year ago, are the leading edge of that expansion. Moving the G2000 average from $2.4 million toward $8–10 million does not require a new product, it requires the existing platform to penetrate deeper inside accounts that have already committed.

What the market prices correctly: the multiple. At 13.6x FY27 EV/Revenue with 89% buy ratings and a $295 consensus target, the repricing from AI victim to AI beneficiary is done.

Three Answers to the Question

The stock is not pricing product revenue growth. It is pricing the answer to one question: does AI make Snowflake more central, merely more useful, or less differentiated? Each answer implies a different business over the next three years.

• The Control Plane Compounds. CoCo and CoWork become the primary interfaces for enterprise data and AI work. The data loop proves durable. Revenue compounds at ~30% to ~$9.8 billion by FY29. NRR re-accelerates above 130% as CoWork expands the user base. Operating margins reach 25% as AI productivity raises the ceiling. At 14x FY29 product revenue: $420–450. Requires CoCo sustaining above 15,000 accounts, CoWork achieving departmental scale, and competitors failing to replicate the governed context advantage. Roughly 25% probability, real but demanding.

• The Better Data Platform. Snowflake is a premium platform with useful AI features but CoCo does not create a second S-curve. Revenue compounds at ~25% to ~$8.8 billion. NRR holds 124–127%. Operating margin reaches 20%. At 11x: $290–320. A high-quality compounder, and at $240 you are paying a fair price for it. Roughly 50% probability.

• The Commodity Layer. Open formats erode governance differentiation. Hyperscalers absorb the context layer. CoCo plateaus as customers throttle AI spend. Revenue compounds at ~19% to ~$7.5 billion. Gross margins compress to 72%. NRR erodes below 120%. At 7x: $140–160. Roughly 25% probability.

In the Q4 piece I estimated $260 from $164, or 58% upside. The thesis has strengthened. The asymmetry has narrowed. That is the tradeoff that comes with being early and staying through the confirmation.

The New Signposts

Six items, each testing a specific leg of the thesis.

Q2 product revenue against the $1.42 billion guide, above $1.46 billion confirms CoCo durability; in-line means expectations have outrun reality. CoCo accounts through FY27, above 12,000 by Q4 sustains the curve; below 9,000 suggests a single-quarter spike. NRR toward 128% by mid-FY28, the strongest possible re-rating signal; below 123% means the expansion engine is weakening. Product gross margin, two quarters below 74% confirms AI is margin-destructive at scale. CoWork departmental deployments of 1,000–10,000 users, validates the second interface collapse for business users. Explicit AI revenue disclosure, “embed first, extract later” is credible but has a shelf life of two quarters before absence becomes a red flag.

Where We Stand

The thesis I held through $118 is closer to consensus at $240. That is a different kind of discomfort than a drawdown.

Conviction is higher than three months ago. Every structural signal, RPO conversion, interface collapse adoption, operating leverage, governed context positioning, moved in the direction I anticipated. The fundamental question has shifted from “is AI good or bad for Snowflake?” to “how far does the compounding loop extend?” That is a better question to sit with.

But at $118 the market was pricing impairment. At $240 it is pricing validation. The thesis is stronger, and the stock now reflects part of that strength. That does not make Snowflake less interesting, it makes the next evidence more important. The next two quarters will show whether the compounding loop is durable enough to justify a further re-rating, or whether Q1 was the high-water mark of a single-quarter adoption spike.

The best investment insights feel obvious in retrospect and unsettling in real time. The governed context layer thesis was unsettling at $118. It is approaching obvious at $240. I have learned to be more suspicious of the latter feeling than the former. The question is no longer whether AI helps Snowflake. It is whether the help compounds.

Subscribe now

• Four quarters ago, I argued that CrowdStrike was not simply a security software company, but a data infrastructure company operating inside cybersecurity. Now management is making a version of that argument on the earnings call. That should feel like vindication. It does not.

• The thesis is stronger than it was at $391, but the opportunity is harder at $659. Q1 made the architecture more credible: AI agents, Mythos-style vulnerability discovery, SIEM, AIDR, Identity, Cloud, and Flex all point in the same direction. But the stock had already rallied on that possibility, and the quarter did not deliver enough ARR proof to justify the move immediately.

• CrowdStrike beat consensus, but consensus was not the bar. The company missed the buyside imagination. The question now is no longer whether Falcon is a platform. The question is whether Falcon becomes the control and remediation layer for AI-era security.

The Inversion

To understand CrowdStrike’s Q1, you first need to understand what happened to software before the company reported.

Earlier this year, software and cybersecurity stocks were in a penalty box. Claude Code and other agentic AI products made the threat feel existential: if AI agents can write code, operate tools, and automate workflows at developer speed, what happens to seat-based SaaS consumption? What happens to developer tools, IT management platforms, or even security software if the human workflow becomes less central?

That was the first narrative: AI compresses software.

Then came the inversion. Snowflake helped show one version of the reversal: AI could increase consumption for the right platforms rather than destroy them. If enterprises need more data, governance, compute, and workloads to build AI, then the right infrastructure companies become more important, not less.

Project Glasswing gave cybersecurity its own reversal, but the point needs to be precise. This was not simply “Anthropic needs CrowdStrike and Palo Alto to secure Mythos.” The more important point was that Mythos demonstrated a capability shock: a frontier model could find and help remediate vulnerabilities in complex, old software that had eluded humans. Anthropic’s response was to work with major technology and security partners so critical systems could be hardened before Mythos-like capabilities became broadly available.

That changed the cybersecurity debate. AI was no longer only a possible threat to security vendors. It was also a threat multiplier for the entire software estate. If frontier models can uncover hidden vulnerabilities at machine speed, then the world does not need less cybersecurity. It needs faster discovery, better prioritization, continuous remediation, and real-time response.

CrowdStrike reported after this regime had already turned. The stock had rallied sharply from the March lows because the market moved from fear to belief. Q1 was supposed to confirm that belief. It partly did. That was the problem.

The Bridge

In March, around $391, I argued that CrowdStrike’s base case offered little return unless the market reclassified the company from security software to security data infrastructure. The structural observations were right: Falcon’s telemetry pipeline, SIEM momentum, Flex consumption behavior, and platform consolidation all mattered more than the quarterly optics.

What I underestimated was the possibility of an external catalyst. Mythos was not in the model. It could not have been. But the right lesson is not “models are useless.” The lesson is that when a company owns a genuine control point, the upside scenarios are often larger and arrive faster than a spreadsheet allows.

At $391, the market was still not paying much for the data-pipeline thesis. At $659, it is paying for a meaningful part of it. That is the uncomfortable update. The thesis is stronger. The margin of safety is not.

The Question Changed

The old question was whether Falcon was becoming security data infrastructure. Q4 pushed that answer toward yes. SIEM made the architecture visible. Flex made the economics more repeatable. Customers were not merely buying endpoint protection; they were embedding Falcon into more of the SOC workflow.

Q1 introduced a harder question: does that security data pipeline become the control and remediation layer for AI-discovered software risk?

That is the only question that really matters. AI agents are not just tools that help humans work faster. They are becoming actors inside the enterprise. They access files, call APIs, inherit permissions, move data, touch cloud workloads, generate logs, and interact with other systems. At the same time, Mythos-like models can make long-buried vulnerabilities visible. The attack surface is not just expanding at the edge; it is being rediscovered inside the software base enterprises already depend on.

This makes CrowdStrike’s architecture more relevant. Falcon sits where a lot of this activity happens: on the endpoint, at runtime, where processes run, files get accessed, APIs get called, and data moves. The sensor creates telemetry. Telemetry creates insight. Insight creates response. Response creates workflow dependency. Workflow dependency creates expansion.

That loop is now being extended from human workflows to machine workflows.

More AI adoption creates more machine activity. More machine activity creates more attack surface. More AI-discovered vulnerabilities create more remediation urgency. More attack surface requires more telemetry, identity governance, detection, response, and data protection. More Falcon modules create more data. More data improves detection and response. Better detection and response make Falcon more central to the SOC. A more central Falcon creates more expansion through Flex.

That is the structural mechanism. AIDR is not the thesis. AIDR is the first obvious product expression of the thesis.

What the Numbers Actually Say

The quarter supported the architecture thesis. It did not fully prove the acceleration thesis.

Revenue grew 26% to $1.39 billion. Ending ARR reached $5.51 billion, up 24%. Net new ARR was $256 million, up 32%. Free cash flow was $468 million, or 34% of revenue. Management raised FY27 net new ARR growth guidance by 520 basis points to 27.7% at the midpoint.

Those are strong numbers. They were not enough.

The reason is simple: the stock was not trading against published consensus. The buyside wanted Q1 net new ARR above $275 million. CrowdStrike delivered $256 million. Worse, $7.8 million of net new ARR came from acquired ARR, which means organic net new ARR was roughly $248 million. On that basis, the headline beat was much less impressive. The company beat the model, but it did not beat the expectation embedded in the multiple.

That is why the stock fell. The market did not reject CrowdStrike’s AI-security thesis. The market had already begun to believe it.

The leading indicators were stronger than the lagging indicators. AIDR ending ARR grew more than 250% sequentially, although the base was not disclosed. Next-Gen SIEM exceeded $600 million of ARR. Combined Cloud, Identity, and Next-Gen SIEM crossed $2 billion. Flex ARR surpassed $1.9 billion, growing 99% year over year. Re-Flex customers reached 480, with a 26% average uplift. More than 130 customers have Re-Flexed multiple times, with a 51% uplift versus the original Flex contract. Subscription gross margin held around 81%, which weakens the argument that Flex is just a discounting mechanism.

Those numbers matter because they describe the flywheel before it fully appears in ARR. Flex removes procurement friction. Customers activate modules faster. Operational workflows adapt to Falcon. Re-Flex becomes a larger renewal because the platform has become more embedded. That is the commercial side of the architecture.

But net new ARR is still the scoreboard. It may no longer be the earliest signal in a Flex world, but it is still the cash-register receipt. If Flex, AIDR, SIEM, and Mythos-driven remediation are truly accelerating demand, they must eventually show up in ARR. Q1 said the flywheel is turning. It did not prove it is accelerating fast enough.

The Snowflake Contrast

Snowflake matters here not because it is the same business, but because it shows the importance of classification.

Snowflake is already accepted as data infrastructure. The market debates how fast it grows, not what it is. When Snowflake beats and raises in an AI consumption cycle, the stock can work because the valuation framework is settled.

CrowdStrike is different. The market still has to decide what CRWD is. Is it a premium cybersecurity platform? Is it security data infrastructure? Or is it the control and remediation layer for AI-era enterprise risk?

That classification question explains the post-earnings reaction. CrowdStrike is not merely being graded on whether it beat consensus. It is being graded on whether the quarter moves it into a higher valuation category. Q1 gave evidence. It did not settle the debate.

The New Misclassification

The old misclassification was product versus platform. The market viewed CrowdStrike as a security product company when Falcon was becoming security data infrastructure.

The new misclassification is platform versus infrastructure. Consensus now understands that CrowdStrike is a high-quality platform. The question is whether Falcon deserves infrastructure status in the AI era.

If AIDR is simply another module, the stock is expensive. If Falcon becomes the runtime control layer for agents, identities, endpoints, cloud workloads, AI-discovered vulnerabilities, and the AI SOC, the addressable role inside the enterprise expands materially.

This is why I am more confident in the architecture and more cautious on the stock. At $391, the market was not paying for the reclassification. At $659, it is paying for enough of it that the next three years require proof, not just architecture.

Three Futures

All prices below are pre-split.

In The Control Plane, CrowdStrike becomes the default runtime security and remediation layer for enterprise AI. AIDR scales, SIEM becomes the AI SOC data layer, identity becomes more important because of non-human agents, and Flex converts new security concerns into platform expansion. Revenue grows roughly 23–25% annually through FY30, reaching $12–13 billion. Free cash flow margins reach 39–42%, producing $4.8–5.5 billion of FCF. If the market values that at 55–65x FCF, the stock could be worth roughly $1,050–1,400. This requires sustained ARR acceleration, visible AIDR scale, SIEM displacements, and Flex continuing to compound without margin erosion.

In The Great Security Platform, CrowdStrike remains one of the best cybersecurity companies in the world, but the infrastructure reclassification remains incomplete. Flex works, SIEM grows, AIDR is real but not transformative, and free cash flow remains strong. Revenue grows around 18–20% annually through FY30, reaching $10–11 billion. FCF margins reach 35–37%, producing $3.5–4.1 billion of FCF. At 35–45x FCF, the stock is worth roughly $500–725. This is the uncomfortable middle case: the company remains excellent, but the stock may not work much because excellence was already priced in.

In The Product, Not the Platform, AI-security urgency creates pipeline but not enough ARR conversion. AIDR remains early, SIEM growth slows, Microsoft and Palo Alto compete aggressively, and CrowdStrike remains best-in-class but loses the infrastructure narrative. Revenue grows 14–16% annually, reaching $8.5–9.5 billion. FCF margins settle around 30–33%, producing $2.6–3.1 billion of FCF. At 25–30x FCF, the stock is worth roughly $275–375. The company survives and generates cash. The stock loses because the multiple changes.

What Resolves It

Q2 matters more than Q1.

Net new ARR above $300 million would validate the AI-security ramp. Below $275 million would pressure the FY27 acceleration story. The second half needs roughly $750 million of net new ARR; below $700 million weakens the guide. AIDR needs dollar-scale disclosure, not just growth rates and pipeline. SIEM needs to sustain growth above 50–60%. Re-Flex cadence and uplift need to remain strong. Subscription gross margin should stay near or above 80% while Flex scales. Endpoint momentum must continue if endpoint is truly the AI runtime control point. RPO and billings need to look cleaner next quarter.

At $550–580, I would be more willing to build. At $659, I would hold but not add aggressively. Above $750 without corresponding ARR proof, I would trim. The thesis is intact. The flywheel is compounding. The margin of safety is not.

The Bottom Line

CrowdStrike’s Q1 did not change the thesis. It changed the burden of proof.

The company is no longer trying to prove that it survived the outage. That question has been answered. It is no longer merely trying to prove that Falcon is a platform. That case has strengthened over multiple quarters.

The new question is harder and more valuable: can Falcon become the control and remediation layer for AI-era security?

Q1 made that question worth asking. It did not fully answer it.

This is where I land: the thesis is stronger than it was at $391, but the opportunity is worse. CrowdStrike’s business is proving durability. Its strategy is reaching for infrastructure status. The stock is priced for proof.

Q2 and the second half will decide whether those three statements can coexist.

Subscribe now

• Broadcom’s quarter was not weak. It was strong versus reported numbers, but not strong enough versus the buy-side AI bar that had moved ahead of consensus.

• The prior thesis still holds: Broadcom is not a normal semiconductor cyclical. AI scale makes its custom silicon, Ethernet networking, SerDes, optics, packaging, supply-chain execution, and VMware cash engine more valuable.

• What changes after 2QFY26 is the framing. Broadcom is not the prime contractor for the rack. It is the industrial silicon and networking layer that makes other people’s AI infrastructure deployable, and increasingly financeable.

The Hardest Updates

The hardest earnings updates are not the ones that prove you wrong. Those are easy. The facts change, the thesis breaks, and the only honest response is to move on.

The harder updates are the ones where the thesis was mostly right, but the market caught up, the language became too easy, and the metaphor started doing more work than the evidence could support.

That is where Broadcom is after 2QFY26.

Across our previous Broadcom work, the central idea was that Broadcom lived in the physics economy while much of AI investing lived in the narrative economy. The narrative economy runs on frameworks, partnerships, press releases, optionality, and the emotional value of being adjacent to the next AI model. The physics economy runs on supply-chain allocation, packaging capacity, purchase orders, delivery schedules, power availability, yield, and revenue recognition.

Hock Tan’s great advantage was that he refused to confuse the two. When the market wanted OpenAI’s 10GW framework to become near-term revenue, he clarified that it was not. When investors wanted every AI announcement to sound like contracted backlog, he separated aspiration from delivery. That honesty was punished at first. Then it became credibility.

In December, Broadcom was punished because the market did not appreciate the discipline. In March, Broadcom was rewarded because the market finally did. In June, the same discipline created a different problem: investors had moved from disbelief to expectation.

Broadcom did not stop being the physics economy. The issue is that, for the first time, the expectations economy caught up.

The Twelve-Quarter Contract

This was a strong quarter by any normal standard. Revenue reached a record $22.2 billion. Adjusted EBITDA was $15.2 billion, or 69% of revenue. Free cash flow was $10.3 billion. AI semiconductor revenue reached $10.8 billion, up 143% year-over-year. Q3 AI semis are guided to $16 billion. FY26 AI semis are expected around $56 billion. FY27 remains above $100 billion.

Those numbers would have seemed absurd two years ago. Yet the stock sold off sharply.

The reason is simple: Broadcom was no longer trading against normal standards. It was trading against the AI whisper. The market wanted Q2 AI semis above the buy-side bogey, Q3 AI semis closer to the high-teens, FY26 AI revenue closer to $60 billion, and a cleaner reason to believe FY27 was not merely above $100 billion but materially above it.

Instead, investors got a good quarter, a good guide, a reiteration of FY27, a messy call, a comment about Google source diversity, and a clarification that Broadcom is not selling racks.

That explains the stock. It does not fully explain the company.

The expectations miss matters because stocks trade on expectations. But the deeper question is whether the structure of AI compute demand is changing in a way that makes Broadcom more valuable over time. This is where the quarter may have revealed more than the market reaction suggested.

From Capex Cycle to Financeable Infrastructure

For most of the last three years, AI compute demand has been understood as a hyperscaler capex cycle. Google, Microsoft, Amazon, Meta, and a few others decided how much to spend. Everyone else in the AI supply chain lived beneath that decision. If the hyperscaler CFO got cautious, the order book slowed. If data-center budgets paused, semiconductor revenue paused.

That is what it means to be in a capex cycle: someone else’s budget is your demand function.

The interesting development in Broadcom’s Q2 call was that this model may be changing at the margin. Anthropic, OpenAI, and other frontier labs are not simply internal hyperscaler workloads. They are AI-native demand sources with their own usage curves, token economics, and capacity requirements. They need compute not because a hyperscaler wants to spend, but because users are consuming intelligence and that usage requires infrastructure.

The Apollo and Blackstone AI XPU platform matters because it points to a new capital stack. When financial investors help fund AI compute capacity, they are not merely buying a semiconductor cycle. They are trying to underwrite infrastructure: power, capacity, utilization, customer demand, and yield over time.

That does not make AI compute risk-free. It is still early. Token economics still need to prove themselves. Utilization matters. Power matters. Deployment timing matters. But the conceptual shift is important. AI compute is beginning to look less like discretionary tech capex and more like a financeable infrastructure asset.

Broadcom sits at the conversion point. It is not the owner of the application. It is not the consumer interface. It is not the balance sheet funding every data center. Its role is more specific and, in some ways, more powerful: it helps turn financial capital into deployed AI capacity.

That is a different kind of value than “we design good chips.” It is the value of being the industrial layer that makes the delivery schedule bankable.

What We Got Right

The structural direction was right. Broadcom is not a normal semiconductor cyclical. It is increasingly an AI infrastructure company with three connected engines: AI semiconductors, AI networking, and infrastructure software.

The quarter confirmed that AI semis are not a side business anymore. They are becoming the center of the company. It also confirmed that networking remains essential. Broadcom’s AI stack is not just custom XPUs. It is Ethernet, SerDes, DSPs, optics, PCIe, Tomahawk, Jericho, and the connective tissue that lets distributed compute function at scale.

This is the part of the thesis that remains strongest. AI does not scale like normal compute. Every new gigawatt adds coordination problems: latency, power, packaging yield, optical reach, memory bandwidth, cluster topology, supply timing, and multi-data-center synchronization. At small scale, the chip matters. At massive scale, the system around the chip matters just as much.

That is Broadcom’s mechanism. Every new customer program deepens three advantages. First, manufacturing locks: supply-chain capacity committed years ahead. Second, IP accumulation: SerDes, packaging, networking, and co-design knowledge that improves with each generation. Third, cross-sell density: networking, optics, and interconnects that travel with the compute.

These advantages reinforce each other. Capacity commitments enable investment. Investment improves execution. Execution wins the next program. The next program increases attach opportunities. That is the compounding loop.

The question is not whether Broadcom can sell more AI chips. The question is whether each new gigawatt makes Broadcom harder to replace.

What We Need to Correct

Honesty also requires updating the parts of the thesis that were too generous.

The biggest correction is the prime-contractor metaphor. In our prior work, we leaned into the idea that Broadcom was becoming the AI infrastructure prime contractor, not just selling chips, but increasingly responsible for the system. That captured something real: Broadcom does coordinate critical layers of the AI stack that customers cannot easily assemble alone.

But the metaphor went too far.

This quarter, management clarified that Broadcom is not selling full racks. It is not capturing full rack-level economics. It is in the chip business: chips, connectivity, switching, optics, and networking infrastructure.

That distinction matters. If one assumed Broadcom was capturing full rack economics, the dollars-per-gigawatt model was too generous. If one assumes Broadcom captures the hardest silicon and networking layers inside and across the rack, the thesis remains large, but cleaner.

The right framing is not that Broadcom owns the rack. The right framing is that Broadcom makes other people’s racks deployable.

That may actually be more durable. It avoids rack-level balance sheet risk and operational complexity, while preserving the highest-value silicon and networking layers. Broadcom does not need to own the full system to be essential to the system.

We pushed the metaphor too far. The insight survives, but the model has to be cleaned up.

Google Diversification

The Google comment matters. It is the cleanest new bear hook.

Broadcom’s longest and most important custom silicon relationship is not immune to second sourcing. That should not surprise anyone. Hyperscalers hate dependency. The more important AI infrastructure becomes, the more rational it is for Google to diversify across vendors, internal programs, and alternative SKUs.

This weakens the easy version of the bull case. The easy version said Broadcom was irreplaceable. The more accurate version is that Broadcom is very difficult to replace at scale, across generations, in the most complex parts of the roadmap.

That is still a powerful position. But it is not the same as monopoly.

The nuance is important. Google diversification is not evidence that Broadcom is irrelevant. It is evidence that the problem Broadcom solves has become too important to single-source. The bear case is that Broadcom’s share is capped. The bull case is that the pie is expanding fast enough for Broadcom’s dollars to grow even if share is not absolute.

Diversification caps the monopoly narrative. It does not automatically cap the revenue opportunity.

Networking: Inseparable, Not Bigger

We also need to refine the networking thesis.

Our prior work was directionally right that AI networking is the hidden asset. The market tends to focus on XPUs because custom accelerators are easier to understand. But as clusters scale, the network becomes the bottleneck. The more distributed the compute, the more valuable the fabric.

The correction is that networking does not need to become larger than compute to be the moat. It only needs to become inseparable from compute.

Every gigawatt of XPU capacity requires switching, optics, SerDes, DSPs, PCIe, and interconnect. Those products are broader, more reusable, and often richer margin than the custom XPU itself. They also make Broadcom less dependent on any single customer’s accelerator architecture.

The custom chip wins the headline. The network may win the economics.

VMware’s Second Act

The software business was supposed to be the ballast. It was the high-margin cash engine that let Broadcom fund AI R&D without relying on capital markets or narrative finance.

This quarter suggests something more interesting: VMware may also become an AI beneficiary.

That sounds counterintuitive because VMware is not what investors think of when they think of AI. But enterprise AI is not only frontier model training. It is private cloud, inferencing, data governance, workload orchestration, security, and infrastructure control. AI agents also increase compute complexity inside enterprises. More CPU cores, more heterogeneous workloads, more need to manage private infrastructure, that is not bad for VMware.

This does not make VMware the main growth story. AI semis are still the growth story. But it gives Broadcom something rare: a software annuity that both funds the AI cycle and may participate in it.

VMware was supposed to finance the AI bet. It may now be one of the ways Broadcom monetizes AI.

The Stock: Reclassification, Not Just Revision

This is where the investment debate has changed.

Earlier, the opportunity was misclassification. The market valued Broadcom like a cyclical semiconductor company while it was becoming an AI infrastructure platform. That gap has narrowed. The market now understands the story.

The new opportunity, if it exists, is reclassification. Is Broadcom still an AI semiconductor name that missed the buyside bogey? Or is it becoming an AI infrastructure platform whose demand is increasingly tied to financeable compute capacity, not just hyperscaler capex budgets?

That distinction matters for valuation. At roughly $386, Broadcom is not obviously expensive if FY28 and FY29 earnings are real. Bloomberg consensus already implies a dramatic ramp, with revenue moving from roughly $105 billion in FY26 to more than $220 billion by FY28 and adjusted EPS moving from roughly $11.5 to more than $25.

The bear case is not simply that AVGO is expensive. The bear case is that the out-year numbers are too high. Google share may leak. Content per gigawatt may disappoint. Customer deployments may slip. Networking mix may not offset lower-margin XPUs. Operating margin may compress.

The bull case is that FY27 is not the peak; it is the bridge. Meta, OpenAI, Anthropic, Google, and the additional customers scale into FY28 and FY29. Networking remains strategic. VMware contributes more than expected. The Apollo platform repeats. And the market reclassifies Broadcom from semiconductor cyclical to financeable AI infrastructure layer.

Broadcom is cheap only if the future arrives on time.

Three Futures

In The Reclassification, Broadcom is no longer treated as an AI semiconductor supplier that happened to catch a powerful cycle. It is re-rated as the industrial layer of financeable AI infrastructure. FY27 AI semis exceed $100 billion comfortably, FY28 and FY29 accelerate as Meta, OpenAI, Anthropic, Google, and the other customers ramp together, networking stays around 40–50% of AI revenue, software grows double digits, and operating margin remains above 60%. That could support FY29 EPS power of $38–42. At 20–22x, the stock could be worth $760–900.

In The Long Bridge, the thesis works, but it takes longer than the market wants. FY27 lands around the $100 billion-plus target, FY28 grows but with lumpiness, Google diversifies but Broadcom dollars still grow, networking remains meaningful, and operating margin settles around 57–60%. Broadcom remains a great company, but the market stays in “show me” mode rather than fully reclassifying the business. That could support FY29 EPS power of $30–34. At 16–18x, the stock could be worth $480–610.

In The Capex Reversion, the market decides that AI compute has not become financeable infrastructure after all. It is still a hyperscaler capex cycle, vulnerable to timing slips, customer concentration, and margin pressure. Customer deployments slip, Google share loss becomes meaningful, content per gigawatt disappoints, networking mix declines, and operating margin compresses toward the low-to-mid 50s. FY29 EPS power may be closer to $21–25. At 12–14x, the stock could be worth $250–350.

The range is wide because the business has become more powerful and more concentrated at the same time. That is usually what happens when a company moves from misunderstood to important.

What We Track

The first signpost is Q3 AI semis. Around $16 billion keeps the thesis intact; a miss would raise timing concerns. The second is FY26 AI semis around $56 billion and continued confidence in FY27 above $100 billion. Any softening there would be thesis-level.

The third is FY28 language. The market needs to believe FY27 is a bridge, not a peak. The fourth is Google: normal second sourcing is manageable; meaningful Broadcom content loss is not. The fifth is networking mix. Around 40% of AI revenue is healthy; moving toward 50% would strengthen the margin and moat argument.

The sixth is operating margin. Gross margin can move lower; operating margin needs to remain in the mid-to-high 60s. The seventh is VMware. The Q3 software guide needs to convert because software is now part of the AI resilience thesis. The eighth is infrastructure financing. A second Apollo-style tranche would suggest the phase transition is repeatable, not one-off.

The final signpost is Hock Tan himself. The stock had been awarded a credibility premium because Broadcom was precise for twelve quarters. This quarter dented that premium. Restoring it requires a clean guide, cleaner language, and evidence that the $100 billion FY27 path is not the ceiling but the bridge.

The Phase Transition

Broadcom did not stop being the physics economy. It still operates in delivery schedules, supply chains, engineering execution, and capacity commitments. But the market is no longer treating that as a secret.

The old opportunity was that investors misunderstood what Broadcom was becoming. The new opportunity, if it exists, is that investors may underestimate what happens when AI compute becomes financeable infrastructure.

That is a harder phase. It requires less storytelling and more proof. It also creates a larger prize. Capex cycles are bounded by budgets. Infrastructure markets are bounded by demand, utilization, financing capacity, and return on capital.

Broadcom’s thesis survived the quarter. The expectations premium did not. The next question is whether AI compute has truly crossed from discretionary capex into infrastructure, and whether Broadcom is the industrial layer that makes that transition real.

The floor became the ceiling for one quarter. The structural question is whether it becomes the floor again.

Subscribe now

• SpaceX looks wildly expensive on conventional numbers, but the real IPO question is whether Musk can once again turn belief into capital, capital into capacity, and capacity into proof.

• Colossus may matter more than Grok: the AI segment’s most valuable asset could be SpaceX’s ability to build powered, cooled, usable compute capacity faster than hyperscalers can.

• The stock’s fate depends on whether Anthropic and Google-style demand proves temporary “bridge capacity” or the beginning of a durable scarcity business; if proof lags belief, financial gravity returns.

There was a long period when Tesla was almost impossible to value correctly.

That does not mean the sceptics were wrong. They were often right. The company was too expensive on near-term numbers. Production was messy. The promises were too large. The balance sheet mattered. Dilution mattered. Execution risk mattered. At various points, the distance between what Tesla was and what Tesla claimed it would become seemed almost indefensible.

And yet, that was never the whole story.

Tesla’s stock was not merely a claim on current car deliveries, gross margins, and free cash flow. It was also a mechanism. Belief in Elon Musk’s future gave the company access to capital. That capital funded factories and batteries and software and manufacturing scale. Those physical assets made parts of the original belief look less absurd after the fact. Share issuances that should have diluted the stock instead increased it, because investors were not pricing a financial instrument, they were propagating a conviction. The stock was not merely reflecting reality. At times, it was helping create the reality it was supposedly discounting.

This is not a comfortable way to think about investing. It sounds circular, and sometimes it is. Narrative can fund genuine progress, but it can also postpone accountability. Belief can become capacity, or it can become overpayment. The difficult part is that both things can be true at different times in the same company’s life.

SpaceX now presents the same problem, on a much larger canvas.

The Obvious Observation and the Interesting One

The ordinary observation about SpaceX is that it is expensive. That is true. At the expected IPO price of $135 per share, the company is being valued at approximately $1.75 trillion on $18.7 billion of 2025 revenue, roughly 94 times trailing sales, on a consolidated income statement that shows a net loss of $4.9 billion. On traditional measures, that is not a valuation. It is an act of faith.

But that is the least interesting observation about the IPO.

The more interesting question is: what kind of faith?

Investors are not merely being asked to believe in rockets. They are being asked to believe in a sequence: launch capability funded Starlink, Starlink funds Starship, Starship enables next-generation satellites and satellite-to-mobile and eventually compute in orbit. They are being asked to believe that the same company which made access to orbit cheaper can use that advantage to build the next set of scarce physical assets the world needs. They are being asked to believe that, once again, Musk can turn belief into capital, capital into capacity, and capacity into proof.

That is the real question at the heart of the SpaceX IPO:

Can SpaceX turn its narrative advantage into scarce physical capacity faster than global demand for AI compute, connectivity, and launch can be satisfied by everyone else?

Everything else is a sub-question.

SpaceX’s First Loop

SpaceX’s original job was not “space exploration” in the abstract. It was more specific: move mass to orbit more cheaply, more reliably, and more frequently than the legacy aerospace system could manage.

That distinction matters. The customer hiring SpaceX was not buying poetry. A government customer needed mission assurance. A satellite operator needed delivery to orbit. NASA needed crew and cargo transport. The job was to remove the launch bottleneck. And SpaceX removed it: the S-1 reports approximately 620 Falcon 9 orbital launches with an over-99 percent mission success rate, a first-stage booster reflown as many as 34 times, and more than 80 percent of global mass to orbit each year since 2023.

Once that bottleneck was solved, SpaceX did something more important than sell cheaper launches to third parties. It became its own best customer.

Starlink was the proof.

The conventional way to understand Starlink is as satellite broadband. That is accurate but incomplete. Starlink is what happened when cheaper launch created a business that would not have made sense under the old cost curve. Thousands of satellites in low-Earth orbit are not an incremental improvement over geostationary broadband; they are a different architecture for making geography less relevant. The rural household does not hire Starlink because it likes satellites. It hires Starlink because fibre will not come soon. Airlines do not install Starlink because they want to make Musk happy. They install it because passengers expect the internet in the air to feel more like the internet on the ground.

This created a compounding loop: cheaper launch enabled more satellites, more satellites improved coverage, better service attracted more subscribers, more subscribers funded more launches, and more launches made the system progressively harder to replicate. Starlink’s 2025 numbers show the loop working: $11.4 billion of revenue, up roughly 50 percent year on year, at $7.2 billion of segment Adjusted EBITDA, with 10.3 million subscribers across 164 countries.

The pattern to hold in mind is not the specific numbers. It is the sequence. SpaceX’s most important strategic move is not winning a market. It is solving one physical constraint and then building the next business on top of the solved constraint.

The S-1 suggests SpaceX is attempting to do this again. But the next constraint is not launch. It is physical AI compute capacity.

The Strange S-1

At first glance, the SpaceX S-1 looks like several companies stitched together.

There is the launch business: Falcon, Dragon, Starship, government contracts, mission success rates. There is the connectivity business: Starlink broadband, enterprise, aviation, maritime, satellite-to-mobile. Then there is the AI business: xAI, Grok, X, Colossus, Colossus II, and eventually compute in orbit.

A note on reading these numbers: the S-1’s consolidated financials were retrospectively recast to include xAI and X, because the transactions were between entities under common control. That means the historical comparisons are not as clean as a typical company going public with an organic segment history. When you see “2023 revenue: $10.4 billion,” that includes xAI and X revenue retroactively attributed to that period. The trends reflect a reconstructed combined entity, not SpaceX’s standalone trajectory.

The easy critique is that xAI is being wrapped inside SpaceX because xAI alone would be much harder to justify. That critique should not be dismissed. The AI segment brought a $6.4 billion operating loss in 2025, consumed $12.7 billion of capital expenditure, and houses a frontier model that has not established itself at the top tier. The founding research team departed. X, the social network folded into the segment, is a complicated asset. If one were simply underwriting Grok against Claude, Gemini, and GPT-4, SpaceX would look less like a clean space-and-connectivity story and more like an extraordinary company carrying a speculative AI burden.

But that framing may miss the point.

The better question is not whether Grok is the best model today. It is whether Grok is the business, or whether Grok caused SpaceX to build something more valuable.

That something is physical compute capacity. Colossus, SpaceX’s flagship AI compute site in Memphis, came online in 122 days from an empty factory shell. Colossus II followed in 91 days. For context, the industry benchmark for bringing a comparable greenfield data centre online is approximately two years. The S-1 notes that construction costs for Colossus II were “considerably lower than industry benchmarks on a per megawatt basis.” Together, the two sites provide approximately 1.0 gigawatt of compute power.

The AI story in the S-1 is not really about a chatbot. It is about chips, power, cooling, land, networking, and construction speed, physical bottlenecks that happen to sit squarely in Musk’s demonstrated area of strength. There is a useful historical echo here: Samuel Insull dominated American electricity in the early 1900s not because his power was better, a kilowatt-hour is a kilowatt-hour, but because he built generation capacity faster than anyone else. In a market where demand outpaces physical supply, the fastest builder earns the scarcity premium. That principle may now apply to AI compute.

Musk is inconsistent in unconstrained software markets. X has made that obvious. But physical constraints have historically been his best arena. Rockets are physical. Cars are physical. Factories are physical. Batteries are physical. Data centres are physical. Power is physical. The world of AI may be discovering that the most important software market in history is increasingly limited by the least software-like problems.

The Market Tells Musk What Business He Is In

The Anthropic and Google deals are the hinge of the SpaceX story.

xAI was always two businesses sharing a balance sheet. There was the model business, Grok, competing against Claude, Gemini, and GPT on talent, data, and algorithmic insight. And there was the physical-capacity business, Colossus, competing on construction speed, power availability, and deployment timeline. These businesses have opposing incentives. The model business wants exclusive access to every GPU to improve Grok. The physical-capacity business wants maximum utilisation at the highest price from whoever will pay. Every GPU rented to Anthropic is a GPU not training Grok. Both cannot get what they want.

The Anthropic deal resolved the tension. In May 2026, Anthropic, a direct competitor in the frontier model race, signed a contract for access to approximately 325,000 NVIDIA GPUs across Colossus and Colossus II, at $1.25 billion per month through May 2029. SpaceX does not need Anthropic to prefer Grok. It only needs Anthropic to need compute. That is a profound difference.

The Google deal, announced in early June via a regulatory filing, made it structural. Google is, by Epoch AI’s April 2026 analysis, the single largest owner of AI compute on the planet, holding roughly 25 percent of global capacity, equivalent to approximately 5 million H100-class processors. The five largest U.S. hyperscalers collectively hold an estimated 71 percent of global AI compute, up from 63 percent in early 2024. Their combined annual capital expenditure exceeds $380 billion. And Google still needed to rent 110,000 GPUs from a rocket company at $920 million per month, describing the arrangement as “bridge capacity” for demand that was “even higher than we expected.”

When the largest owner of AI compute on Earth cannot build fast enough to meet its own demand, the shortage is not a temporary blip. It is a physical constraint. And the nature of that constraint matters: power permitting takes three to five years in many jurisdictions. Grid interconnection queues are multi-year. Water rights for cooling face growing political opposition. Environmental review adds 12 to 18 months. Community resistance to large-scale data centre projects is intensifying. These bottlenecks do not respond to larger cheques. They respond to time. And AI demand, by Epoch AI’s estimate, doubles roughly every seven months.

The combined contracts represent approximately $26 billion in annualised revenue for a segment that generated $3.2 billion in all of 2025.

There is an uncomfortable irony in Grok’s competitive position that may actually strengthen the investment thesis. If Grok were dominant, investors would have reason to worry that the leasing business might be deprioritised in favour of internal model training. Grok not being dominant makes the pivot to selling scarce physical capacity credible and likely durable. The most important customer for SpaceX’s AI buildout may not be Grok. It may be Grok’s competitors.

The Variant Perception

The consensus view is straightforward: SpaceX is an extraordinary launch and connectivity company, with a money-losing AI business attached.

The variant view is different. The AI segment may be misread. The near-term value may not rest on Grok winning the model race, but on SpaceX becoming the fastest route from demand to usable compute.

This matters because AI is moving from a purely software race into a race for physical capacity. The model companies can raise capital, hire researchers, and find users. They cannot instantly create powered, cooled, networked compute capacity at scale. The bottleneck is not just chips. It is chips plus power plus buildings plus interconnect plus time.

If that bottleneck persists, something interesting follows: competition among AI model companies may actually benefit SpaceX. If one model provider wins everything, SpaceX has a single dominant customer. If many model providers compete aggressively, all of them need compute. Competition in the model layer becomes demand for the physical layer. SpaceX does not need to pick the winning model. It can sell to all of them.

The 90-day termination clauses in both contracts deserve honest treatment. They reduce revenue durability, these are not long-term locked commitments. But they also enable repricing in a persistent shortage. Which reading is correct depends on whether scarcity is temporary or structural. That is the core bet.

The orbital compute option, deploying AI workloads on satellites powered by solar energy, should be treated as a call option, not a base case. The S-1 targets initial deployment as early as 2028, and the engineering challenges are substantial: heat rejection in the vacuum of space, radiation effects on calculations, power density per kilogram, and the sheer scale of deployment required. But it is an option only SpaceX can plausibly exercise, because it requires launch economics, satellite manufacturing capability, and operational experience in orbit that no other entity currently possesses. If terrestrial power and permitting constraints prove genuinely binding at the scale AI demands, this option moves from speculative to necessary.

What Bulls May Be Missing

The danger is that the thesis sounds too clean.

Contract durability is the most important near-term risk. The Anthropic and Google deals are not the same as a permanent annuity. They may be bridge capacity. They may be renegotiated. They may be cancelled. Both contracts are terminable on 90 days’ notice after initial periods. Customers may build their own capacity within two or three years. GPU generations depreciate quickly, today’s H100s will be two generations old by 2029. A data centre that looks scarce today can become less scarce if supply catches up or if model efficiency improves dramatically. The most important mistake a bull could make is treating every dollar of compute revenue as if it deserves the same multiple as durable recurring cloud revenue. It probably does not.

The cash burn is not just high, it is accelerating. Q1 2026 capital expenditure was $10.1 billion, of which $7.7 billion was AI, against only $1.0 billion of operating cash flow. Cash fell from $24.7 billion to $15.9 billion in three months. Annualise the Q1 2026 net loss of $4.3 billion and you get roughly $17 billion, more than triple the full-year 2025 loss of $4.9 billion. The company is simultaneously funding Starship development, AI compute buildout, a $19.6 billion spectrum acquisition (EchoStar, approved but not yet closed, expected November 2027), and inherited debt from the xAI and X absorption. Total principal debt stood at $29.1 billion at March 31, 2026.

The IPO is partly a debt cleanup, not purely a growth capital raise. The S-1 discloses a $20 billion bridge loan in March 2026, with proceeds used to extinguish X and xAI debt, including X B-1, X B-3, xAI fixed and floating loans, and xAI 12.5% senior secured notes. The repayment amount was approximately $18.9 billion, including a $1.16 billion prepayment penalty. Part of what investors are funding when they buy the IPO is the financial consequences of absorbing xAI and X, not only the next generation of Starship and AI compute.

Starship is a single point of failure for too many futures. Nearly everything in the long-term thesis, next-generation V3 Starlink satellites, satellite-to-mobile at scale, cost-per-kilogram reductions, orbital compute, depends on one vehicle that has completed 11 test flights and has not yet delivered a commercial payload to orbit. The S-1 targets the second half of 2026 for payload service. If Starship is delayed meaningfully, multiple revenue lines delay with it.

Starlink’s cash engine is strong but the price trajectory is worth watching. Blended ARPU has declined every year, from $99 per month in 2023 to $66 in Q1 2026. The decline is mix-driven, reflecting a deliberate land-grab in newer, lower-priced geographies, and the expectation is that V3 satellites and premium aviation, maritime, and enterprise plans will stabilise or improve the trajectory. But V3 depends on Starship. If ARPU continues declining and subscriber growth alone has to carry the revenue line, the funding engine for everything else weakens.

Revenue concentration after the deals is extreme. Two customers will account for the vast majority of AI segment revenue, both with short-term termination rights. If both walked simultaneously, unlikely but not impossible, SpaceX would go from $26 billion in annual AI revenue to near-zero within a quarter, while still carrying the capex burden of facilities built to serve that demand.

Signal quality varies across the business. SpaceX is now part of a broader Musk ecosystem in which Tesla, xAI, X, and SpaceX transact with one another. Tesla’s 2025 10-K/A discloses $573 million in revenue received from Musk-controlled entities, including $430 million in Megapack sales to xAI and $143 million from SpaceX vehicle purchases. These transactions may be commercially rational and were reviewed by Tesla’s Audit Committee at arm’s-length terms. But internal ecosystem demand is less informative than independent third-party demand. That is precisely why Anthropic and Google matter so much: they are cleaner external validation. The more SpaceX’s growth comes from independent customers rather than intra-ecosystem activity, the stronger the thesis becomes.

Musk himself is the narrative engine and the governance risk. SpaceX benefits from his credibility, ambition, and talent magnetism. But the dual-class structure concentrating approximately 82 percent of voting power in his hands limits minority shareholder influence. His operational attention is split across multiple companies. His public role can attract capital and talent, or it can repel customers and partners, depending on the moment. The IPO structure reinforces this: investors are not buying ordinary governance rights. They are buying exposure to a company where one person’s judgment is inseparable from the thesis.

The danger is not that SpaceX is fake. It is obviously not fake. The danger is that investors treat several futures as though they have already arrived.

How to Follow a Narrative-Driven Stock

For most companies, belief follows proof. For Musk companies, belief often arrives early and pays for the proof. That is the uncomfortable mechanism. Narrative lowers the cost of capital, attracts talent, pressures suppliers, and gives management permission to make long-term bets that would look irrational inside a more conventional company.

This can be an advantage. It can also be a trap. If belief funds physical capacity, and that capacity wins customers, the narrative becomes self-confirming. If belief funds capacity that customers do not want, or that arrives too late, financial gravity returns.

The right way to follow SpaceX is not simply to ask whether revenue grew this quarter. It is to ask whether proof is accumulating in the places that matter.

Third-party compute demand. Do Anthropic and Google expand, renew, reduce, or terminate? Do new customers appear? Does “bridge capacity” become recurring dependency? Is pricing holding or compressing? Does SpaceX disclose utilisation rates? Is new demand coming from independent third parties or from within the Musk ecosystem?

Starship. Does it deliver payload to orbit on a repeatable basis? Does reuse become real? Does cadence increase meaningfully? Do V3 satellites begin deployment? Are there external commercial payloads, or is the manifest mostly internal Starlink missions?

Starlink. Are subscribers still growing? Is average revenue per user stabilising after the mix-driven decline from $99 to $66 per month? Are aviation and maritime customers increasing? Does direct-to-cell move from promise to meaningful use? Does the EchoStar spectrum acquisition close on schedule?

Orbital compute. This should be tracked as an option. Can SpaceX demonstrate useful compute from orbit? What is the power per kilogram? How effectively is heat removed? What are the error rates? Does the concept produce useful work at a cost that can eventually compete with terrestrial alternatives?

The narrative itself. This may sound soft, but it is not. Does the market increasingly describe SpaceX as a supplier of scarce AI compute capacity, or does it continue to see AI as a Grok story? Which analysts cover the stock, and what comparison set do they use, aerospace names at 15 to 25 times earnings, or AI and compute names at 50 to 100 times? Does the stock correlate with AI sentiment or aerospace sentiment on volatile trading days? What does Musk emphasise on the first earnings call?

Only after these questions should one return to the traditional metrics: revenue by segment, capital expenditure per megawatt, revenue per megawatt, EBITDA, free cash flow after growth spending, debt trajectory, dilution, and stock-based compensation. The numbers will eventually decide the stock. But the narrative will decide where the numbers are allowed to go first.

Three States of Belief

The usual labels, bull, base, and bear, are not quite right for SpaceX. This is a stock that will trade on states of belief as much as financial estimates. The better framing is to ask which regime the stock is operating in, and to watch for evidence of transitions between them.

Financial Gravity

In this state, proof fails to keep pace with belief. Starship slips. Compute contracts prove temporary or reprice lower. AI capital spending remains heavy without a clear return. Grok continues to consume resources without strengthening the broader thesis. Starlink remains an excellent business, but not enough to carry every future embedded in the IPO price. The cash burn, $8.9 billion in Q1 alone, persists while contracted revenue proves less durable than hoped. Orbital compute remains a slide in a presentation.

This is not a “SpaceX fails” case. That distinction matters. SpaceX can remain one of the most important companies in the world and still disappoint public-market investors if the stock capitalised too much too early. In this state, SpaceX is valued as Starlink plus launch plus a discounted AI option. An indicative three-year range: $60 to $90 per share.

The Bridge Holds

In this state, proof accumulates at roughly the pace required to sustain belief. Starlink keeps compounding. Starship begins useful deployment, even if slower than hoped. V3 satellites improve capacity economics. Anthropic and Google-style compute revenue mostly holds. New compute customers appear, though investors still debate contract duration and margin quality. Grok improves enough to remain strategically relevant without becoming the centre of the thesis. The inherited debt from xAI and X is managed down without further distress.

The IPO valuation was demanding, but not fatal, because the AI segment starts to look less like a cash drain and more like a capacity business with identifiable customers. An indicative three-year range: $150 to $220 per share.

Escape Velocity

In this state, proof and belief reinforce each other. Starship cadence becomes real. V3 changes Starlink economics. Compute customers expand or are joined by new entrants. SpaceX becomes the fastest route to scarce AI capacity, and the market stops treating the AI segment as “Grok plus losses” and starts treating it as a way to monetise the physical constraints of AI demand. Orbital compute moves from concept to working prototype.

Belief funds capacity. Capacity wins customers. Customers validate belief. That validation supports the stock. The stock supports further capacity creation. The loop that worked in a smaller way for Tesla, and then for launch and Starlink, begins to operate again.

In this state, investors stop asking whether the multiple is high and start asking how large the next market can become. An indicative three-year range: $350 to $500 per share.

These are not predictions. They are states. The article’s purpose is to give the reader a framework for recognising which state they are in as evidence arrives, not to declare which state is most likely.

The First Future That Has to Arrive

The SpaceX IPO asks investors to pay for several futures at once.

One future is Starlink becoming a much larger global connectivity business. Another is Starship reducing the cost of orbit enough to change satellite economics again. Another is direct-to-cell expanding the reach of mobile networks. Another is Grok becoming a meaningful AI product. Another is Colossus becoming a durable compute business. Another is compute in orbit.

Not all of these futures need to arrive for the stock to work. But the first one does.

The first future is not Mars. It is not even Grok. It is proof that the Anthropic and Google deals are not merely temporary overflow demand, but early evidence of a repeatable business in selling scarce compute capacity. If that proof arrives, if contracts hold, if new tenants sign, if the physical bottleneck persists, the rest of the thesis gets room to develop. If it does not, financial gravity returns.

SpaceX is almost certainly expensive on normal near-term financial measures. It may also be one of the few companies where normal near-term measures do not capture the mechanism investors are actually buying. The right stance is neither blind faith nor reflexive dismissal.

The right question is whether proof is accumulating fast enough to keep belief productive.

Subscribe now

• Backward-looking quality metrics measure the old route, not the next one. High margins, stable earnings, and strong returns can persist long after a profit pool has started migrating.

• The key question in disruption is not whether a moat is real, but whether it travels. Teece’s insight matters: value accrues to whoever owns the scarce complementary assets needed in the new structure.

• AI will not kill all incumbents or save all challengers; it will create dispersion. The winners will be companies whose trust, workflow ownership, data, distribution, or ecosystem become more valuable as the route changes.

For centuries, Venice was the highest-quality asset in Europe.

It had the location, the fleet, the merchants, the institutions, the financing networks, and the Arsenal, perhaps the most advanced industrial operation of its time. More importantly, it had the route. Goods moving between East and West passed through the Adriatic, and Venice sat at the tollbooth. Its wealth was not simply a function of what it produced; it was a function of where it stood.

A fifteenth-century quality screen would have loved Venice. Durable competitive position, recurring trade flows, pricing power, institutional depth, high returns on capital, and centuries of proof. This was not a speculative asset. It was not a turnaround. It was the definition of proven quality.

Then the Portuguese rounded Africa.

Venice did not suddenly become incompetent. Its merchants did not forget how to trade. Its ships did not stop floating. Its institutions did not collapse overnight. The problem was simpler and more devastating: the route moved.

Once that happened, the very things that made Venice great became less relevant. The Arsenal could keep building ships. The bankers could keep financing commerce. The politicians could keep managing the republic. But the flow of value no longer needed to pass through the Adriatic. Venice’s quality had not disappeared. It had been stranded.

This is the central problem with quality investing during disruption. The spreadsheet measures the city. It does not tell you where the route is going.

The Missing Half of the Quality Trap

In my earlier piece, The Quality Trap, I argued that traditional quality screens are often museum catalogs. They identify the winners of the last era with remarkable precision, then mistake that precision for foresight.

The argument was straightforward. Companies with high returns on equity, stable earnings, low leverage, strong margins, and pricing power often look safest at precisely the moment their profit pools have begun to migrate. Newspapers in 2000 were wonderful businesses on paper. IBM in 2010 looked far cleaner than Microsoft. Legacy consumer brands had historical margins, distribution advantages, and brand equity that had compounded for decades. The numbers were not wrong. They were simply describing an economic structure that was already changing.

I still think that argument is right. But it was incomplete.

If every incumbent with great historical metrics is suspect when the route changes, then the conclusion is too blunt. Walmart survived Amazon. The New York Times survived digital media. Microsoft survived the transition from desktop software to cloud computing. In fact, Microsoft is the complication hiding inside my own argument: in 2010 it was itself a legacy incumbent, apparently past its prime, and yet it became one of the defining winners of the next era.

That means the real question cannot be “incumbent or challenger?” It cannot even be “quality or disruption?” Those are lazy oppositions. The harder question is this: when the route changes, which advantages travel?

Some moats become islands. They remain real, but the traffic leaves. Other moats become bridges. They allow a company to cross from one profit pool to the next.

The investor’s job is not to admire the moat. It is to determine whether anyone still needs to cross it.

Teece and the Theory of the Route

The cleanest way to understand this distinction was laid out by David Teece in his 1986 paper, Profiting from Technological Innovation.

Teece’s insight was simple, powerful, and still underappreciated: innovation and value capture are not the same thing. The company that invents a technology does not necessarily earn the profits from that technology. Those profits often accrue to whoever owns the complementary assets required to commercialize it.

The classic example is EMI and the CT scanner. EMI helped invent one of the most important medical technologies of the twentieth century. It had the breakthrough. It had the intellectual property. It had the technological lead.

What it lacked was the surrounding system. Hospitals did not just need a machine. They needed sales support, training, maintenance, procurement confidence, servicing infrastructure, regulatory credibility, and trust. General Electric had those assets. EMI had the invention; GE had the complements. The technology mattered, but the complement determined who got paid.

This is the intellectual foundation for thinking about quality during disruption. When a new technology appears, the exciting thing is rarely the only valuable thing. Often it is not even the most valuable thing. The durable economics accrue to the bottleneck around the technology: the asset that remains scarce after the innovation spreads.

This is why “follow the profit pool” is necessary but not sufficient. You also have to ask who owns the scarce complement needed to capture that profit pool.

Sometimes the complement belongs to the innovator. Sometimes it belongs to the incumbent. Sometimes it belongs to neither. The answer has little to do with who is old, who is new, who invented the technology, or who has the cleanest historical metrics. It has everything to do with who controls the bottleneck in the new structure.

This explains why Walmart survived and Borders did not. Both faced e-commerce. Borders had stores, inventory expertise, and a retail brand. Walmart had procurement scale, logistics infrastructure, supplier leverage, local presence, and physical nodes that could be integrated into an omnichannel model. In one case, the assets were stranded. In the other, they could travel.

It also explains why The New York Times diverged from most newspapers. The internet destroyed the local newspaper bundle, but it increased the value of trusted national brands. In a world of infinite content, scarcity moved from distribution to trust. The Times owned an asset that became more valuable after the route changed.

This is the refinement to the quality trap. The question is not whether the old moat was real. It usually was. The question is whether that moat controls a scarce complement in the new structure.

Why Traditional Metrics Invert

The Sparkline piece on AI disruption is valuable because it gives empirical scaffolding to what otherwise sounds like a collection of anecdotes. (https://www.sparklinecapital.com/post/ai-disruption)

Its core finding is simple and important: traditional value metrics work reasonably well in insulated industries, but struggle badly in industries exposed to technological disruption. That should not be surprising, but it is still underappreciated. In stable industries, low multiples often reflect mispricing. In disrupted industries, low multiples often reflect the market recognizing that current earnings are not durable.

The mechanism is brutal. Price moves first. Fundamentals move later.

A company facing disruption can look cheap for years. Revenue does not collapse immediately. Customers take time to leave. Contracts roll off slowly. Management cuts costs. Buybacks support earnings per share. Margins can hold up even as the strategic position deteriorates. The multiple compresses against earnings that still look real. That is the value trap.

But the same logic applies to quality.

A quality trap happens when the metrics remain beautiful because the old route has not fully decayed. Return on equity, operating margin, free cash flow, and earnings stability are all lagging indicators. They are evidence of what the company has been able to extract from the previous structure. They are not, by themselves, evidence that the next structure will require the same company.

This is why backward-looking metrics do not merely fail during disruption; they can invert. They select for the companies that look best under the old regime. But in a route change, the best-looking old-regime companies may be the most exposed. They are optimized for a path customers are abandoning.

That does not mean traditional quality is useless. It means its usefulness is conditional. In a stable structure, quality metrics are evidence. In a changing structure, they are questions.

Why are margins high? Because the company has a durable advantage, or because it is underinvesting? Why is free cash flow strong? Because the business compounds internally, or because management has no better use for capital? Why is leverage low? Because the company is disciplined, or because it is harvesting? Why are returns on capital high? Because the moat is widening, or because the denominator is being starved?

The same number can mean opposite things depending on whether the route is stable or moving.

Sparkline’s “intangible value” framework is interesting for precisely this reason. I would not treat it as a magic formula; any proprietary backtest deserves skepticism. But the direction is right. It attempts to measure the assets that Teece would have told us to look for: intellectual property, brand, human capital, and network effects. In other words, it tries to identify the complements that traditional accounting either misses, expenses, or misclassifies.

That is the real lesson. The answer to disruption is not to abandon valuation or quality. It is to redefine what must be valued.

Static, Stranded, and Traveling Quality

There are three kinds of quality during disruption.

Static quality is the easiest to identify. The business has great historical metrics. It screens well. It has high margins, stable earnings, low leverage, and a long record of profitability. Static quality is not fake; it is simply incomplete. It tells you the company was advantaged in the structure that produced the data.

Stranded quality is more dangerous. This is the business whose moat remains real but is attached to the old route. The company may still have brand, scale, distribution, assets, and customer relationships. The problem is that the customer’s route to value is changing in a way that makes those assets less central. The moat protects something, but not the bottleneck.

Traveling quality is what investors should be looking for. These are companies whose advantages become useful in the new structure. They may be incumbents or challengers. They may screen well or badly. The defining feature is that their scarce assets migrate with the profit pool.

This framework is more useful than the standard debate between quality and growth. Growth investors can overpay for companies that have the technology but not the complement. Quality investors can overpay for companies that have the complement to an old world. Value investors can buy optically cheap companies whose earnings are about to be structurally impaired.

The right question is more specific: what remains scarce after the technology spreads?

If the answer is “nothing the company owns,” the moat is stranded. If the answer is “the company’s trust, data, distribution, workflow, regulatory position, ecosystem, or implementation capability,” the moat may travel.

Software Is the Test

The software selloff is the cleanest current test of this framework.

The market’s fear is understandable. If AI makes software dramatically cheaper to build, then the economics of software must be re-examined. A product that once required large engineering teams may be built by a smaller team with AI assistance. Features that took quarters to develop may take weeks. Internal tools that companies previously bought may become easier to generate. The scarcity of code is falling.

If code was the moat, this is a problem.

This is why treating beaten-down software as a single bargain basket is dangerous. Some software companies are cheap because the market is panicking. Others are cheap because their business model depended on a scarcity that AI is dissolving.

The distinction is not obvious from the multiple. A software company trading at a low forward earnings multiple may be a bargain or a trap. The question is what the customer was really paying for.

Some software is vulnerable because it is a feature bundle. It automates a narrow task, has shallow workflow integration, limited proprietary data, weak switching costs, and pricing linked to work AI can now compress. These products may still be useful, but usefulness is not the same as pricing power. If AI makes alternatives easier to build, the product becomes less scarce.

Other software is different. It is a system of record. It sits inside a company’s daily operations. It carries permissions, history, approvals, audit trails, integrations, reporting, compliance, customer data, and employee training. The software is not merely a tool; it is a place where work happens.

For these companies, AI may be disruptive, but not necessarily destructive. If the vendor owns the workflow, AI can make the workflow more valuable. If the vendor owns the data context, AI can make that context more useful. If the vendor has trust in a regulated or mission-critical environment, AI can increase the premium on reliability. If the product is deeply integrated, AI features from the incumbent may be adopted faster than standalone alternatives.

This is the Teece question applied to software: if AI makes code abundant, what is the scarce complement?

For some companies, there may not be one. For others, the answer may be workflow ownership, proprietary data, compliance, trust, implementation capability, ecosystem depth, or customer context.

This leads to a practical test: if the cost of creating software fell by 90%, would the customer still choose this vendor?

If the answer is no, the moat was probably software scarcity. If the answer is yes, the moat is somewhere else.

That is why AI may create enormous dispersion within software. It will compress some companies and strengthen others. It will reveal which products were merely expensive because building software used to be hard, and which products were expensive because they occupied scarce positions inside customer operations.

The market’s first reaction is to sell the category. The investor’s job is to sort the category.

Adoption Versus Exposure

One of the more interesting points in Sparkline’s work is that software is not merely exposed to AI. It is also adopting AI aggressively.

That distinction matters. Exposure tells you where the threat is. Adoption tells you who is responding.

A company facing disruption has three choices. It can deny the shift, harvest the old model, or cannibalize itself. The third path is the only one that produces traveling quality, but it is also the hardest to recognize in traditional financial statements. It often depresses margins, increases expenses, and creates uncertainty before it creates proof.

This is why capital allocation matters more than commentary. Every management team says it is embracing AI. The question is whether the business is being rebuilt around it. Where is the company hiring? What is being integrated into the core product? Is AI a feature, a pricing lever, or a restructuring of the workflow? Is the company willing to weaken a legacy revenue pool to strengthen its position in the new one?

Microsoft’s cloud transition worked because it was not a press release. It was capital allocation. Azure absorbed investment before it produced the financial profile investors later celebrated. The old company funded the new route.

That is the pattern to look for today. The winners in software will not be the companies that mention AI the most. They will be the companies whose existing complements, data, workflow, trust, distribution, developer ecosystems, customer relationships, make their AI adoption more valuable than a competitor’s demo.

The losers will not necessarily be the companies that lack AI. Some will have AI features. Some will have impressive demos. The issue will be that their complements are weak. They will have product but not position.

Where the Route Goes

Venice’s tragedy was not incompetence. That is what makes the story useful. The city did not fail because it forgot how to be Venice. It failed because being Venice mattered less once the route changed.

That is the risk in every quality portfolio today. The businesses may be well managed. The margins may be real. The returns may be high. The balance sheets may be clean. The moats may even be deep.

But if the route is moving, the key question is not how good the city looks. The question is whether the route still runs through it.

That is the refinement to the quality trap. Backward-looking metrics fail because they measure the quality of the old route. The investor’s job is to determine whether the company’s moat is stranded on that route or whether it owns the scarce complement required on the new one.

Some companies will be Venice: excellent, historic, and bypassed. Some will be Walmart: attacked by the new route but able to bend its assets toward it. Some will be Microsoft: written off as a legacy incumbent while quietly owning the relationships, talent, and infrastructure needed for the next platform shift. Some will be EMI: inventors of the future without the complements needed to profit from it.

AI will create all four.

It will create companies with breathtaking technology and weak value capture. It will create incumbents with beautiful financials and stranded moats. It will create apparent bargains whose earnings have not yet caught down to reality. It will also create survivors whose old advantages become more valuable when carried into the new structure.

The old quality question was: is this a great business?

The better question is: will the things that made it great still matter where the profit pool is going?

That is the difference between static quality and traveling quality. Static quality tells you the city was rich. Traveling quality tells you the route still passes through it.

The spreadsheets will be the last to know.

Subscribe now

• The proof may not come as a software revenue line: We were waiting for Johansson to disclose Saab’s software/integration economics, but Q1 suggests the platform is first becoming visible in product architecture, especially C-UAS, where sensors, command-and-control, effectors, and upgrades are sold as one integrated system.

• Saab’s real flywheel is sovereign architecture: The bull case is not simply that Saab sells more defence equipment; it is that each radar, C2 layer, Gripen system, naval combat system, or counter-drone deployment makes the next Saab product easier to adopt by deepening training, doctrine, interoperability, maintenance, and political switching costs.

• Q1 strengthened the base case, not the victory lap: Organic growth, EBIT leverage, Surveillance strength, and operational cash flow all support the thesis, but free cash flow was still negative, Aeronautics remains a drag, and software disclosure is still missing, so Saab remains an “own and verify” stock, not a thesis fully proven.

I ended the last Saab piece waiting for a number.

Not a revenue number. Not an EBIT number. Not another backlog number. Saab already had those. I was waiting for the number that would either vindicate or embarrass the thesis we had been building for a year: how much of Saab is really software, systems integration, electronic warfare, data links, combat management, AI, and upgrade economics hiding inside a company still reported like a traditional defence contractor?

Q1 did not give us that number.

That should have been a problem. Instead, it made the story more interesting.

Our prior Saab arc was straightforward. Saab was no longer just a Swedish defence contractor. It was becoming Europe’s “sovereign stack”: a company whose value came not merely from selling fighters, radars, missiles, or submarines, but from connecting those systems into a NATO-compatible architecture that countries could operate and upgrade with greater independence. The missing proof point was disclosure. At some point, we expected Micael Johansson to point to a slide and say: here is the software layer, here are the margins, here is why Saab should not be valued like a normal defence contractor.

He still has not said it.

There are two ways to respond to that. The easy way is to say the thesis remains unproven. The better way is to ask whether we were looking for proof in the wrong place.

What If Software Is Not a Segment?

The mistake may be importing a software-company disclosure expectation into a defence architecture business.

Microsoft can report Azure because Azure is a product line. Adobe can report subscription revenue because the business model makes the software line visible. Saab is different. Its software is not consumed separately by a procurement officer with a credit card and a dashboard. It sits inside radars, fighters, electronic-warfare systems, combat-management systems, counter-drone architectures, naval platforms, training routines, command structures, and national doctrine.

The customer is not buying “software.” The customer is buying the ability to operate a defence system that can sense, decide, adapt, and remain under sovereign control.

That is why the software number still matters, but may not be the first place the truth appears. If Saab is right, the economics will show up first in how customers buy: not isolated equipment, but architectures. Not a radar, then a separate command system, then a separate effector, but an integrated chain from detection to decision to action. Not one national project, but a series of adjacent procurements that become easier because the country is already inside the Saab architecture.

The question after Q1, then, is not “where is the software disclosure?” It is: are customers behaving as if Saab’s integration layer is becoming the product?

Q1’s answer was not definitive. But it was suggestive.

The SEK 2.6 Billion Order That Matters More Than Its Size

The most important order in Q1 was not large enough to move the model.

Sweden’s Counter-UAS order was worth SEK 2.6 billion, with deliveries planned for 2027-2028. In a company with a SEK 274 billion backlog, that number is not the point. The architecture is the point.

Saab described the system as an integrated Counter-UAS solution, from “sensor to shooter,” modular, adaptable to evolving drone threats, built with Saab and selected third-party technologies, and including further development and integration.

That is the sovereign stack in miniature.

A drone appears. A sensor detects it. A command system classifies it. An effector is selected. The architecture adapts as the threat evolves. The customer does not simply buy a radar, a cannon, a missile, or a vehicle. The customer buys the system that makes all of those pieces work together.

On the call, Johansson described the system as flexible and effector-agnostic — able to work with different effectors, from Bofors rounds to Trackfire and eventually Nimbrix. That is the right kind of agnostic. Saab does not need to own every fist if it owns the brain that decides where the fist lands.

This is where Saab differs from the usual European rearmament story. Rheinmetall’s question is capacity: can it produce enough shells, vehicles, and land systems into a shortage? Saab’s question is architecture: can it become the integration layer for countries that want NATO-compatible capability without total dependence on any one superpower or prime contractor?

Capacity still matters. But for Saab, capacity is the cost of the thesis, not the thesis itself.

The Flywheel Is Boring, Which Is Why It Works

Defence switching costs are not created by delight. They are created by doctrine.

A country buys a Saab radar. That radar connects into Saab command-and-control. The next counter-drone system is easier to buy from Saab because the sensor layer already exists. The next naval system is easier if the combat-management layer is trusted. The next aircraft, electronic-warfare package, or upgrade becomes less like a fresh procurement and more like an extension of the architecture already in place.

This is not a social network effect. No one is posting selfies on Giraffe 1X. It is slower, more political, more bureaucratic, and much harder to see. But in defence, those features can make it more durable.

Training compounds. Maintenance routines compound. Data links compound. Local industrial participation compounds. Software upgrade paths compound. Operational trust compounds. The switching costs are not just financial; they are institutional.

That is why Q1’s Giraffe 1X commentary mattered. Johansson said Saab had manufactured some Giraffe 1X systems partly “on speculation,” rather than only waiting for signed contracts, which enabled fast deliveries to customers facing urgent missile and drone threats. In a bespoke defence world, that is an unusual posture. It suggests some Saab systems are becoming repeatable nodes in an architecture, not one-off national projects that begin from zero every time.

Naval points in the same direction. Saab’s new Naval business area, formed after the quarter, combines Kockums with Naval Combat Systems. That is not just reorganization. It is strategy made visible: platforms and combat systems belong together if the value is integration.

The Quarter After the Fireworks

Only now do the Q1 numbers matter.

Q4 was the fireworks quarter: Colombia Gripen, France GlobalEye, Poland submarines selected, Ukraine letter of intent, record orders, and a backlog that made the sovereign-stack thesis feel almost too neat. Q1 was quieter. Order bookings fell 5% to SEK 18.2 billion, and book-to-bill was below 1x. A sell-side recap would start there.

If this were merely a defence-cycle story, that would matter more. If it is an architecture story, the more interesting fact is that the system kept moving without another megadeal.

Sales rose 21% to SEK 19.2 billion, organic growth was 23.6%, EBIT rose 32%, and operational cash flow improved to SEK 1.0 billion. All business areas and Combitech delivered double-digit sales growth. Surveillance grew sales 32% and EBIT 52%, helped by Giraffe 1X deliveries and C-UAS. Dynamics remained the profit engine, with a 17.5% margin.

That is what we needed to see. Not perfection. Repeatability.

Saab did not need a giant new order to grow faster than 20%. It did not need a software disclosure slide to show operating leverage. It did not need a fighter headline for Surveillance to look increasingly like the place where sensors, software, command-and-control, and threat adaptation collapse into one business.

This is what I mean by the architecture becoming observable.

The Part That Still Bothers Me

This is where I do not want to let the thesis get too elegant.

Free cash flow was still negative SEK 301 million. Inventories rose. Tangible fixed assets rose. Saab is investing heavily in capacity, inventory, suppliers, and facilities because architecture still has to be manufactured.

Some of that is exactly what we want. If speed wins defence orders, and if customers need systems now, then carrying inventory can be strategic. The Giraffe 1X example suggests Saab is willing to take working-capital risk where it has demand visibility.

But there is a line between strategic inventory and bad inventory, and the income statement will not tell you where that line is. Cash will.

Saab’s medium-term target is cumulative cash conversion above 60%. Q1 moved in the right direction, but one quarter does not answer the funding question. If Saab’s architecture story requires too much capital to scale, the platform multiple will have a ceiling.

Aeronautics is the second constraint. Sales grew 15%, but EBIT fell 11%, and the margin compressed to 6.3%, hurt by R&D amortization, currency, and T-7 start-up costs. Gripen is central to the sovereignty narrative, but aircraft programs remain slow, complex, and capital-intensive. The base case can tolerate that. The bull case cannot ignore it.

What Changed in My View

Before Q1, I thought software disclosure was the next decisive gate. After Q1, I still think it matters, but I no longer think it is the only gate.

The cleanest proof would be a number. The earlier proof may be customer behaviour.

If C-UAS exports follow Giraffe 1X installations, if Naval shows better economics after combining platforms with combat systems, if Surveillance sustains margins above 10%, if countries that buy one Saab node return for adjacent nodes, then the architecture thesis will be proving itself before accounting catches up.

That is the update. We were waiting for Saab to separate software from hardware. Q1 suggests customers may increasingly be buying the place where software and hardware become inseparable.

Variant Perception

The consensus view is reasonable: Saab is an excellent European defence company, growing rapidly, but priced for a lot of success. Order intake was down, free cash flow was negative, Aeronautics disappointed, and the software number still does not exist.

The variant perception is not to dismiss those issues. It is to say they are not the central issue.

The central issue is whether Saab is becoming an architecture layer. If it is, every system sold increases the probability that the next system is also Saab. A radar creates pull-through for C2. C2 creates pull-through for counter-drone. Counter-drone creates pull-through for effectors. Gripen creates decades of training, maintenance, upgrades, weapons integration, and electronic-warfare dependence. Naval combat systems can do the same at sea.

The bull case is not that Saab sells more equipment. The bull case is that each piece of equipment makes the next piece easier to sell.

Three-Year Scenarios

Using a reference price around SEK 514, the setup is attractive but not obvious.

I would not raise the bull probability yet. Q1 did not give us software disclosure, a major new platform award, or clean free cash flow. What it did give us was a stronger base case and a lower probability that Q4 was merely order-book euphoria.

This remains an “own and verify” stock. The expected value is attractive, but the proof points still matter.

What To Watch

The first proof point is C-UAS exports. Sweden proves the concept; export orders prove the architecture.

The second is Surveillance margin. Above 10% sustained suggests integration economics are scaling. Below 8.5% suggests Q1 was mix.

The third is Naval margin from Q2 onward. The reorganization only matters if platforms plus combat systems produce better economics than platforms alone.

The fourth is Aeronautics margin. Above 8% would show Gripen economics surfacing. Below 6% would mean T-7 and program drag are still eating the story.

The fifth is cash conversion. Above 60% keeps the thesis self-funded. Below 50% turns the platform story into a capital-intensity problem.

And yes, the sixth is still software disclosure. It is no longer the only proof. It remains the cleanest one.

The Number Can Wait. The Architecture Can’t.

Q1 did not answer the question we asked. It suggested a better question.

We wanted Saab to prove the software layer by disclosing it. Instead, Q1 showed the software layer becoming visible in the architecture customers are buying: Giraffe 1X into C-UAS, C-UAS into command-and-control, Naval into combat systems, Surveillance into the integration layer.

That is not yet disclosed software economics. It is not durable free cash flow. It is not permission to ignore valuation.

But it is the right kind of evidence.

Saab is no longer trying to prove that countries want sovereignty. Q4 did that. Q1 asked whether sovereignty compounds when the headlines go quiet.

The answer is not final. But it is getting harder to dismiss.

Subscribe now

• The thesis moved from funding to throughput. Q4 proved sovereign customers would pre-fund Rheinmetall’s capacity; Q1 asks whether that funded capacity can turn into accepted deliveries, cash flow, and delivery credibility fast enough.

• The moat is intact, but the governor is visible. Demand and funding are not the issue; the bottleneck is the conversion layer, customer acceptance, lot testing, procurement timing, inventory-to-revenue mechanics, and management bandwidth across Naval, Air Defence, Digital, autonomy, and deep strike.

• At ~€1,207, the price finally makes the debate investable. The stock is no longer pricing most of the bull case, but Q2 is the proof point: if inventory converts and Q2 revenue validates management’s >50% growth claim, the flywheel works; if not, the “sovereign infrastructure” thesis starts looking like industrial overreach.

Rheinmetall reported Q1 sales of €1.94 billion, operating result of €224 million, an 11.6% operating margin, operating free cash flow of -€285 million, Rheinmetall Nomination of €4.87 billion, and backlog of €73.0 billion. That is the earnings recap. It is also the least useful way to understand the quarter.

The right place to start is with our own evolution. In Q3, we worried that Rheinmetall’s story was running into industrial physics: backlog was enormous, but cash burn and ammunition margin pressure suggested that being indispensable did not automatically mean being cash-generative. In Q4, we changed our view. The important discovery was not that free cash flow improved; it was how it improved. Sovereign customers were prepaying because they needed Rheinmetall’s capacity more than Rheinmetall needed incremental orders. We upgraded the moat from regulatory entrenchment to sovereign dependency: governments were financing the industrial capacity they could not replace.

Q1 does not reverse that conclusion. It makes it more precise. The funding debate is no longer the center of the thesis. The throughput debate is.

Q4 proved Rheinmetall’s customers would fund the moat. Q1 asks whether Rheinmetall can operate it.

The Factory Outran the Bureaucracy

The sell-side version of Q1 is easy: sales growth disappointed, cash flow was negative, and the company promised a stronger Q2. That is true, but too shallow.

Management’s explanation was unusually specific. Around €200 million of trucks were ready but awaiting German customer delivery timing; another €100 million of powder-related revenue slipped because lot acceptance testing pushed delivery into Q2. Papperger therefore guided to Q2 growth of more than 50%.

That is the key point: not the miss, but the specificity of the test. A vague “second-half recovery” is management asking for faith. “More than 50% growth in Q2” is management putting a date on credibility.

The quarter’s core tension is that the factory appears to be scaling faster than the conversion layer around it: customer delivery windows, acceptance tests, procurement procedures, and inventory-to-revenue mechanics. CFO Klaus Neumann said roughly €0.5 billion of inventory went into finished and pre-finished goods ready for Q2 shipment; almost half of the €1 billion inventory increase was tied to goods prepared for Q2.

If that inventory ships, Q1 was the balance-sheet shadow of growth. If it does not, Q1 was the first warning that the capacity flywheel has a governor.

The Mechanism Is Intact, Not Yet Confirmed

The fundamental question is whether Rheinmetall can turn customer-funded capacity into a durable sovereign infrastructure role. Europe has moved from a just-in-time defense model to a just-in-case security model. In the old world, budgets were the bottleneck. In the new world, budgets are increasingly visible; the bottleneck is deliverable industrial capacity.

Rheinmetall’s core capability is not “making tanks” or “making shells.” It is capacity credibility: powder, ammunition, propulsion, vehicles, air defense, digital systems, autonomy, and now naval, assembled into something governments can actually procure, deploy, and maintain.

That mechanism compounds if backlog and prepayments fund capacity, capacity improves delivery credibility, credibility wins larger integrated orders, and those orders deepen Rheinmetall’s role as sovereign defense infrastructure. It breaks if inventory does not convert, if acceptance delays become structural, if Weapon & Ammunition margins normalize faster than volume scales, or if Naval and Digital absorb capital without proving platform economics.

Q1 gave us evidence that the mechanism is still working. It did not prove it.

Operating profit grew 17% on 8% sales growth. Vehicle Systems grew only 3%, but margin improved to 9.6%. Weapon & Ammunition was flat on sales, but held 19.4% margin. Digital Systems grew 16%, with margin rising to 5.2%. Air Defence was the standout: more than 30% growth, with margin expanding from 12.5% to 15.6%. Naval contributed its first month at €77 million of sales and roughly 10% profitability.

That is not a broken quarter. It is an unfinished quarter.

The Integration Layer

The most important strategic development in Q1 was not in the income statement. It was in the widening of the architecture.

Naval closed in February. Management said the first month was in line with expectations, at about €80 million of sales and 10% EBIT, with profitability expected to move over time toward 15%. F125, F126, and Romania are the first tests of whether this becomes a real Sea domain or merely an expensive shipyard acquisition.

The same logic applies to autonomy and deep strike. Rheinmetall is not trying to invent everything internally. It is using partnerships as time arbitrage: Boeing for Ghost Bat, Lockheed for CCA cooperation, Indra for Spain, Kraken for unmanned surface vessels, Destinus for cruise missiles and ballistic rocket artillery. Papperger’s point was explicit: partnerships are faster than development programs, and Germany does not have time to wait for organic development of every new capability.

Viewed individually, each initiative can be justified. Viewed collectively, they reveal both the bull case and the risk. Rheinmetall is trying to become the industrial integration layer across vehicles, ammunition, air defense, ships, drones, missiles, sensors, and command systems. That is a better business than a product supplier. It is also a harder business to execute.

The company is becoming more important because it is becoming broader. It is becoming riskier for the same reason.

Variant Perception: Rearmament Is Not the Edge

The market already understands that European defense spending is going up. That is not the variant perception.

The better question is whether Rheinmetall is becoming the sovereign infrastructure layer that converts budgets into capability. The loop is:

backlog → prepayments and inventory → shipments → cash → capacity → delivery credibility → larger integrated orders → more backlog.

Q1 matters because it tests the middle of the loop. Demand is visible. Funding was validated in Q4. Now comes the harder-to-model part: physical throughput.

At €1,542 after Q4, we argued that the business had improved but the entry had not. The ceiling rose, but much of the upgraded thesis was priced. At roughly €1,207 now, the setup is different. The stock is no longer asking investors to pay for most of the bull case. It is closer to paying for the base case while preserving upside if Q2 proves the governor is bureaucratic rather than structural.

That does not mean the stock is obvious. It means the debate is finally worth having.

Three Years, Three Outcomes

By 2028, this should be clear. Capacity will either translate into output or it will not. Digital will either show platform economics or remain project revenue. Naval will either integrate or absorb management time. Arminius, ammunition, Air Defence, and Middle East demand will either convert or stay in procurement fog.

Assuming a current price around €1,207:

The probability-weighted value lands roughly in the high-€1,700s to low-€1,800s before dividends. That is a materially better setup than at €1,542. But it is still not a sleepy value stock. You are being compensated for execution risk, not spared from it.

What Has to Happen Next

The thesis is now unusually measurable.

First, Q2 revenue needs to validate management’s timing explanation. Above €3.2 billion would be a strong signal; below €2.8 billion would suggest the governor is more structural than bureaucratic. Second, operating free cash flow must improve in Q2 and H2; inventory has to be a bridge, not a trap. Third, Weapon & Ammunition margins need a clean baseline under the new structure; 28%+ would restore confidence, while below 25% would reopen the Q3 concern. Fourth, Arminius language matters: any move from Q2/Q3 to 2027 raises the bear probability. Fifth, Digital Systems needs at least one platform KPI by Q3, software or services mix, TaWAN attach rate, D-LBO milestone, lifecycle revenue, because without it the bull multiple is narrative, not evidence. Sixth, Naval margins above 8% through the year would suggest the acquisition works; below 5% would suggest shipyards are doing what shipyards often do.

What Changed

Our view did not reverse after Q1. It became more precise.

Q3 warned that backlog could run into industrial physics. Q4 showed sovereign customers would finance the capacity they could not replace. Q1 now asks the next and harder question: can Rheinmetall turn funded capacity into delivered capability fast enough to make the dependency permanent?

Rheinmetall’s problem is no longer finding demand. It is no longer, at least for now, finding funding. The problem is throughput.

Throughput is where rhetoric ends. Trucks must be accepted. Powder must be qualified. Naval programs must be integrated. Digital systems must prove recurring economics. Air defense conversations must become contracts. Inventory must become cash.

At €1,207, the price finally acknowledges some of that risk. Whether the business validates the price arrives in Q2. Everything else is commentary.

Subscribe now

• Baidu has crossed the first threshold: AI is no longer a side story. Core AI-powered revenue is now the majority of Baidu General Business, confirming our prior view that Baidu is not defending search so much as rebuilding itself around AI.

• The market may be using the wrong scoreboard: Baidu is not the largest Chinese cloud player: Alibaba has scale, Tencent has economics, Huawei has infrastructure, but Baidu may be much more relevant in the narrower AI-native compute layer where its full-stack approach matters.

• The second threshold is economics: AI Cloud proves demand, but applications, margins, and free cash flow will decide the multiple. The bull case requires Baidu to turn infrastructure growth into application-layer lock-in, not merely replace high-margin search ads with lower-margin compute revenue.

When we wrote “Baidu: The Elephant’s Dance,” the central argument was not that Baidu’s next quarter would be good. It was that the market was using the wrong mental model. Investors were judging Baidu as a declining search company; we argued that Baidu was doing something more deliberate and more dangerous: using search, distribution, data, and cash flow to build a vertically integrated AI company. The IBM analogy mattered because IBM’s transformation under Lou Gerstner was not an optimization exercise. It was an identity change. Baidu, we argued, was attempting the same: not defending search, but dismantling it to build something else.

1Q26 is the first quarter where that argument no longer lives only in metaphor. The elephant is not merely dancing in management commentary; it has crossed an accounting line. Baidu’s Core AI-powered Business reached RMB13.6 billion, up 49% year-over-year, and for the first time exceeded half of Baidu General Business revenue. AI Cloud Infra revenue rose 79% year-over-year to RMB8.8 billion, while GPU Cloud revenue grew 184%.

That is not a normal quarterly datapoint. It is the moment when a transformation moves from “management story” to “income statement reality.” But it also makes the debate harder, not easier. Before this quarter, the market could ask whether Baidu’s AI business was real. After this quarter, that question is less interesting. The better question is whether the AI business Baidu is building is economically superior to the search business it is replacing.

That is the second threshold.

The question is not growth. It is compounding.

The structural mechanism that could make Baidu more valuable is not simply “AI revenue growth.” The mechanism is a loop: intent data creates demand for AI answers and actions; those answers and actions create inference and cloud workloads; those workloads improve infrastructure utilization and model/application performance; lower cost and better performance attract more enterprise and consumer usage; and that usage produces more intent and workload data.

This loop strengthens at scale if Baidu’s stack — Kunlun chips, PaddlePaddle, ERNIE, AI Cloud, search distribution, agents, Wenku, Drive, and Apollo — makes each incremental AI interaction cheaper, faster, and more useful. It breaks if Baidu is merely replacing high-margin search ads with lower-margin compute revenue.

This distinction matters because there are two very different ways to read 1Q26. The first is simple: AI revenue is now big. The second is more important: Baidu may be trying to turn intent into infrastructure, and infrastructure into applications. The former is a revenue story. The latter is a platform story.

The metric is the message.

The most revealing part of Baidu’s release is not the headline revenue beat. It is the metric Baidu wants investors to focus on: Core AI-powered Business. Companies do not introduce new metrics by accident. They do it when old metrics no longer capture the story they want the market to understand — or when old metrics reveal too much pain.

That makes Baidu’s new frame both self-serving and useful. The old Baidu was online marketing plus iQIYI plus optionality. The new Baidu wants to be AI Cloud Infra, AI Applications, AI-native Marketing, and Apollo Go, with legacy search slowly receding into the background.

Source: Baidu 1Q26 results.
This is the quarter’s essential message: Baidu General Business barely changed in size, but it changed dramatically in composition. The company did not beat because the old business recovered. It beat because the new business became large enough to cover more of the old business’s decline.

Crossing 50% does not prove Baidu is an AI winner. It proves Baidu can no longer be analyzed as merely a declining search company.

The search business is not being saved.

The prior article argued that Baidu was not trying to preserve search in its old form. 1Q26 supports that view, but it also forces a more disciplined update: intentional cannibalization is not automatically value creation.

Online marketing revenue fell 22% year-over-year in 1Q26 and dropped to 48% of Baidu General Business revenue, down from 63% a year earlier. That decline should not be treated as an unfortunate side issue. It is the cost of the strategy.

The old Baidu monetized intent through links. The new Baidu wants to monetize intent through answers, agents, workflows, cloud workloads, and eventually transactions. A traditional search query ended with a click. An AI search interaction may end with a completed booking, a healthcare consultation, a shopping decision, a document generated in Wenku, a file workflow in Drive, or a business process executed by an enterprise agent.

That is a better strategic ambition than defending ten blue links. But it is also a harder business model to prove. Clicks were simple to monetize. Actions are more valuable, but also more complex: they require trust, orchestration, partner integrations, attribution, and new pricing models.

So yes, the first part of the thesis is working: Baidu is dismantling the old search model. The unresolved question is whether the replacement model captures more value than the one being dismantled.

The wrong scoreboard.

The biggest missing piece in a simple Baidu earnings analysis is the competitive scoreboard. Baidu is not transforming in a vacuum. It is fighting Alibaba, Tencent, Huawei, ByteDance, and others for position in China’s AI stack. The mistake is to treat that as one race.

On the broad cloud scoreboard, Baidu is not the winner. Alibaba has the scale. Its Cloud Intelligence Group generated RMB41.6 billion of revenue in the March quarter, up 38% year-over-year; external cloud revenue rose 40%, and Cloud Intelligence adjusted EBITA rose 57% to RMB3.8 billion. Alibaba is building AI on top of an existing cloud empire. That is a formidable starting point. (Morningstar, Inc.)

Tencent starts from a different fortress. It does not need AI to become a different company; it can insert AI into WeChat, games, advertising, fintech, productivity tools, content, and cloud. Tencent’s 1Q26 revenue was RMB196.5 billion, gross margin was 56.6%, non-IFRS operating profit was RMB75.6 billion, and free cash flow was RMB56.7 billion, even as capex reached RMB31.9 billion. More revealingly, Tencent disclosed non-IFRS operating profit excluding new AI products of RMB84.4 billion, which effectively shows a very profitable core business funding the AI build-out. (Tencent)

Huawei is harder to underwrite because the public numbers are opaque, but strategically it matters because it sits where sovereignty, chips, devices, telecom infrastructure, government relationships, and enterprise trust intersect. ByteDance has the consumer distribution, recommendation engine, and product velocity, but its advantage is more naturally consumer-AI than enterprise infrastructure.

Baidu’s position is narrower than Alibaba’s and less comfortable than Tencent’s. It is not the biggest hyperscaler. It does not have Tencent’s cash-flow fortress. What Baidu has is urgency and coherence: AI is not a feature being added to the business; AI is becoming the replacement operating system for the business.

The more interesting scoreboard is therefore not broad cloud. It is AI-native compute. Third-party summaries of Frost & Sullivan’s self-developed GPU cloud market report put Baidu Intelligent Cloud first with 40.4% share, ahead of Huawei at 30.1%. This is a narrower market than overall cloud, and the definition matters: self-developed GPU cloud is not the same thing as all public cloud. But that is precisely the point. Baidu looks small on the old scoreboard and much more relevant on the AI-specific one. (C114Pro)

That distinction reframes the quarter. Alibaba is building AI on top of cloud. Tencent is using AI to reinforce a cash-flow machine. Baidu is using AI to become a different company.

Compute is proof. Applications are control.

AI Cloud Infra is where Baidu’s full-stack thesis first shows up. Management has been explicit about this. On the 4Q25 call, Baidu described AI accelerator infrastructure as the primary growth driver of AI Cloud, emphasized its full-stack end-to-end architecture, and pointed to heterogeneous computing, unified scheduling, and Kunlun chips as sources of performance and cost efficiency. Robin Li also made the more important strategic point: “applications matter more than models.”

This is the correct answer to the model race. In a world of DeepSeek, Qwen, Doubao, Kimi, and open-source releases, simply having a model is not enough. Models alone are products with uncertain durability. The value is in deployment: how cheaply inference can be served, how reliably enterprise workloads can run, how quickly applications can be built, and how deeply those applications become embedded in customer workflows.

Baidu’s 1Q26 AI Cloud Infra number is therefore significant. It suggests customers are buying the infrastructure. But infrastructure is only the first proof. The final form is applications.

The distinction is crucial. Infrastructure revenue proves customers need Baidu’s compute. Application revenue proves customers are locked into Baidu’s platform.

That is where Baidu has more to prove. AI Applications revenue was RMB2.5 billion, approximately flat year-over-year and down 10% quarter-over-quarter. AI-native Marketing Services grew 36% year-over-year, but fell 15% sequentially. These are not fatal numbers, but they are the missing link in the bull case.

AI Cloud can make Baidu relevant again. AI applications are what can make Baidu valuable again.

The bill comes due.

The uncomfortable part of the quarter is that AI revenue is not free. The old Baidu was a high-margin intent monetization business. The new Baidu, at least in its current phase, looks more like infrastructure: servers, chips, depreciation, bandwidth, financing, and capex.

Baidu’s cost of revenue rose 7% quarter-over-quarter, primarily due to costs related to AI Cloud. Operating income recovered sequentially, and operating cash flow was positive at RMB2.7 billion, but the quality debate remains open. Bloomberg Intelligence’s post-result comment was blunt: better-than-expected earnings “masked” a 450 bp gross-margin miss, illustrating the dilutive impact of low-margin AI sales.

This does not invalidate the thesis. It clarifies it.

The absolute capex number matters less than the burden it carries. For Tencent, AI capex enhances the fortress. For Alibaba, it defends cloud scale. For Baidu, it funds the escape route from search. That makes the same kind of spending feel different. Tencent can spend into AI while the core throws off enormous free cash flow. Alibaba can spend because cloud scale is already a strategic pillar. Baidu has to spend because the old profit pool is being cannibalized.

That is the second threshold in financial form. The first debate was whether AI could become material. The second is whether material AI revenue can become high-return AI revenue.

What changed in our view.

1Q26 makes us more confident that the prior strategic diagnosis was right. Baidu is not pretending to be an AI company. The revenue mix has genuinely changed, and AI Cloud Infra is growing fast enough to matter. The market’s old question — “can AI offset search revenue?” — is increasingly stale.

But the quarter also makes us less willing to jump straight to the bull case. The prior article leaned into the elegance of the full-stack thesis: intent, model, framework, chip, infrastructure, application. That architecture still matters. It may matter more after 1Q26. But full-stack integration only creates value if it shows up in margins, cash flow, and application-layer monetization.

So the updated thesis is this: Baidu has crossed the first threshold. AI is now large enough to redefine the company. The second threshold is harder. AI must become profitable enough to re-rate it.

The variant perception is therefore not “Baidu AI is underappreciated.” That is too simple. The sharper version is: Baidu’s AI revenue is no longer the controversy. The controversy is whether that revenue compounds like a platform or depreciates like infrastructure.

Three futures.

At roughly HK$135–136, the stock still embeds skepticism. That skepticism is not irrational; it is the correct price for an unresolved transformation. The next three years depend less on whether Baidu can grow AI revenue and more on what kind of revenue AI turns out to be.

This keeps the spirit of the prior scenario framework, but raises the bar. The bull case is more visible after 1Q26, but it is not yet earned.

The scoreboard from here.

The key markers are not complicated. AI-powered revenue must remain above 50% of Baidu General Business; otherwise the new identity looks premature. AI Cloud Infra needs to remain above 50% year-over-year through 2026 to prove the Q1 step-up was not pull-forward. Gross margin needs to recover above 40%; below 38% keeps the utility bear case alive. Free cash flow needs to turn sustainably positive by 2H26. Online marketing decline needs to narrow to better than -15% year-over-year, or else cannibalization remains ahead of monetization.

Most importantly, AI Applications need to return to growth. That is the real tell. If applications stay flat while AI Cloud grows, Baidu may become a respectable infrastructure vendor. If applications reaccelerate, Baidu may become a platform.

That difference is the investment case.

Baidu has crossed the first threshold: AI is now large enough to replace the story of search. The second threshold is harder: AI must become profitable enough to replace the economics of search. The wrong scoreboard matters because Alibaba has scale, Tencent has economics, Huawei has national infrastructure, and ByteDance has attention. Baidu’s edge is narrower but potentially powerful: the conversion of intent into AI workloads, and eventually into actions.

The elephant is no longer standing still. The question is whether it is walking toward a better business, or merely away from the old one.

Subscribe now

• Right on the business, early on the stock: NU has compounded revenue, ARPAC, customers, Mexico, and AI; but the stock is flat because the market still prices it as a credit-sensitive bank.

• Q1 made the thesis testable: 15–90 NPL hit 5.0% and RA-NIM fell to 9.5%; management’s explanation is credible but now needs proof.

• Re-rating requires decision quality: NU gets a platform multiple only if AI/data-driven underwriting shows up in RA-NIM recovery, NPL normalization, and Mexico earnings.

A year ago, the Nubank thesis was elegant. Trust would compound into engagement; engagement would compound into data; data would improve underwriting; better underwriting would expand monetization; and higher monetization, layered on a low-cost digital architecture, would create one of the rare financial platforms that could grow like a technology company while earning returns like a great bank.

A year later, the business has largely followed that script. The stock has not.

That is the uncomfortable starting point for any honest Q1’26 update. Nubank’s quarterly revenue has moved from roughly $3.4 billion in Q1’25 to $5.3 billion in Q1’26. ARPAC has risen from $11.6 to $15.9. Customers have increased from 118.6 million to 135.2 million. Mexico has gone from promising expansion market to break-even business. AI has moved from a strategic claim to a production tool inside credit decisioning. Net income is up sharply year-over-year.

And yet the stock remains near the same low-teens range that framed much of our earlier work.

The easy answer is that the market does not understand Nubank. The better answer is more uncomfortable: the market may understand the upside, but it is not yet willing to underwrite the downside.

That is the tension Q1 brought back to the surface. We have been broadly right about the operating trajectory. We were too early — perhaps too optimistic — about when the market would pay for it.

What We Thought We Owned

Our view of Nubank has evolved over the past year.

In The Purple Way, the thesis was that Nubank was not simply a digital bank. It was a new financial institution built around customer trust, low-cost distribution, and a culture of product simplicity in markets where incumbents had trained customers to expect friction. The purple card was not just a product; it was a wedge into a banking system that had grown fat on fees, branches, and inertia.

By Q2’25, Mexico became the key proof point. The question was whether Nubank’s international growth was being bought with promotional deposit rates or earned through product utility. When Mexico customers largely stayed after yields were cut, the conclusion was powerful: trust had become a funding advantage. The cheapest capital was not a teaser rate; it was habitual use inside an app.

By Q3, the thesis moved from distribution to decision quality. NuFormer made the argument more ambitious: Nubank was not only acquiring customers more cheaply; it was potentially making better financial decisions because it had proprietary data, a unified technology stack, and machine-learning models trained on behavioral signals incumbents could not replicate.

Q4 sharpened the debate. The market sold higher provisions. We argued that the provision increase was largely mechanical: NuFormer-driven credit limit expansion front-loaded IFRS 9 expected losses before the associated revenue appeared. But we also said the important question was whether that limit expansion would convert into NII without credit deterioration.

Q1 did not ask a new question. It asked the same question again, with less room for elegance.

The Question Q1 Forces

The fundamental question is simple:

Is Nubank’s scale making its underwriting engine smarter, or merely making its credit book larger?

This is the entire investment debate.

If Nubank’s scale improves underwriting, then every new customer makes the system more valuable. More customers create more data; more data improves credit decisions, personalization, fraud detection, collections, and cross-sell; better decisions raise ARPAC and reduce losses; higher profitability funds more product development and market expansion. That is the compounding loop.

If, however, scale simply allows Nubank to extend more unsecured credit faster, then the story is less special. The company may still be an excellent digital bank, but it deserves to be valued as a bank: profitable, growing, but exposed to credit cycles and capital constraints.

This is why AI matters. Not because it makes customer service cheaper, although it may. Not because it makes the app better, although it may. AI matters because the core product of banking is judgment under uncertainty. Every credit limit, every loan approval, every collection intervention, every deposit rate, every product offer is a decision.

The old unit of competition in banking was the branch. Then it was the app. Nubank is arguing that the new unit of competition is the decision.

Q1 made that argument measurable.

Why the Market Sold First

The initial selloff after the Q1 print was not irrational. Revenue was strong, but the first-pass market reaction was not about revenue. It was about risk-adjusted profitability.

Risk-adjusted NIM fell to 9.5% from 10.5% in Q4. The 15–90 NPL ratio rose to 5.0% from 4.1%. Credit Loss Allowance rose to $1.79 billion. Adjusted net income was $937 million, down slightly from Q4 despite record revenue.

Those numbers trigger bank-investor muscle memory. The concern is straightforward: perhaps Nubank is not using AI to price risk better; perhaps it is using AI to expand risk faster.

Management’s response was credible. The company argued that the 15–90 NPL move was mostly seasonal, that growth and mix drove provisioning, and that 90+ NPLs actually improved to 6.5%. On the call, management said risk-adjusted NIM should move back toward second-half 2025 levels as Q1 dynamics normalize.

That explanation matters. It is also not enough.

The market is not simply pricing this quarter’s outcome. It is pricing the width of the distribution. Nubank’s AI underwriting may be superior. Cohort economics may be healthy. Coverage may be strong. But the market has not yet seen this model through a genuine adverse credit cycle.

That is the discount.

Why the Stock Came Back

The recovery from the lows should not be interpreted as the market accepting the platform thesis. It was more modest than that. Investors rejected the worst-case interpretation.

They had reasons. The credit bridge was detailed. Late-stage delinquencies improved. The seasonal pattern in early delinquencies had precedent. Mexico reached break-even. AI disclosures were more tangible than in prior quarters. And the broader market backdrop was supportive enough that buyers were willing to underwrite an explanation rather than extrapolate the first headline.

Still, the stock finished down. That matters too.

Q1 did not resolve the debate. It made the debate measurable.

What Changed in Our View

The long-term thesis is not broken. It is more conditional.

Before Q1, it was reasonable to argue that the market was over-penalizing provision optics. After Q1, that may still be true, but it now requires proof. The burden of proof has shifted from demonstrating the flywheel to narrowing the left tail.

That is a subtle but important change.

Revenue growth is no longer enough. ARPAC growth is no longer enough. Mexico reaching break-even is not enough. AI adoption metrics are not enough.

The next phase of the thesis depends on whether decision quality shows up in risk-adjusted returns. The market does not need Nubank to eliminate credit risk; that would be impossible. It needs evidence that Nubank’s data, underwriting, and cost structure make credit outcomes more predictable than they would be at a traditional lender.

That is also where our own framework needs humility. We have been right about the business direction. We were early on the re-rating. The market has not refused to see Nubank’s progress; it has refused to capitalize that progress at a platform multiple before the credit engine is proven through stress.

The variant perception, therefore, is not that Nubank deserves no credit discount. It is that the discount may be too large if the company can demonstrate that its data loop narrows the left tail.

Mexico Is the Counterweight

The reason Q1 is not simply a credit-risk story is Mexico.

Brazil proved that Nubank’s model can scale. Mexico is testing whether it can travel. Q1 was a real milestone: Mexico crossed 15 million customers, became the third-largest financial institution in the market, and reached break-even.

That does not yet force a re-rating. Break-even is validation, not materiality. But it changes the probability distribution. A year ago, Mexico was mostly evidence of transferability. Now it is beginning to look like the second earnings engine.

That matters because the bull case for Nubank has never been only Brazil. It has been that the company’s core capability — trust, data, underwriting, low-cost product manufacturing — can travel across markets. Mexico is the first serious test of that claim. The U.S. is still an option. Colombia is still early. Mexico is the bridge between ambition and evidence.

What Would Close the Gap

There are only three credible ways for the valuation gap to close.

First, Nubank needs a visible credit test. Not seasonal Q1 noise, but an actual period of macro deterioration where industry NPLs rise and Nubank’s outcomes prove structurally better. That would be the definitive evidence, but it cannot be manufactured.

Second, Mexico must become material. Break-even validates the playbook. A business generating $100–200 million of quarterly earnings would force investors to model a second Brazil.

Third, the company needs a pattern: three or four consecutive quarters of rising earnings, stable or improving credit metrics, and risk-adjusted NIM back above 10%. One quarter can be explained. A pattern gets capitalized.

This means the re-rating is probably not immediate. Nubank may be mispriced, but the catalyst is evidence accumulation, not one earnings call.

Scenarios

The scenarios are not just financial. They are classification outcomes.

The bear case is not that Nubank fails. It is that Nubank remains a bank. The bull case is not simply that revenue grows. It is that the market changes the category.

What We Track Now

The next few quarters need fewer metrics, not more.

The 15–90 NPL ratio needs to fall below 4.7% in Q2; if it stays near or above 5.0%, the seasonality argument weakens. Risk-adjusted NIM needs to move back above 10%; if it stays below 9.7%, growth quality remains in question. 90+ NPLs need to stay below 6.7%; a move above 6.8% would challenge the underwriting story.

ARPAC should hold above $15.5 and move toward $17 over time. Mexico needs to remain profitable. The efficiency ratio can normalize toward 20%, but core efficiency should continue improving. And AI needs to show up in cohort behavior, not just adoption metrics.

Early, Not Exonerated

The best theses are not the ones that never get challenged. They are the ones where the challenge clarifies what must be true.

Nubank’s Q1 did not break the thesis. It made it more honest. The company has executed like a platform, but the market is still valuing the unknowns like a bank. That is frustrating, but it is not irrational.

I still think we are early, not wrong. But early no longer means waiting for investors to notice revenue growth, ARPAC, Mexico, or AI. It means waiting for enough credit evidence to narrow the left tail.

If that evidence arrives, the platform multiple becomes available. If it does not, the market will have been right to keep Nubank in the bank bucket all along.

Subscribe now