• MongoDB may be in the wrong valuation category. Atlas is not primarily paid per seat; it gets paid for reads, writes, storage and compute. If agents create more machine work, MongoDB can benefit from the very force hurting traditional seat-based software. Pasted markdown

• AI is upside, not the underwriting. At today’s compressed multiple, roughly 20% growth, continued margin expansion and controlled dilution produce a base case of about $640 in three years. AI-native customers, Agent Engine and hyperscaler distribution are what could force the re-rating, not what the base case requires. Pasted markdown

• The real risk is not the CEO vacancy. It is whether MongoDB becomes the default persistence layer for agents. AI-native customers are only 3–5% of Atlas today; Postgres, AWS and Microsoft remain credible threats. By June 2027, we should know whether the reclassification thesis is becoming real.

Snowflake Lost Its Closer. The Price Followed Something Else.

On the afternoon of February 28, 2024, Frank Slootman did something celebrated CEOs rarely do on an earnings call: he announced he was leaving, explained why, and handed the microphone to his successor in the same breath. Slootman had taken Snowflake public in the largest software IPO ever, and investors treated him as the company’s closer. His explanation was blunt:

With the onslaught of Generative AI, Snowflake needs a hard-driving technologist to navigate the challenges the new role represents. Sridhar’s vision for the future and his proven ability to execute at scale made it clear to us as a board that he is the right executive at the right time to lead Snowflake. This marks my retirement from an operating role. I will remain on duty as Chairman of the Board and look forward to working with Sridhar and the team going forward.

The successor was Sridhar Ramaswamy, who had arrived nine months earlier when Snowflake bought Neeva, his AI search start-up. The same release guided product revenue growth to 22% against a Street wanting about 31%. Snowflake closed at $230.00 that day, $188.28 the next, and $167.75 within a week, down 27%. The market priced one story: Snowflake without its closer.

Here is the part that matters. The CEO change did not set Snowflake’s price over the next two years. The stock kept sliding and hit about $121 in April 2026, in the middle of the “AI eats software” sell-off, with Ramaswamy two years into the job. Then it rose 182% to $341.04 by October 2, 2026. Nothing about him changed in either stretch.

What changed was the drawer the market filed Snowflake in. In early 2026 it sat with every other software company: businesses whose customers pay per person, and whose customers need fewer people once agents do the work. By autumn it had been moved: a business that gets paid for the work a computer does and so gets paid more when agents do more work. The CEO discount was the sideshow. The classification set the price.

Thirty-one months later, the same script opened at MongoDB, with one difference: the market has not yet decided which drawer MongoDB belongs in. So here is the question this piece answers: not whether MongoDB has an AI product, but whether MongoDB gets paid when AI does more work.

MongoDB Lost 29% Without Losing the Business

From its August 13 high of $472.29 to its September 28 close of $334.68, MongoDB lost 29% of its value in six weeks, concentrated in two bad days. Neither was about the business getting worse.

On September 1 it reported a quarter most companies would frame: revenue of $771.8 million, up 30.5%; Atlas, the managed cloud service that is about three-quarters of revenue, up 29% for the sixth quarter running; operating margin of 24% against a 21% consensus; the full-year Atlas guide raised by 300 basis points. The stock fell 13.5%, because buyers positioned for acceleration wanted Atlas growth starting with a three.

On September 28, CJ Desai, CEO for less than eleven months, resigned with immediate effect to run Meta’s new enterprise business, reporting directly to Mark Zuckerberg, the day before MongoDB’s own Investor Day. The stock fell 18.5% on 7.8 times normal volume.

Then look at what followed. Guidance was reaffirmed the same day. Dev Ittycheria, CEO from 2014 to 2025, returned as interim. At Investor Day the next morning, management raised its three-year targets: revenue growth of 20% or more, Atlas in the mid-20s, margins expanding 100 to 200 basis points a year. The board added $1 billion to the buyback. FY28 EPS revisions stand at 36 up, 1 down. None of that is a company whose CEO saw trouble coming. It is a company whose CEO got an offer from Mark Zuckerberg.

The Street reads it as a holding pattern. Stifel says shares won’t re-rate “pending signs of durable acceleration for Atlas and a permanent CEO.” Capital, One calls it “a crowded long” that lost the man responsible for its alpha. That story, a good business with a leadership hole, is true. It is also the less important half.

Billed Like Snowflake, Filed Like Salesforce

MongoDB started in 2007 as a database developers chose because it matched how they wrote code: flexible documents instead of rigid tables. In 2016 it launched Atlas, the managed cloud service that runs on all three hyperscalers. In 2018 it changed its license to stop those clouds reselling its software; by then it was already selling the service itself.

The thing that matters about Atlas is how it bills. Not per person who logs in, but for the work the database does: reads, writes, storage, compute. A company that charges per seat loses when software replaces the people in the seats. A company that charges for work done earns more when software does more work. As a chain: more agents means more applications, more applications mean more database calls, more calls mean larger clusters, and larger clusters mean more Atlas revenue, with a lag of a quarter or two at the cluster step, because most Atlas spending runs through provisioned capacity rather than per-operation metering. Nearly every software company calls AI a tailwind. Very few can name the metered unit that increases when an agent acts. MongoDB can.

The market ignored the distinction in early 2026. When Anthropic’s Claude Cowork plugins landed, the IGV software ETF’s holdings lost nearly $1 trillion of value in a week, and MongoDB fell with everything else, to an intraday low of $215.68 on April 10. The recovery then sorted the sector: from the April lows to October 2, Snowflake rose 182% and Datadog 163%. MongoDB rose 66%. You can see the misfiling in the Street’s own words; KeyBanc, staying sidelined after a strong quarter, was

cautious about competition from Postgres and how MongoDB’s subscription model compares to consumption-based peers.

Atlas is the consumption model. It is billed like Snowflake’s and filed like Salesforce’s. That gap is the opportunity.

Agents Are Better Customers Than People

The case for reclassification starts with management’s own framing. Ittycheria at Investor Day:

So, when it comes to AI, I think it’s very important for all of you to understand why AI is a tailwind for three basic reasons. One, more workloads. With AI, you can produce far more software far more quickly. So by definition, more software applications require more databases and more data persistence stores. So that’s point number one and you’re going to hear a lot more about this. Two, new workloads. We’re going to address a broader set of workloads around agents, memory, knowledge work. The breadth of use cases expands not just from software but taking over knowledge work that people were doing manually. And the third is a brand-new set of customers. We’ve already expanded into the frontier labs. We’ve seen a lot of AI native companies start to use MongoDB.

The evidence, in order of how much it proves. Customers with at least one AI use case now account for about 40% of Atlas annual recurring revenue, up from 30% a year ago; that means AI has reached the customers who generate 40% of the revenue, not that AI generates it. Among Atlas customers above $100,000 a year, multi-feature use, mostly full-text and vector search, has climbed from 34% to 42% to 48% over two years. Vector search customers have grown fourfold. And the company disclosed a cohort of about 2,200 AI-native customers: 3% to 5% of Atlas revenue, with growth of 67%, then 129%, then 165% over the last three periods.

The arithmetic on that last number is the core of my view. Let the cohort grow 100% a year, well below its current pace, while Atlas grows 25%: it reaches about 6.4% of Atlas in one year and about 10% in two. Hold the current 165% and it passes 8% within a year. At 10%, AI stops being an argument on an earnings call and becomes a revenue line, and the market has to reclassify the company that owns it. One caution: the cohort is defined by who the customer is, not what the workload is. It measures the pull of AI companies toward MongoDB, not agent traffic itself.

Now the bear case that actually matters, and it is not the CEO. Agents are good for databases. That does not make them good for MongoDB. The real contest is the question my earlier piece was built around: who owns the default persistence layer for agentic applications? Today, coding agents often reach for Postgres first, and the biggest data companies paid to own that reflex, Databricks about $1 billion for Neon and Snowflake about $250 million for Crunchy Data. Second, AWS and Microsoft could keep agent memory inside their own stacks; MongoDB says it is in discussions with both about building its Agent Engine memory layer directly into their AI services, and discussions are not contracts. Third, MongoDB’s own pricing: every Atlas customer is already entitled to Agent Engine, “no new contract, no new sales motion,” so MongoDB earns nothing unless agents actually burn more usage, and management’s own slide puts meaningful Agent Engine revenue in FY29. The company has lived the gap between winning workloads and earning from them: in May 2024 Ittycheria admitted “we really indexed on volume,” and the workloads won were “growing a little slower than we expected.”

My position: agents will use databases far more intensively than people do; the open question is whose. The evidence for MongoDB is real and small, 3% to 5% of Atlas, and I think it becomes undeniable within two years. The end of this piece says exactly how to know if I am wrong.

The Street’s Math Doesn’t Add Up

The Street’s own model offers a cross-check. I want to size it honestly, because it is not a pillar.

Consensus is not blind on margins. FactSet has revenue of $3.02, $3.61 and $4.34 billion for FY27 through FY29, with operating income of $635 million, $797 million and $1.03 billion: margins of 21.0%, 22.1% and 23.8%, expanding about 110 and then 170 basis points a year, inside management’s guided range. The margin story is in the price.

The anomaly is one floor down: the Street’s operating line and its EPS line disagree. Take its own FY29 operating income of $1.03 billion, add roughly $140 million of interest income on $3 billion-plus of cash, tax at the 20% non-GAAP rate, divide by 86 million shares: about $10.85. Consensus EPS for the same year is $9.82. Making both true requires a 27% tax rate or a share count near 95 million while a $1.35 billion buyback runs. Even before any reclassification, the Street is slightly understating the earnings conversion. Caveat: 13 analysts’ model FY29 operating income against 37 for EPS.

On top of that sits a habit. Berry at Investor Day:

So our commitment was on average 100 basis points to 200 basis points a year. We went from 14.9% in 2025 to 18.5% in 2026 and we’re guiding to approximately 21%. That was 360 basis points growth from 2025 to 2026 and 250 basis points for this year. We were super clear last year, the 100 basis points to 200 basis points was not a ceiling.

Delivered 360 and 250 against a promise of 100 to 200, and the initial revenue guide beaten by 5% to 13% every year. Earnings quality is turning too: trailing free cash flow of $659 million now exceeds stock compensation of $564 million for the first time, stock comp has fallen from about 30% of revenue to 19%, and the trailing year is GAAP-profitable. Expensive on owner earnings, no longer imaginary.

MongoDB Needs a Seller, Not Another Visionary

The next six to twelve months of revenue do not need a CEO. They run on existing customers expanding (net revenue retention is 122%) and self-serve sign-ups (about 2,900 net new customers a quarter). And the interim is not a caretaker: Ittycheria took the company from about $35 million to $2.3 billion-plus in annualized revenue and built Atlas.

What left with Desai was access, not product. Three weeks before resigning he told a Goldman conference that C-suite awareness of MongoDB was low, and that AI architecture decisions are made top-down by chief data and AI officers. Analysts credited his relationships for the record $90 million and $100 million-plus deals, which sit mostly in Enterprise Advanced, a quarter of revenue. His departure revealed the one thing MongoDB now lacks, and it is not vision but executive distribution: the ability to be in the room where the database behind a company’s agents is chosen. The product bench, two new Chief Product Officers, is in place. The board should hire for the room.

There is one signal I cannot explain away. Insiders filed 102 stock transactions in the last six months, all sales, none open-market buys, including after the 18.5% drop. If the people closest to the business thought $335 was absurd, one of them could have said so with their own money. None did.

You Don’t Need the AI Bull Case to Own It

MongoDB trades at about 8.4 times next-twelve-months sales and 48 times earnings. Across sixteen software names, its 19% expected growth should fetch about 10.2 times sales; 8.4 prices it as a 17% to 18% grower. The consumption cohort trades far higher: Snowflake 15.9 times, Datadog 19.0.

Read the table below as two lists. What you are paying for at today’s price: roughly 20% growth, margins expanding as guided, a buyback that holds the share count, at the multiple the CEO exit already compressed. What you are not paying for: the AI-native cohort scaling, Agent Engine shipping inside AWS or Microsoft, and the market refiling the company. The first list is the base case. The second list comes free, and it is the bull case.

Methods: exit multiples apply to the next twelve months at October 2029 (¼ FY30 + ¾ FY31); the sales multiple is enterprise value to NTM revenue, net cash added back; EPS is operating income plus interest income, taxed at 20%, over the shares shown. Shares fall to 84M where the buyback offsets dilution, rise to 92M in the bear where it cannot.

The bear is the Teradata path, replacement rather than supplement, and its falsifiers are named below. The bull needs the reclassification, at a multiple still below where Snowflake and Datadog trade today. Probability-weighted: about $630 in three years, about $445 in twelve months.

March Will Look Worse Than the Business Is

In March 2027, the fourth quarter laps last year’s $90 million and $100 million-plus multi-year deals, booked upfront under the accounting rules, so headline growth prints near 14%. The FY28 revenue guide will start at or below consensus, because the guide always starts low. Likely still no permanent CEO. It will read like the bottom of the story while the underlying ARR has not changed, which means the worst headline of the next twelve months may coincide with the best price of the next three years. The number to read that day is the margin guide: consensus embeds about 110 basis points for FY28; 150 or more means management intends to beat the Street again on the day the narrative looks worst; below 100 means the earnings cross-check fails, and I will say so.

The Next Re-Rating Is About the Business Model, Not the CEO

Two and a half years after Slootman handed over the microphone, nobody prices Snowflake on who replaced him. They price it on what the business became. MongoDB’s answer arrives the same way, on dates already known:

My call, in a form you can check: by June 30, 2027, AI natives are at least 7% of Atlas revenue, which requires the cohort to keep compounding above roughly 115% a year, and MongoDB’s agent memory is generally available inside AWS or Microsoft, within the window Berry gave. Both happen, and the stock trades above $500. Neither happens, and the base case still holds, because it never needed them; the bull case, though, is dead.

That is the asymmetry. MongoDB does not need agents to work for the stock to work. It needs agents to work for the market to realize it has been valuing the company incorrectly.

The market priced MongoDB’s CEO in September. Over the next two years it will price the drawer.

Subscribe now

TL; DR

• Micron has converted scarcity into unusually strong forward visibility: strategic agreements now cover an estimated 35%+ of revenue through 2030, financial commitments reached $32 billion, and more than 75% of FY27 output is already committed. That improves planning without guaranteeing equivalent profit protection. Pasted markdown

• Differentiation and near-term economics are not the same thing: HBM deepens customer dependence and brings Micron into co-design with NVIDIA, but conventional memory is currently enjoying even stronger scarcity margins. The key test is whether technical differentiation ultimately earns superior returns. Pasted markdown

• The investment case now shifts from peak earnings to cash retention: Micron says it intends over time to return 100% of excess cash to shareholders. The cycle does not need to disappear for value creation to persist if exceptional cash generation is converted into sensible repurchases before supply catches up.

A customer signing a five-year memory agreement is making a judgment about more than memory prices. It is deciding that the cost of missing supply could exceed the savings from waiting for a better deal. For an AI infrastructure operator, that calculation includes processors, networking and power infrastructure whose productivity depends on having enough memory. Micron’s opportunity begins with that customer calculation; its investment case depends on what the company does with the resulting bargaining power.

In January, I argued that AI was making memory more strategically important and its most advanced forms harder to replace. In June, The Cycle Gets a Contract described Micron’s attempt to turn that importance into customer commitments before competition and new capacity caught up. Q4 advances both arguments but introduces a third: exceptional profits need not persist indefinitely if enough cash reaches shareholders while they last.

That is the position I would take after this quarter. Micron has bought greater visibility into future demand, and customers have accepted more obligations to secure supply. The task now is to convert that visibility into disciplined investment and capital returns. A longer shortage helps, but shareholder value ultimately depends on how much of its proceeds the company retains and how intelligently it uses them.

The Calendar and the Contract

Micron reported $54.23 billion of revenue, an 87% adjusted gross margin and $33.2 billion of adjusted free cash flow. The next-quarter revenue guide is $61.5 billion. On the surface, sequential growth falls from 31% to 13%; however, Q4 contained fourteen weeks and Q1 contains thirteen.

Revenue per week rose approximately 21.5% in Q4, and the Q1 guide implies another 22.1% increase. That is a materially stronger operating trajectory than the headline comparison suggests. It gives us evidence of continuing momentum, although one actual quarter and one forecast cannot establish a lasting growth rate. The more consequential evidence concerns what customers are committing to beyond the next report.

Strategic customer agreements increased from sixteen to twenty-six, covering an estimated over 35% of revenue through 2030. Financial commitments rose from $22 billion to $32 billion, predominantly deposit commitments; approximately $12.7 billion of customer deposits had been received by fiscal year-end. Management also disclosed approximately $150 billion of remaining performance obligations, calculated using committed volumes and minimum pricing for agreements with determined pricing frameworks.

The company describes why these agreements matter:

These multi-year take-or-pay agreements sharpen our long-term supply planning and enhance the durability and predictability of our strong financial performance. Further, they provide our customers supply assurance and deepen technology roadmap collaboration.

The final phrase connects the commercial change to January’s technical argument. Co-design can produce a qualified product that a customer is reluctant to replace; commitments make investment in that product easier to justify; successful manufacturing adds experience and improves delivery. If each generation strengthens the relationship, Micron acquires something more durable than an attractive spot price.

The boundaries matter because June’s language sometimes ran ahead of the disclosures. Financial commitments are not all collected cash, and a target for greater coverage is not existing protection. Today, three-quarters of agreement-covered revenue has a defined pricing framework, most such frameworks have floors and ceilings, and the remainder is periodically repriced. Separately, more than 75% of fiscal-2027 output is committed across agreement and non-agreement customers. That establishes substantial volume visibility without establishing equivalent profit protection.

The Price of Differentiation

The strongest challenge to an uncomplicated AI story appears in the Cloud Memory business:

CMBU gross margins were 83%, flat sequentially, driven by higher pricing, offset by higher HBM mix.

More HBM diluted the segment’s gross-margin percentage relative to the alternative mix. Conventional memory was earning such exceptional scarcity margins that the more specialized product was less profitable on this measure. That distinction matters: the products that deepen customer dependence need not be the products extracting the greatest immediate benefit from a shortage.

Calendar-2027 HBM agreements include significant price increases intended to narrow that gap. Beyond repricing, Micron’s NVHBM work with NVIDIA places it inside the design of future computing platforms. Technology chief Scott DeBoer described that work on the analyst call:

We’ve been working with NVIDIA for over a year on HBM4E, what’s called NVHBM. We see substantial opportunity there for us in the co-design of that product with obviously a key customer to have this be a product that delivers substantial value beyond standard HBM4E.

The opportunity is real, but its economics must be negotiated. HBM4E’s base die uses a foundry process, unlike Micron’s in-house HBM4 approach, adding another participant with a claim on the product’s value. BNP Paribas analyst Karl Ackerman asked who captures the economics of customization; management expressed confidence in attractive returns without quantifying the division. The relevant test is whether pricing and manufacturing execution turn technical participation into profitable differentiation.

NAND complicates the commodity-versus-franchise distinction further. NAND accounted for most of the revenue surprise against our frozen expectations, but data-center SSDs represented more than two-thirds of NAND revenue. Controllers, firmware and qualification matter there too. Contract coverage and differentiation are separate dimensions of the business: conventional products can carry long commitments, while sophisticated products can retain pricing risk.

The Factory Is an Option

Customer commitments also make new capacity easier to finance. Micron’s spending outlook implies more than $50 billion of net capex in fiscal 2027, with much of the incremental construction supporting cleanrooms available in late 2028 and beyond. This extends June’s supply clock, but does not assign an expiration date to scarcity. Demand and the pace of equipment installation still determine what those buildings produce.

CFO Mark Murphy made that distinction when Wolfe Research’s Chris Caso asked about construction spending:

I think you made a very important point that the spend there doesn’t translate into bits, and that these fabs will be made. We put the equipment in the fabs and produce wafers when needed based on our view of the market. These SCAs are a good way for us to keep a pulse on the market and make sure that we get a return on that CapEx.

Construction creates an option to add supply; installing equipment exercises it. This gives Micron discretion, although buildings commit capital before production begins. Contracts could improve that discretion by connecting investment to customer obligations. They could also encourage several suppliers to expand simultaneously, which is why equipment decisions will eventually matter more than descriptions of construction as disciplined.

Meanwhile, capacity is not the only claimant on the profits. Management identified roughly $1 billion of additional Q1 costs from incentive compensation, startup and other items. Taiwan union representatives are separately seeking a permanent arrangement allocating 15% of operating profit to employee bonuses. The demand is not a settlement, and the company’s global compensation changes cannot simply be attributed to it; nevertheless, the dispute illustrates how exceptional profitability attracts competing claims from within the production system.

What Shareholders Keep

This brings us to the disclosure that changes the investment argument. Management intends to increase capital returns from December 9, and Murphy stated:

Over time, we expect to return 100% of our excess cash to shareholders.

The word “excess” carries the investment decision. Customer deposits have repayment conditions, factories require funding, and an authorization is not a completed repurchase. But if Micron produces substantial cash after those obligations, buying shares at prices that discount a severe earnings decline can improve the remaining shareholders’ economics even when profits subsequently normalize.

That is our investment argument: the shortage does not have to become permanent for some of its benefits to become permanent per share. It requires cash conversion and sensible repurchase prices. A valuation that ignores interim cash misses’ part of this opportunity; one that adds cash already spent on buybacks counts the same benefit twice.

The following fiscal-2029 scenarios therefore track cash and shares together. They are our assumptions, with returns measured from the September 30 pre-release close of $1,065.11 and revenue CAGR from fiscal-2026 actual revenue of $133.19 billion.

The bear assumes earnings settle at a substantially lower level; its multiple does not presume a rapid recovery from a temporary trough. The base assumes strong growth followed by margin normalization, while the bull requires differentiation and contracts to preserve unusually high profitability. A 65% operating margin makes even the base case demanding. The higher bull repurchase price also means that spending more retires proportionately fewer shares.

The cash bridge starts with $55.5 billion of net cash after reserving received customer deposits. Shares start at 1.15 billion; we assume a 17% tax rate, zero net interest, 30 million shares of cumulative dilution and $3 billion of dividends. Across bear/base/bull, three-year gross capex is $150/$185/$220 billion, depreciation is $63/$69/$72 billion, and net working-capital cash use is $0/$9/$21 billion. Repurchases reduce both cash and shares. Terminal cash reserves of $35/$50/$65 billion are excluded from value, and only remaining excess cash is added to capitalized operating earnings. Deposits and grants do not fund the modeled shareholder returns; dividends are excluded from displayed returns.

The Window

Our view has strengthened because the commitments are broader and the route from earnings to shareholder value is clearer. At the $1,065 reference price, I favor a measured position, with additions below $1,000 if the operating evidence remains intact. That stance depends on cash actually being generated and allocated well; it does not require us to declare the memory cycle abolished.

There are specific tests. Q2 gross-margin guidance above 86.25% would support management’s expected recovery after Q1. First-half free cash flow after gross capex reaching at least 70% of adjusted net income would support conversion through the investment increase. By the first-half results, I would want evidence of executed repurchases and a quantified capital-return framework, rather than another statement of intention. These are our underwriting hurdles, not additional company promises.

Contract concessions or repeated realized-price declines alongside rising supply would weaken the case more fundamentally. A modest quarterly margin miss would prompt investigation; persistent failure to translate commitments into post-investment cash would change the thesis. June showed that customers would sign for future supply. September makes the shareholder opportunity more concrete: Micron can use the interval of scarcity to strengthen both its customer relationships and each remaining share’s claim on the business. The contracts create the opportunity. Capital allocation will determine how much of it shareholders keep.

Subscribe now

• The key shift is from communication to execution: Meta already creates the introduction and hosts the conversation between customers and businesses. Agents could reduce the work that still sits between an enquiry and a completed order. Pasted markdown

• The business agent may be what makes the personal agent dependable: a customer’s agent can ask, but only the business side can make authoritative commitments on stock, price, delivery and policy. Giving both sides agents reduces ambiguity rather than merely adding more automation. Pasted markdown

• The real thesis is a network effect, not an AI product bundle: Meta becomes more valuable if each participating business expands what customers can reliably accomplish and reusable capabilities make later deployments cheaper. If every new merchant adds bespoke work instead, the network does not compound.

Imagine a woman selling sarees from her home in Surat. She posts a video of a new collection on Instagram and pays to promote it. A customer in Pune taps through to WhatsApp, where the familiar questions begin: the price, the available colours, more photographs, whether delivery by Friday is possible. The seller checks her stock, works out the shipping charge and sends payment details. After the money arrives, she packs the order and handles the messages about tracking, alterations and exchanges.

In this transaction, Meta has helped create the introduction and provided the place where the relationship develops. The seller and customer still carry information between messages, payment applications and delivery services themselves. What looks like one conversation is a small operating process held together by two people. The seller does not need more ways to communicate. She needs less work between an interested customer and a completed order.

That distinction changes how we read Meta’s recent announcements. In June, Meta invested $900 million in CRED and appointed its founder, Kunal Shah, to lead WhatsApp. On September 8, it launched Muse, a personal agent with a dedicated computer in the cloud. On September 28, it announced an enterprise business under CJ Desai, bringing personal and business agents together with developer tools. Each development fitted a familiar category: payments, consumer software, enterprise software. Their possible relationship was less obvious.

We did not anticipate these moves and initially treated the enterprise initiative as a potential distraction from the consumer product. That may have been the wrong starting point. When Apple introduced the iPhone through three familiar descriptions in 2007, the significance lay in their combination. The comparison is not a prediction of equivalent success; it is a reminder to ask whether apparently separate capabilities make one another more useful. For Meta, the question is whether agents can turn its network of people and businesses into a network for getting things done, where participation on each side improves the other and creates value Meta can capture. Pasted text

The ambition that failed

Facebook had an earlier answer to where the relationship should live. In April 2013, it introduced Home, an Android experience that placed friends’ updates on the lock screen and home screen. Its launch announcement asked, “if sharing and connecting are what matter most, what would your phone be like if it put your friends first?” The ambition was to make Facebook more central to the device by reorganising the experience around relationships rather than applications.

The problem was that the applications performed jobs. Someone looking for a boarding pass or transferring money was not expressing a preference for a less social phone; they were trying to accomplish something specific. Home confused the importance of the relationship with the usefulness of placing that relationship first. Putting friends ahead of the applications did not relieve the user of operating them, and could put another experience in the way.

Muse proposes a different arrangement. Meta describes it as an agent that can use a browser, operate connected services and continue working after the person closes the application, returning when approval is required. Instead of asking the user to put Meta ahead of the tools, it offers to use the tools on the user’s behalf. The computer doing that work runs in Meta’s cloud, although access to devices and outside services remains consequential.

The distinction is between making the relationship prominent and making it useful. A customer might express an intention through a conversation and let the agent determine which applications to use, what information to collect and when a decision is needed. The interface can be generated or selected around the task instead of becoming another system the customer must learn. But there is a limit to what the customer’s agent can achieve alone: it cannot know what the seller is genuinely prepared to do merely by becoming better at asking.

The other side of the conversation

The enterprise announcement is more interesting when read from the customer’s end of the problem. Rather than begin with a new technology market, Zuckerberg begins with relationships Meta already serves. His opening paragraph places the business ambition directly beside those existing relationships:

We believe superintelligence will create significant new opportunities for all people and businesses. Meta already serves billions of people at scale and helps hundreds of millions of businesses reach customers. Today we are starting the next major pillar of our business, Meta Enterprise Platform, to help businesses use AI to grow and transform in new ways as well. Pasted text

Our reading is that the business side could help fulfil the promise made to the individual. The customer in Pune can ask her agent for a maroon silk saree below ₹5,000 that arrives before the weekend. The agent can organise the search and send the questions, but it still needs valid information about stock, price and delivery. A reply is not yet an agreement, and a fluent answer cannot reserve a product.

Now give the seller an agent connected to her catalogue, available stock and operating rules. It could respond with a particular saree, an all-in price, delivery terms supported by the courier’s service and a time-limited reservation. The customer’s agent can compare that offer against the request and obtain approval. The business agent does not remove physical uncertainty; it makes the commitment explicit and connects it to what the business can authorise.

Nothing about the saree has changed. What changes is how confidently the parties can proceed. The customer’s agent has less ambiguity to resolve, while the seller no longer must reconstruct the same information for every enquiry. Straightforward checks can remain ordinary software operations; the agent handles the varied request and coordinates the appropriate actions. Giving the business an agent may therefore be part of making the customer’s agent useful, not simply another opportunity to sell computing capacity.

A recent difficulty shows why these matters. Reuters reported that Meta tested human contractors for some Muse calls, then rolled back the test following employee objections. An internal executive said human callers achieved success rates of 95% to 98% in some tests, without providing a numerical AI-only baseline. An employee also reported that an insurer kept hanging up when it recognised an AI caller. This does not prove the proposed solution, but it shows that completion depends partly on the counterparty, not only on the model making the request.

OpenAI’s March change to Instant Checkout offers another caution. Its announcement said the initial implementation lacked the desired flexibility and shifted toward merchant-owned checkout experiences, including a Walmart environment supporting accounts, loyalty and payments. The episode points to a requirement, not a predetermined winner: connecting a consumer interface to commerce means accommodating the business’s operations and its relationship with the customer.

Seen this way, Kunal, Muse and CJ could address complementary requirements: familiarity with commerce, execution for the individual, and dependable business participation. Their combination remains an interpretation rather than a disclosed organisational blueprint. Its attraction is that improving one side could improve the other, without requiring every participant to adopt simultaneously. A human customer benefits from a capable business agent; a personal agent can still use conventional websites and services.

What becomes more valuable

Meta’s existing reach can help people encounter an agent. That is a distribution advantage, but it is not yet the reason the product should improve with scale. The stronger mechanism begins when participating businesses make useful actions available: customers can complete more work, successful delegation encourages repeat use, and the resulting demand gives additional businesses a reason to participate. Their participation expands what the service can accomplish.

Personal memory is different. Remembering a customer’s preferred colours makes her agent better for her but does not necessarily improve anyone else’s experience. A merchant offering dependable stock and fulfilment information can improve the experience for relevant customers across the network. Reusable connections can also reduce the cost of serving additional businesses. Personalisation, network effects and engineering economies reinforce one another, but they are not the same advantage.

Nor does every additional merchant make every customer better off. Another seller matters when it adds a relevant product, useful availability, better service or an alternative the customer can use. Counting accounts would obscure the real asset: the range of outcomes that participants can complete dependably. More messages are not sufficient evidence, particularly when the purpose of the product is to reduce how many messages a person must manage.

This creates a demanding test. Does bringing a business into the service make it more useful to existing users, or simply give Meta another customer to support? For a small seller, installation must become inexpensive enough to justify the expected return. Larger customers can support substantial implementation work, but that work should create reusable capabilities or sufficient profits. Self-service is an attractive route, not a universal requirement; continually rebuilding the same functions is the warning sign.

The customer is already here

The seller in Surat already has a reason to consider the product: she wants more of her enquiries to become orders. A large consumer company can approach it for the same reason. Marketing wants better returns from customer acquisition, sales want completed transactions, and service wants fewer unresolved interactions. The company need not first decide to replace its internal technology; it can begin with a bounded opportunity in how it serves customers.

Meta has been developing that commercial connection. Its July 2025 announcement brought marketing across WhatsApp, Facebook and Instagram into Ads Manager, with common creative, setup flows and budgets. The same announcement discussed business AI that could recommend products, facilitate sales and follow up in WhatsApp. This predates the new enterprise division and provides a concrete route from an existing commercial relationship into a broader service.

In that setting, the chief executive’s question is how to pursue a customer opportunity within acceptable limits. Technology leadership still matters, but an architectural preference is not a complete reason to forgo profitable business. Equally, commercial enthusiasm is not evidence of incremental revenue. Both teams must account for their preferred decision: the risks of implementation and the business lost while waiting.

This also explains why Meta’s financial opportunity need not depend on building a large standalone subscription business. Suppose, purely illustratively, the seller receives twenty enquiries from a promotion and converts five. If faster responses, accurate availability and reliable follow-up help her convert eight at comparable margins, the same advertising expenditure becomes more productive. She might choose to spend more, pay for business capabilities or do both.

The agent could therefore improve the economics of Meta’s existing relationship rather than merely be subsidised by it. That is not automatic: an order moved from a website into WhatsApp is not necessarily a new order, and money shifted between Meta products cannot be counted twice. The relevant gain is incremental commercial value after delivery costs. Nor does the mechanism require private personal-agent information to enter advertising systems; Meta says Muse’s conversations and VM data are not shared with those systems.

The department she never had

The same seller also reveals why the consequences could extend beyond companies with large technology budgets. An established retailer has specialised teams for sales, service and operations; our home-based merchant has herself. An agent might make a department more efficient in the first case and make some of its capabilities affordable for the first time in the second. Alexandr Wang’s explanation of Muse reaches beyond an efficiency calculation:

Most people never say what they want. It’s not because we don’t want. We want fiercely and constantly. It happens in the moments before falling asleep or while daydreaming on the car ride home. We want to spend more time with our families, overcome social anxiety, eat healthier food, open a bakery, build an app, or just find a few more moments of calm in a hectic life. But rarely do we get to even express these dreams, and even more rarely do they ever come true. Pasted text

The claim is aspirational, but the business implication is concrete. Our seller might handle more enquiries, communicate with customers in another language and manage repeat orders without first building a larger organisation. Meta’s May 2026 Business AI announcement in India describes multilingual customer responses, lead capture and appointment booking within WhatsApp Business. Its selected merchant testimonials are not representative outcome data, but the product is explicitly aimed at this gap in capabilities.

That is the geographic possibility worth keeping in view. An economy need not produce frontier models to benefit from making more of its businesses commercially capable. The constraint may be neither an absent product nor an uninterested customer, but the difficulty of bringing them together reliably. Better participating businesses would widen what consumers can accomplish, feeding the same network mechanism rather than creating a separate national-growth thesis.

When more makes it worse

The customer chooses maroon, then changes to green after the seller has started preparing the order. The price differs, the original courier arrangement may no longer work, and the customer’s approval covered the previous version. Now someone must determine what was agreed, what can be changed and whether Friday delivery remains possible. The quality of the next sentence matters less than the accuracy of the underlying state.

This is where a network can accumulate obligations instead of advantages. If every merchant brings a different collection of exceptions requiring bespoke engineering and continual intervention, growth makes the service harder to operate. Desai’s potential contribution is not merely enterprise credibility; it is helping turn business requirements into dependable products. The question is whether later deployments reuse what earlier ones required Meta to build.

Meta’s technical design recognises part of this distinction. A separate system, Sentinel, decides whether proposed actions may proceed, should be denied or require approval. That separates the intelligence proposing work from the authority permitting it. The purpose is not to interrupt every step, but to put checks where they matter and allow appropriate routine actions to proceed. Building the control is necessary; demonstrating dependable outcomes under changing conditions is the harder test.

A second constraint is loyalty. The customer wants the right saree on acceptable terms; the seller wants a profitable order. Meta can help them reach an agreement without pretending their interests coincide. But if the customer’s agent favours the merchant paying Meta most, it risks weakening the relationship on which future delegation depends. Privacy protects information, while loyalty concerns whose interests determine the choice. The two promises must not be confused.

Finally, useful activity can remain expensive to provide. Computing, failed attempts, support, human intervention and capital investment all belong in the calculation. Neither high usage nor an expanding transaction count establishes an attractive return. The thesis requires repeat activity and reusable capabilities to support economics that improve, while recognising that agents on both sides can communicate without being owned by the same company. Meta has an advantageous starting relationship, not a guaranteed claim on every outcome.

The pattern and the proof

The evidence through September 2027 should begin to distinguish three separate propositions. The first is behavioural: do people entrust agents with recurring responsibilities and spend less total effort supervising them? The second is the network mechanism: as relevant businesses participate, do comparable tasks complete more reliably, require less intervention a`nd encourage further delegation? Improvements caused solely by a better model would strengthen the product but would not by themselves establish a benefit from business participation.

The third proposition is financial. By March 2028, can Meta demonstrate that it captures incremental profit through business services, transactions or improved advertising economics after the full delivery cost? New disclosures about delegated tasks would help identify management’s priorities, but definitions and outcomes would matter more than the decision to report them. Insufficient disclosure leaves a question unresolved; it does not turn an attractive narrative into evidence.

Our expectation is that this develops through bounded responsibilities rather than universal autonomy. Businesses make specific commitments executable, people delegate more after repeated success, and reusable capabilities connect those responsibilities into increasingly useful arrangements. Meta should prioritise those connections over the breadth of its product catalogue. If business adoption grows without improving consumer outcomes or delivery economics, the network thesis is wrong even if individual products remain useful.

Return to the seller in Surat. At the beginning, her conversation with the customer was a record of work they were doing elsewhere: checking stock, agreeing terms, confirming payment and arranging delivery. The opportunity is to make that conversation a dependable means of performing the work, giving the seller greater capacity and the customer fewer reasons to intervene. If each additional capable participant improves that experience for others, Meta’s existing relationships take on a new economic role.

Facebook Home tried to make the relationship central by placing it ahead of the applications. Agents offer a different route: make the relationship more useful by operating the applications and involving businesses that can act on the other side. Whether Meta delivers that outcome remains to be seen, but it is the right ambition to evaluate. The prize is not another application people open; it is a network through which they can accomplish more together.

Subscribe now

• On has done something unusual in footwear: it scaled from CHF725 million to CHF3.0 billion while raising gross margin from 52% to 65% and largely avoiding the markdown cycle that has historically destroyed challengers. Pasted text

• The mechanism is more than product innovation: athlete credibility supports premium pricing, full-price discipline protects desirability, and high margins fund the next round of product, stores and marketing. The crucial gauge is whether discounts stay absent as the company scales. Pasted text

• 2027 is the real test: Nike has returned, US run-specialty weakened, inventory is running ahead of sales, and On has already sacrificed shipments to protect full-price integrity. If the refreshed running range restores sell-through without markdowns, the engine is real; if not, the missing discount was merely delayed.

The bill from 1972

Phil Knight had a problem that every fast-growing shoe company eventually meets: he was selling shoes faster than he could pay for them. Nike in the early 1970s had to pay its Japanese factories months before American retailers paid Nike, the banks had stopped extending credit, and growth itself was pushing the company toward insolvency. His solution, described in Shoe Dog, was to change who carried the risk:

Why not go to all of our biggest retailers and tell them that if they’d sign ironclad commitments, if they’d give us large and nonrefundable orders, six months in advance, we’d give them hefty discounts, up to 7 percent?

Nike called the program Futures, and it worked so well that it stopped being a program and became the industry. As late as 2015, 87 percent of Nike’s wholesale footwear shipments still flowed through Futures orders, four decades after the cash crunch that invented them.

But Futures carried a clause nobody wrote down. When a retailer commits to shoes six months before they arrive, somebody is guessing what people will want to wear half a year in the future, and the guess is always wrong by something. When it is wrong in the expensive direction, the extra pairs do not disappear; they go to the back room, then the sale rack, then the outlet, then the off-price chains. The discount was never really a marketing decision. It is how footwear pays for its forecasting errors, and the bill lands hardest on challengers, who must buy shelf space with volume before they know what sells. That is a large part of why the list of new footwear brands that reached even $4 billion in the last fifty years is so short, and why the ones that did mostly died on the way up. Under Armour got there in 2016, hit a wall in US wholesale, pushed the excess through off-price chains, and taught its customers to wait; a decade later its sales are below where they started. Allbirds, worth $4 billion at its 2021 IPO, marked its inventory down until gross margin fell from 53 to 41 percent, and agreed this March to sell its footwear business for $39 million.

Which is what making the company this article is about so strange. Between 2021 and 2025 a Swiss firm grew from CHF 725 million of revenue to CHF 3.0 billion, roughly 43 percent a year, much of it through the same wholesale doors, under the same advance-order system, against the same incumbents. And the bill never showed up: no sale rack, no off-price channel, a gross margin that rose from 52 to 65 percent while the average shoe climbed from $145 to $170. Either On Holding built a machine the industry has never seen, or it caught the best weather the industry has ever offered. The halved stock price says weather. Working out which it happens to be worth about thirty dollars a share.

What On was actually selling

Start with the founding insight, because it is more precise than the garden-hose legend suggests. In 2010 Olivier Bernhard, a retired Swiss Ironman champion, did glue sections of garden hose to a running sole in pursuit of a soft landing and a firm push-off in the same stride. But his co-founder Caspar Coppetti brought something else to that kitchen table, and he described it at last week’s Investor Day:

I had also brought a spreadsheet to that kitchen. We had studied the market, and we had found something very surprising. Passionate runners were investing more into their gear, into their sport every year. They were basically willing to pay for progress. The most valuable customers in the category. However, there was no brand that catered to them exclusively. I think that was the moment where it all made sense to us. We would disrupt running with innovation, and we would only sell at the premium end.

Everyone believed running shoes were a saturated market because Nike and Adidas served everybody. That was exactly the opening: brands that serve everybody cannot cater to the customer who wants to pay more, because their whole apparatus, from factory volumes to wholesale terms to $60 price points, is built for the middle. The premium runner was underserved in plain sight, and a company with no factories and no middle to defend could serve her exclusively.

But the insight that made On a niche winner is not the one that made it a CHF 3 billion company. The hollow pods that Bernhard engineered for cushioning turned out to have a second property that mattered more: you can see them from across a room. Most performance technology is invisible, a foam formula buried in a midsole, which is why shoe marketing historically needed athletes and airtime to explain it. CloudTec explained itself. The shoe was its own advertisement on the shelf, on the train, on a colleague’s feet, and its design chief later compressed the whole approach into a sentence: engineering solves a functional requirement, and design makes the solution desirable. What that combination produced was a product most of its buyers never use as intended. The people in Clouds and Cloudtilts at the office, the airport and the school run are not runners; they are buying quiet, affluent, health-adjacent identity with a Swiss engineering story attached, a signal of self-care that a logo sneaker cannot send. The performance is the alibi; the identity is the purchase. That is why the price could rise without volumes falling, and it is the first reason the discount went missing: the person buying identity does not comparison-shop the way the person buying cushioning does.

Today that product is a $10 billion company by market value, selling in more than 90 countries, still 92 percent footwear, with the Americas slightly over half of revenue, Europe about a quarter, and Asia-Pacific, the fastest-growing fifth, compounding at over 50 percent last quarter. Apparel, with accessories, is about 7 percent and targeted to triple. The mix matters later; hold onto it.

The engine

At the Investor Day, Allemann posed the question every investor asks and gave an answer worth quoting in full, because the rest of this article is essentially a test of whether it is true:

People ask us all the time, what makes On On? What is the secret behind the success? Here is the honest answer. A single piece of On can be copied. A foam can be reverse engineered. An athlete can sign with somebody else, right? A beautiful store can be built. What cannot be copied is how the pieces reinforce each other.

Strip the mission language and the pieces do form a loop you can trace through five years of results. Athletes validate the product at the highest level: Hellen Obiri won Boston and an Olympic medal in On prototypes she helped build, Iga Świątek won Wimbledon in On tennis gear, and Federer, who joined in 2019 as an owner and co-engineer rather than an endorser, gave a nine-year-old brand the single most credible name in premium sport. Validation permits the premium price. The full-price rule protects the desirability the price depends on, because nothing corrodes a status good like a sale rack. The 65 percent gross margin that results funds the athletes, the labs, roughly 100 stores in the right cities, and marketing at 14 percent of sales, which produce the next visible product, which earns the next validation. Performance funds the story, lifestyle funds the company, and margin funds both.

Two details show the loop was engineered rather than lucky. The first is sequencing: On earned its stripes in run specialty, the small stores where staff fit serious runners and a brand’s credibility is minted, and only from 2023 scaled into DICK’S, Foot Locker and JD, by which time it arrived as the hot brand rather than the supplicant. The second is that the loop has now run twice. Tennis, entered through Federer in 2019, has compounded at 61 percent a year since 2021 and carries On’s highest customer lifetime value by management’s account; the same playbook, athlete first, premium only, lifestyle crossover after credibility, is now being loaded for football, with Mbappé signed away from Nike, and golf. A flywheel that works once may be circumstance. One that repeats in a second sport is a machine.

And the discount is the gauge on the machine. As long as the loop generates slightly more desire than On supplies, the discount stays missing; the moment supply runs ahead of desire, the discount reappears, first at the edges, then everywhere, and the margin that funds the whole loop unwinds. Watch the gauge and you are watching the engine. Which is why, before August, the reading had been so remarkable: On repeated “full price” on every one of twenty earnings calls, and by 2022 Allemann was saying the quiet part aloud, that “having a certain amount of product scarcity helps us remaining premium and also helps our margin situation.”

The weather

Here is the uncomfortable part for the bulls, and the reason this article exists rather than a victory lap: the decade that machine spun up in was the most favourable stretch of weather a footwear challenger has ever been handed, and every element of it is now expiring.

The first and largest was that Nike left. In June 2017 it announced the Consumer Direct Offense, and that October said it would concentrate on roughly 40 of its more than 30,000 retail partners. The company that invented the wholesale system walked away from most of it to sell directly, emptying exactly the shelves where runners shop, and then spent the following years recycling legacy silhouettes while its running innovation went quiet. A challenger did not have to beat Nike for space and newness; Nike surrendered both. On’s CMO was disarmingly honest at the Investor Day about what that meant:

We do not have an archive. We cannot go down into the dustbins in the basement and pick out something from decades before. But what we know from consumers is that other brands can do that. They want newness. They want innovation. They want freshness. They are actually pretty hungry for a brand like us to come along with something very, very different.

Newness was On’s only inventory, and it happened to be the one thing the incumbent had stopped supplying. Second, the culture moved toward the product: wellness became a status symbol in the pandemic years, applicants to the London Marathon went from 120,000 to 1.3 million, and, as Coppetti argued last week, luxury houses priced out some 80 million aspirational customers whose spending went looking for a new signal. Third, social media collapsed the cost of building a brand; a Swiss startup could reach American buyers through athletes, a Zendaya contract and a Loewe collaboration instead of two decades of television, which is the only reason a challenger could afford the awareness game at all. Fourth, the price lane was pre-cleared: Lululemon had trained the affluent to pay premium prices for activewear and Hoka had normalised the $150 everyday running shoe, so On’s $170 asked customers to step up a rung, not build the ladder.

Count what has expired. Nike has a new chief executive, rebuilt wholesale relationships and, by its own account, gained five points of running share in North America and Western Europe in fiscal 2026, more than any top-five brand; Hoka claims share gains above $120 in the same months. The wellness boom continues but now lifts every boat on the shelf. Attention is no longer cheap: awareness of 30 percent, up from 12 in 2023, must now be defended with that 14 percent of sales against competitors spending multiples in dollars. And the price lane is crowded with the brands that cleared it. The machine, in other words, spent five years running with a tailwind, and nobody, including management, knew its true speed. Then, this spring, the wind stopped.

11 August 2026: the first quarter without weather

The American market turned promotional in the spring of 2026 as competitors cleared aging stock, and the big chains matched prices to keep customers. Ed Stack, executive chairman of DICK’S, described the mechanics from the retailer’s chair:

The margin pressure is real as these legacy silhouettes have slowed down, and then some of the brands started discounting, and the discounting got pretty aggressive. And we felt it was really important for us to stay with the market from a price standpoint... they say: We can buy this shoe for $25 less expensive on such and such a site. We didn’t want them to go someplace else to buy that product.

On’s everyday running shoes, late in their product lives, sat next to rivals that had just gotten $25 cheaper, and the rule forbade a response. Worse, the trouble in the channel that mints credibility had started earlier. At US run-specialty stores, On’s dollar sales fell 19.7 percent in the first nine months of 2025, the worst of any top-ten brand, while Nike’s rose 35 percent, and On’s president Scott Maguire had discovered why in the most analog way possible, spending a day watching shoppers in one of those stores:

We watched as people came in. They picked up our shoes and then put them down. Some chose On, and some chose others. We went super deep on why. Some of the answers were in the product, some were in the storytelling, some were in the channel, even some was the information on the side of the shoebox. The store staff couldn’t quickly work out what product they were pulling. We were losing out exactly when it mattered.

Read that twice, because it locates the problem precisely. On was not losing on price, and the brand was not fading; the company was losing on product, presentation and even the box label, in the store where the performance job is won, while an ageing range met a rejuvenated Nike. The engine’s most important cylinder was misfiring even as the group grew 30 percent.

So, when the promotional storm hit, On faced the exact choice its system was built for, and on 11 August it chose. The company cut its full-year outlook from at least 23 percent constant-currency growth to the low twenties, its first guidance cut as a public company, raised its gross margin outlook in the same release, and explained: “We held back sell-in rather than build inventory in the channel that could potentially compromise On’s full price integrity down the line.” The stock fell 20 percent that day and reached $26.76 by mid-September, 47 percent below its January high. The market looked at a decelerating challenger with a US problem and priced the Under Armour script.

Here is where precision matters more than conviction. Holding back shipments is not avoiding the error, because shipping fewer shoes you have already made is not the same as making fewer shoes. On’s inventory ended the quarter at CHF 473 million, up 31 percent year over year against sales growth of 14 percent, which tells you the shoes largely exist; the company moved the mistake from the retailer’s shelf, where it would have been settled in markdowns, onto its own balance sheet, where On controls the clock. That is the cheaper currency only if the shoes eventually sell at full price. Clearance so far looks normal, the outgoing Cloudsurfer 2 at $95 in a single colourway, but if that 31 percent becomes markdowns in 2027, the discount was never missing, just late. August proved that management will sacrifice growth, its stock price and a five-year beat-and-raise reputation to protect the gauge. It did not prove the protection was free. For what it is worth, the people who can see the inventory voted: all three founders bought stock in the open market in May, and two bought again on 14 August, three days after the cut.

The 2029 wager

Six weeks later, standing in Zurich in front of the investors who had halved its stock, management, run since May by Allemann and Coppetti as co-chief executives with a new finance chief beside them, raised the stakes: at least CHF 5.6 billion of revenue by 2029, gross margin of 65 percent or better throughout, an EBITDA margin of at least 22 percent built from named cost lines rather than price rises, and a first buyback of up to $1 billion. The plan is less heroic than it sounds. It needs the Americas to grow only in the low teens, distribution costs have already fallen from 13.4 percent of sales to 10, and the biggest product refresh in the company’s history, every running franchise renewed around the lessons of Maguire’s store visit, lands in full by mid-2027. This is not a plan that requires new weather. It requires the engine to work at twice the size, and two of its assumptions deserve a harder look than the sell-side has given them.

The first is hiding in the plan’s own arithmetic. On sits in about half its key partners’ doors, up from 20 percent in 2023, and intends roughly 75 percent by 2029, which means growing the door count around 14 percent a year, while the planned channel mix implies wholesale revenue growing about 13 percent. Side by side, those numbers mean the plan assumes revenue per door flat or slipping for three years, across a widening and necessarily weaker set of stores, each new door another advance-order guess of precisely the kind this company is supposed to be taming. Sam Poser of Williams Trading put the objection to management directly, “Getting items placed is different than having good sell-through,” and he is right: flat-per-door expansion is the exact shape the error took at every brand that ended up on the rack. Management’s reply is that the ceiling is deliberate, in Rebecca Cai’s words that “while we could easily be in all doors tomorrow, 75 percent is the premium distribution that we believe fits our brand,” and that weekly colour-level sell-through data lets it see trouble early. Perhaps. The engine has instruments no challenger ever had; the question is whether anyone throttles back when the instruments say so, because the last time the choice arose, in August, they did, and it cost them a third of the company’s value. They will be asked again.

The second assumption is that the credibility engine restarts. Recall the division of labour: performance running is small in units but supplies the belief that lets a $170 lifestyle shoe outsell a $90 one. The run-specialty numbers say that source was shrinking for a year while the lifestyle business boomed, and strong lifestyle demand with a decaying credibility base is not a stable state; it is the first act of Under Armour, whose lifestyle pivot outlived its performance reputation by about two years. On’s answer, the 2027 range led by the rebuilt Cloudsurfer 3, gets its referendum in the same specialty stores in the second half of 2027, and the early signals, demand at six times initial supply for the newest racing shoe, 40 to 50 percent of Cloudmonster sales in higher-priced versions, all come from the seller. Until independent sell-through confirms them, the honest status of the US running problem is: diagnosed, plausibly product-specific, not yet fixed.

I would abandon the thesis if two of the following five occur: markdowns on current-season styles at two or more major US retailers lasting a month or more; Americas growth below 10 percent in the second half of 2027, after the refresh; inventory still growing ten points faster than sales next summer; wholesale revenue growth running persistently below door growth; or gross margin, excluding tariff refunds, below 63 percent for two straight quarters. Any one is noise. Two is the weather thesis winning.

What the price assumes

The strange thing about On’s valuation is that the sceptic is not the analyst community, which rates the stock overwhelmingly a buy, carries a $43.53 average target, and models 2027 revenue almost exactly where the CFO hinted, “we may exceed” $5 billion “somewhere next year.” The sceptic is the price. At $30.16, On trades at 15x times 2027 consensus earnings, in the same neighbourhood as Deckers and Amer Sports on roughly double their growth, and a reverse DCF at a 10 percent discount rate implies growth of about 10 percent a year, a figure that moves between 6 and 14 percent as the rate shifts, so treat it as a range. However you cut it, the market is pricing a company whose rise was mostly weather. My variant view is the one this whole article has been building: the engine is real, it has repeated in a second sport, its gauge never moved even in the storm, and the disagreement with the market is about duration, not next quarter, which management has effectively pre-announced.

The scenarios below are the spine translated into prices, one method throughout: a multiple on fiscal 2030 EBITDA net of stock-based pay, plus accumulated net cash after the buyback, on a flat 335 million shares, at $1.2138 per franc.

The weighted value is about $49, roughly 17 percent a year from here, and the shape of the bear case matters as much as its number: it is not a collapse but an ordinary footwear company, single-digit growth, drifting margins, a multiple to match, which is the outcome the market is already closest to pricing. The calendar of proof is short and specific. November’s third-quarter report should show about 17 percent constant-currency growth, largely pre-announced, with a tariff refund of up to CHF 53 million flattering gross margin that you should strip out. March brings the first annual guide under the new plan, which needs to say high teens. By next August, inventory growth must be converging on sales growth. And through 2027, watch wholesale revenue against door count, and above all watch the specialty-store wall when the new range is fully on it.

Phil Knight solved his cash problem by selling retailers more of the future than anyone could see, and the industry has paid for the wrong guesses ever since, in price, on the rack, every season, a bill that crushed nearly every challenger who tried to scale beneath it. On grew up in a rare decade when the bill collector was distracted, and it used that decade to build something the industry has not seen: a loop of validation, price, margin and reinvestment with a working gauge on the front. August was the first payment demand, and On paid in volume, exactly as designed, at a price the market mistook for failure. If the inventory clears, the new range takes back the running wall, and the doors sell what they stock, then the engine was real all along, and the discount that never appeared will compound longer than any shoe the company ever launches. If not, the weather made the brand, and the bill from 1972 is still out there, accruing interest on 473 million francs of shoes. By the end of 2027, the shelf will tell us which company On has been all along.

Subscribe now

• Cloud is the strongest evidence in the quarter: AI Cloud Infra grew 50%, GPU cloud revenue rose 283%, and Baidu is increasingly proving it can serve large-scale AI demand. But almost all incremental AI revenue came from cloud, not applications or AI-native marketing.

• The original full-stack thesis has weakened: Qianfan can serve ERNIE and competing models, which is good for customers but means Baidu can succeed in infrastructure without its own model, search and applications reinforcing one another economically.

• The missing connection is shareholder economics: search MAUs are falling, application growth is weak, free cash flow remains negative and capex is high. Baidu has shown that AI changes its revenue mix; it has not shown that owning the whole stack makes the combined business more valuable.

Baidu’s potential compounding mechanism is that more AI workloads improve infrastructure utilization and serving efficiency, attracting customers whose usage finances further improvements. Scale strengthens that mechanism only if customers stay and Baidu retains some of the savings after price competition and replacement spending. It breaks if computing capacity becomes interchangeable while search deteriorates and Baidu’s broader AI ambitions consume the profits.

Owning the pieces

Robin Li described the quarter’s strongest business on the earnings call:

AI Cloud Infra delivered another quarter of strong growth, with overall revenue increasing 50% year-over-year, once again outpacing the broader market. Within AI Cloud Infra, GPU cloud revenue nearly quadrupled year-over-year, growing 283% and accelerating significantly from an already strong 184% growth rate last quarter.

I think these results validate Baidu’s engineering more clearly than they validate our original investment thesis. In October’s “The Elephant’s Dance,” we described search intent, models, chips, infrastructure and applications as components of a reinforcing system. By May’s “The Second Threshold,” we had added a qualification: integration had to appear in margins, cash flow and application monetization.

That qualification is now the argument. Of the roughly RMB2.5 billion increase in AI-powered revenue from a year earlier, RMB2.4 billion came from cloud. Applications grew 3%; AI-native marketing was approximately flat. Baidu owns the pieces, but almost all the incremental revenue came from one of them.

The call’s most revealing passage explained Qianfan, Baidu’s platform for serving AI models:

On MaaS, our Qianfan MaaS platform offers one of the most comprehensive model libraries, covering Baidu’s ERNIE family, as well as virtually all of China’s leading models. A key priority for Qianfan is to make model inference at scale more reliable and cost-efficient for customers.

This is a good customer proposition. Enterprises want their chosen models to run reliably and economically. Supporting competing models expands Baidu’s opportunity and reduces dependence on ERNIE winning. But when Baidu sells computing capacity to a customer running somebody else’s model, that transaction does not automatically improve Baidu’s consumer products or bring users back to search.

A valuable business can nevertheless emerge. More workloads can improve utilization; better serving software can lower costs; integration into customers’ operations can encourage retention. This is a different mechanism from the one we originally emphasized. Our variant perception is that Baidu can succeed by serving competing models, but owning its own model and applications does not, by itself, justify a premium valuation.

Alibaba’s 45% external cloud growth in the June quarter also shows that infrastructure demand extends beyond Baidu. Baidu’s advantage must appear in customer economics and returns on capital, not simply participation in a growing market.

The audience is shrinking

Search makes that distinction consequential. Baidu App monthly active users fell from 735 million in June 2025 to 644 million in June 2026, a 12.4% decline. An argument built around capturing consumer intent becomes harder to sustain when the audience supplying that intent is shrinking.

Management attributes advertising pressure both to restrained AI-search monetization and to intensified competition for attention. We previously gave too much weight to deliberate cannibalization. Choosing to show fewer advertisements and losing users to alternatives can happen simultaneously, but they offer different prospects for recovery.

The slowdown in AI-native marketing is particularly uncomfortable. Its growth fell from 262% to approximately zero in three quarters. Comparison bases matter, and this does not establish that AI advertising has reached its ceiling. But it weakens the claim that a new monetization engine is already replacing the old one. Online marketing revenue declined 19% to RMB13.1 billion, improving from Q1’s 22% decline but still deteriorating substantially.

Cloud requires a similarly careful reading. Revenue fell 17% sequentially; management said profits and margins improved year over year. Those comparisons answer different questions. Profitability is reportedly stronger than a year ago, but sequential revenue remains uneven. Without sufficient disclosure of utilization, customer retention and investment payback, we cannot establish how durable the economics are. GPU revenue growth also cannot be treated as evidence of equivalent growth in Kunlunxin’s external chip sales.

AI now represents approximately half of General Business revenue. That confirms its importance, but the percentage rises partly because advertising contracts. The competitive question is whether Baidu is building a stronger customer franchise, not simply a different revenue mix.

The ambition has a carrying cost

Baidu is moving from advertising, which required comparatively little physical capital, toward computing infrastructure, which requires substantial spending before shareholders receive the returns. The transformation therefore changes the balance sheet as well as the income statement.

Q2 operating cash flow was RMB3.4 billion against RMB11.4 billion of capital expenditure, approximately 36% of revenue. First-half free cash flow was negative RMB11.2 billion. Interest-bearing debt increased RMB14.3 billion from December. The RMB1.2 billion quarterly foreign-exchange loss adds earnings volatility, although it is not equivalent to recurring operating cash burn.

Nor should investors spend a dividend before it exists. February’s policy allowed a potential first payment by year-end; it did not declare a distribution. The absence of a payment announcement in these results leaves the timing and amount unresolved.

Meanwhile, Robin Li reaffirmed another claim on capital:

ERNIE has likewise always been an important part of Baidu’s AI strategy and full-stack AI capabilities. We were among the first companies in China to invest in foundation models. There were trials and errors along the way, but our commitment to make ERNIE competitive remains unwavering.

He subsequently described bringing ERNIE “back into the top-tier of foundation models.” Li also described feeding search experience back into model training. That is a plausible improvement loop, and model expertise could help infrastructure efficiency. But the declining audience and weak application growth leave its financial benefits unproven. Shareholders are financing both a platform that serves competing models and an effort to restore Baidu’s own model leadership. The economic connection between those investments still needs demonstrating.

Our May tests remain intact: cloud growth above 50%, consolidated gross margin above 40%, advertising declines better than 15%, stronger application growth and sustainable positive free cash flow in 2H26. This quarter delivered 50%, approximately 39%, a 19% decline, 3% application growth and another cash outflow.

The cash-flow deadline has not passed. However, the spending assumptions below now imply it will be missed. That is a downgrade to our forecast, not permission to move the deadline silently. Baidu’s infrastructure opportunity remains credible; confidence in how quickly it benefits the parent’s shareholders has weakened.

What the balance sheet buys

At HK$89.50, the September 22 close, Baidu sits just below May’s HK$90–105 bear range. That was a conditional future valuation, not a floor. Our former HK$185–205 base range assumed a more successful conversion of AI growth into shareholder earnings than we can presently underwrite.

The revised model separates operating value from financial assets. At June 30, cash, short-term investments and longer-term deposits and held-to-maturity investments, less interest-bearing debt, totaled approximately RMB136.7 billion. That is substantial protection, but it will change with spending, repurchases and operating performance.

These scenarios value the shares in September 2029 using FY2029 operating earnings. They begin with FactSet’s RMB127.4 billion FY2026 revenue estimate. Capital spending and cash flow are explicit assumptions, not management guidance.

Equity value equals after-tax operating earnings times the multiple, plus forecast net financial assets, less a RMB25 billion operating-liquidity reserve and RMB23.3 billion for minority claims. The latter uses current book values as an explicit proxy, not a market valuation. Tax is 25%; the assumed exchange rate is HK$1.15 per RMB. This avoids adding cash to an earnings multiple that already capitalizes its interest income.

Financial assets incorporate the cash-flow paths through September 2029, including three-quarters of 2029’s annual cash flow. Base and bull cases deduct RMB8 billion of cumulative buybacks, assumed to offset issuance and keep diluted ordinary shares at 2.755 billion. The bear assumes no repurchases and 1% annual dilution. No dividends, acquisitions or listing proceeds are assumed. Strategic equity investments are excluded, as is any separate Kunlunxin or robotaxi premium; this is a conservative operating valuation rather than a complete asset appraisal.

The base case requires annual operating cash flow to rise to RMB32 billion by 2029 while capex settles at RMB23 billion, still nearly 15% of revenue. It depends on better utilization, stronger margins and search stabilization, not the disappearance of replacement spending. The bull requires profitable growth across a broader business. The bear assumes continued cash consumption; if that persists indefinitely, even its operating multiple could prove generous.

Investor constituencies will read these results differently. Value investors will focus on financial assets but demand control of cash burn. Growth investors need evidence of durable margins. Event-driven investors can focus on September’s Stock Connect inclusion and Kunlunxin’s proposed listing, but access is not guaranteed buying and a subsidiary valuation is not a distribution. A listed Kunlunxin could also let chip investors bypass the parent. These are interpretations of investor incentives, not verified changes in ownership.

My call is to wait at HK$89.50. The base case offers approximately 12% annual appreciation, against roughly 53% downside in the bear case. I would reconsider around HK$73, where the base case implies approximately 20% annually, provided the operating thesis remains intact. That is a review level, not an automatic purchase or a claim that cash prevents further losses.

The cautious view would weaken if Baidu produced two consecutive quarters of positive free cash flow while sustaining cloud growth above 50%, restoring gross margin above 40% and narrowing advertising declines below 15%. Continued audience losses, flat application revenue and capital spending above operating cash flow would instead strengthen the bear case. Better disclosure of customer retention and investment payback would help distinguish temporary spending from structurally poor returns.

We were right that AI would change Baidu’s business mix. We were too quick to assume that owning the entire stack would make the combined business more valuable. The next increase in conviction should come when growing AI demand leaves more cash for shareholders after financing the rest of the ambition.

Subscribe now

• The recovery is real: Core Local Commerce returned to a 7.9% margin, food and non-food delivery unit economics turned positive, and subsidy discipline improved. But the old margin structure has not returned.

• The fight has moved upstream: Meituan remains excellent at fulfillment, but rivals can attack discovery, local services and customer intent without recreating the entire delivery network. Merchant-services growth of just 4% shows that monetizing this influence remains difficult.

• Staying essential increasingly requires owning more: and sharing more: grocery adds supply control but also capital intensity, while Tencent may become an important source of customer demand. The investment works only if better supply and distribution leave enough economics with Meituan.

In September, we described Meituan as the digital bellhop: valuable because it understood the city and could get things done. We expected its operating advantages to outlast competitors’ subsidies. By March, we had separated the durability of that capability from the margins it could earn; by June, we argued that competition had moved into discovery, grocery supply and the customer relationship.

Q2 passes June’s clearest test. We said Core Local Commerce needed to reach breakeven in Q2 or Q3; it earned RMB5.7 billion. But our earlier confidence that competitors’ financial constraints would settle the contest was excessive. Spending can become more disciplined while the battle for customers continues.

The question governing this update is therefore: can Meituan preserve its influence over local purchasing decisions at a cost that leaves attractive earnings for shareholders? Q2 makes me more confident in the durability of its operating capabilities. The evidence that those capabilities will produce attractive returns after reinvestment remains incomplete.

The recovery has a price

Meituan’s underlying mechanism is straightforward. More orders within the same neighborhoods can improve courier utilization, merchant selection and delivery reliability. Better service encourages repeat purchases, reinforcing local density; the mechanism breaks economically when keeping those customers costs more than the additional density saves.

This quarter showed that Meituan can reduce subsidies and recover earnings. Revenue rose 14.4% to RMB104.6 billion, and adjusted net profit reached RMB2.5 billion against StreetAccount’s RMB736 million expectation. Core Local Commerce supplied 84% of the sequential improvement in group operating profit.

The recovery matters, but the segment includes in-store services, hotels and travel alongside delivery. Management separately disclosed positive food and non-food delivery unit economics. Neither measure establishes that every order is profitable or that the former margin structure has returned.

Regulatory pressure supports greater discipline, but management’s outlook describes a gradual adjustment:

“Even so, we expect UE to stay positive in Q3 as we continue to optimize operational efficiency. Specifically, the industry subsidy level is still much higher than 2024 level, and it will take a few quarters to normalize. At the same time, seasonal headwinds will weigh meaningfully on our UE.”

Summer courier costs, marketing and occupational injury insurance will absorb some gains. Management also expects in-store margins to decline in Q3 and Q4 as investment increases. I expect Q3’s core margin to retreat from 7.9%; sustaining more than 5% would be a reassuring result, rather than company guidance.

The accounting reinforces the need for care. Some incentives reduce reported revenue, so withdrawing them can improve both sales and profit without equivalent transaction growth. Management describes stronger retention, frequency and order mix, but provides insufficient figures to measure how much demand survives without incentives.

The stock’s response is less mysterious when viewed through future earnings. By September 22, the shares were 5.5% below their pre-results close. Between August 27 and September 22, FactSet’s 2027 normalized EPS mean fell 5.8%, leaving the corresponding forward P/E almost unchanged at 16.5 times. The analyst samples changed, and this does not establish trading causality. It does show that the stronger quarter failed to lift the following year’s aggregate earnings expectation.

The lobby is still contested

Our original competitor-exhaustion argument assumed rivals would judge delivery primarily by its own profits. Alibaba’s June-quarter release exposes the weakness in that assumption: quick commerce increased Taobao engagement and contributed incremental customer-management revenue, while delivery unit economics improved without sacrificing market share. My inference is that an order can create value for Alibaba elsewhere in its business, extending its willingness to compete.

Meituan faces a related problem in local services. Completing the transaction and influencing the purchasing decision are different sources of value. A competitor can take valuable restaurant discovery or hotel demand without recreating Meituan’s entire delivery operation. Management described the current attack:

“Over the past few quarters, we’ve seen competitors stepping up investment in local service space through a dedicated shelf-based app. They have been subsidizing heavily to redirect the traffic from their content-driven model to accelerate the adoption of the new app.”

Core Local Commerce merchant-services revenue grew only 4.0%. That is a warning, but it combines commissions and advertising across activities. Without comparable transaction growth and category mix, it cannot establish that Meituan’s pricing power has deteriorated. It does establish that merchant monetization is growing much more slowly than the headline business.

Meituan’s response is to improve supply, merchant tools and customer experience while spending more in-store. These investments could deepen relationships; they could also become permanent costs of defending existing ones. The distinction will appear in retained customers and contribution profit after incentives, rather than in management’s description of its expanding capabilities.

June’s argument therefore survives, with a harder economic test. Meituan still coordinates something difficult and useful. Shareholders benefit when that coordination gives customers and merchants reasons to stay without requiring a matching increase in defensive spending.

Buying the pantry, sharing the front door

The most revealing number this quarter concerns what Meituan increasingly owns. Product sales rose 48.8% to RMB26.7 billion and supplied 66.5% of incremental revenue. Excluding product sales, revenue grew 6.0%. Selling goods adds their full selling price to revenue; arranging a transaction generally adds only the associated fees.

This is why grocery belongs at the center of the investment argument. Customers want reliable availability, quality and delivery. Where independent merchants cannot consistently provide them, owning supply can strengthen Meituan’s service. But that ownership brings inventory, operating complexity and capital requirements that a revenue-growth comparison misses.

Xiaoxiang reached 68 cities. Grocery margins improved, yet absolute losses increased sequentially as expansion continued. Total New Initiatives losses narrowed because improvement elsewhere, including Keeta, offset that spending. Our June test must become more precise: grocery should earn its own justification rather than receive credit for overseas progress inside the same segment.

Keeta provides encouraging evidence that operating knowledge travels. Hong Kong is sustainably profitable, and management says Saudi Arabia reached profitability in July. The disclosed milestones use different definitions, so they do not establish a precise acceleration in time to breakeven. Brazil must still demonstrate its own economics.

AI introduces a different ownership question: who controls the customer’s request? Meituan is developing its own assistants and merchant tools, while pursuing an external route to demand. Management said:

“Our partnership with Tencent is progressing rapidly, with both teams refining the product to deliver faster, more convenient services.”

The opportunity is substantial. A request originating within Tencent’s services could reach Meituan without Meituan paying to acquire that customer directly. But a better distribution channel can also gain negotiating power over the businesses fulfilling its requests. The announcement establishes cooperation; it does not establish default status inside WeChat or disclose the commercial terms.

Meituan’s defense is the difficult work behind the recommendation: accurate availability, trusted merchants and reliable execution. Those capabilities can remain valuable when another company originates the order. Their value weakens if competing providers become interchangeable and the outside assistant controls customer choice.

The pantry and the front door belong in the same argument. Meituan is investing to make its service harder to replace while accepting that it may not control every customer interaction. The return depends on whether differentiated supply and execution outweigh the cost of building them.

Three prices for staying essential

My variant view is that Meituan can deliver attractive returns without recovering its former margins. A larger business earning moderate operating profits, supported by valuable financial assets, can justify more than today’s price. The risk is that maintaining its position consumes both the recovered earnings and part of those assets.

At June-end, net cash before leases was RMB79.6 billion, alongside RMB77.3 billion of investments: roughly HK$30 per current share combined. After lease liabilities, net liquidity was RMB72.5 billion. Those assets matter, but their value to shareholders depends on what management does with them. On potential disposals, management said:

“Going forward, we will weigh market conditions, valuation, funding needs, and our broader capital allocation priorities. When time is right, we are very open to exit or monetizing selected positions to free up capital. This will give us greater flexibility to reinvest in our own business and return value to shareholders.”

The choice between reinvestment and shareholder returns remains open. Q2 operating cash flow recovered to RMB9.7 billion, but the first-half total was only RMB2.7 billion, down 81.8%. Without corresponding capital expenditure, that cannot be presented as verified free cash flow. Asset sales would improve liquidity without proving better operating economics.

The revised valuation separates operating earnings from financial assets. It charges stock compensation and depreciation in operating margins, assumes 20% tax, and excludes financing and investment income from the earnings being capitalized. Cash and investments are then added separately, avoiding double counting.

These are my assumptions, using 6.35 billion diluted shares and HK$1.1703 per RMB. I reserve RMB30 billion of current net liquidity for operations; bear, base and bull cases then credit cash changes of minus RMB20 billion, zero and plus RMB20 billion through 2029. Investment haircuts are 50%, 25% and zero, reflecting realization and tax uncertainty. The base gives no credit for additional cash accumulation. Revenue growth starts from FactSet’s RMB415.9 billion 2026 estimate.

The bear assumes persistent competitive spending and grocery investment consume the benefits of scale. The base requires profitable delivery, resilient in-store earnings and manageable expansion costs. The bull requires grocery, merchant tools and wider distribution to improve customer economics sufficiently to support higher group margins.

At the September 22 close of HK$73.20, the base implies roughly 17% annualized price appreciation over three years; the bear implies approximately 32% downside. The HK$117 base remains within June’s HK$105–135 range. I remain moderately bullish, with a more attractive entry around HK$65–68, rather than assuming a strong quarter has removed execution risk.

The tests are specific. Negative food-delivery unit economics in Q3 would contradict management’s outlook. A core margin below 2% would be a warning; two consecutive quarters below that level alongside renewed subsidy spending would challenge the recovery thesis. Negative H2 free cash flow after disclosed capex would weaken the capital-discipline case. Grocery losses rising without evidence of improving mature-location returns would challenge the case for owning more supply.

The bellhop has demonstrated that he still earns his place in the building. What remains unresolved is how much he must spend on the pantry, the reservation desk and access to the guests. Q2 strengthens my belief that Meituan can remain essential. The investment succeeds when doing so leaves more cash for its owners.

Subscribe now

• Alibaba finally cleared an important cloud-economics test: AI Cloud and Compute revenue grew 45%, adjusted EBITA grew 133%, and margin reached 11.6%, above the 11% threshold set earlier.

• But the cash burden worsened materially: capex reached RMB67.7 billion, free cash flow was negative RMB44.7 billion, and Alibaba subsequently raised HK$80 billion of new equity for AI investment.

• The thesis is now narrower: Alibaba must prove not just that cloud demand is real, but that profits eventually grow faster than the capital required to sustain them. Owning Qwen, applications and infrastructure only matters if enough of the economics ultimately accrue to shareholders.

On August 23, Alibaba announced the pricing of HK$80bn of new equity: 710 million shares at HK$112.70, with the proceeds earmarked for AI. By September 21, the Hong Kong shares had closed at HK$112.60, almost exactly the placement price. That coincidence is not a valuation model, but it captures the question confronting shareholders: how much should they pay today for capacity whose eventual returns remain uncertain?

Our first article in this series appeared on September 23, 2025, when 9988.HK closed at HK$159.40. Bloomberg’s price history shows a decline of 29.4%, excluding dividends, through September 21. Our three-year forecasts have not reached their horizon, but the experience so far has been poor. We identified Alibaba’s strategic direction more accurately than we estimated the cost of travelling in it.

Our January article, The Second Trust Problem, concerned customers trusting AI to execute transactions. The question now extends to shareholders. Alipay’s original escrow arrangement withheld payment to the seller until delivery was verified; Alibaba’s AI transformation requires capital before the financial delivery can be verified.

This quarter supplied meaningful verification. It also increased the amount at risk.

The test we cannot move

In March, we wrote:

“Cloud EBITA margin crossing above 11% in any quarter would indicate that operating leverage is beginning to emerge.”

The June quarter crossed that threshold. AI Cloud and Compute generated RMB48.4bn of revenue and RMB5.6bn of adjusted EBITA, an 11.6% margin. Revenue grew 45%; profit grew 133%. The segment now includes T-Head, so our original standalone-cloud test is not perfectly unchanged, but Alibaba has recast the prior-year comparison. The improvement is not simply an artifact of adding a business to one side of the calculation.

That deserves an upgrade to our view of cloud economics. It does not justify declaring the entire transformation successful.

Our earlier expectations about funding have aged considerably worse. November’s Peak Investment argument assumed financial pressure would force spending to normalize. January placed too much faith in a rapid improvement in commerce economics. By May, we had retreated to The Bills Changed Shape: AI demand was real, but neither lower spending nor stronger shareholder cash generation had followed.

Customer-management revenue supplies an uncomfortable audit of another condition. We said a year-over-year decline would change our view; reported CMR fell 7%. The reported trigger has been crossed. Moving subsidies from marketing expense to contra-revenue complicates comparisons, but management’s comparable figure is only 1% growth—still far below the 8% threshold we once regarded as validation.

We should neither ignore the reported breach nor mistake a presentation change for proof of equivalent economic deterioration. Either way, the expectation of a strong, expanding commerce engine comfortably financing AI is no longer our base assumption.

The revised conclusion is narrower: cloud is beginning to earn its growth, while commerce is providing less support than we expected. We were not merely early on the cash recovery; important assumptions behind its timing were wrong.

What compounds

Alibaba’s opportunity is not simply that customers need more accelerators. It is that AI can make computing part of producing work rather than merely supporting the software around it.

An agent completing a task may generate tokens, retrieve records, execute code, maintain state and call other services repeatedly. Management says agent adoption is increasing demand for CPUs, storage, databases and networking alongside GPUs. This creates an opportunity for recurring, usage-linked revenue across several products—provided customers continue finding the work useful.

That broadens Alibaba’s opportunity to capture each customer’s activity. A customer renting an accelerator can compare prices. A customer running a process across databases, execution tools and model services has more reasons to stay, provided those services work well together.

T-Head could strengthen the economics underneath this relationship. Alibaba reports more than 650 external customers using Zhenwu chips through its cloud services. Management argues that proprietary silicon can reduce dependence on merchant accelerators and improve margins through tighter integration. Customer adoption is documented; the size of the fully loaded cost advantage is not.

The potential mechanism is straightforward: more paying workloads support better utilization and system optimization, which improve price-performance and attract more workloads. This is primarily a scale-and-cost mechanism, not an automatic network effect.

The quarter makes it more credible:

Source: company disclosures; calculations ours. The sum is not a standalone consolidated AI income statement: it excludes unallocated and intersegment effects and benefits elsewhere in commerce.

Cloud earned approximately a 21% incremental EBITA margin on its additional revenue. But its RMB3.2bn profit improvement was outweighed by a RMB10.6bn increase in model-and-application losses. A profitable computing operation is not yet evidence of profitable ownership of the entire AI system.

There is a second test ahead. EBITA already includes depreciation, making the margin improvement meaningful. As additional equipment becomes operational, however, its depreciation enters expenses. We cannot assign the entire capex wave to an invented future start date or assume today’s margin is untouched by it.

The question is whether utilization, pricing, mix and T-Head savings can continue improving economics as the installed asset base grows. That is how an encouraging quarter becomes a durable competitive advantage.

The deposit got larger

Group capex reached RMB67.7bn, against RMB22.9bn of operating cash flow, leaving free cash flow negative RMB44.7bn. The subsequent equity placement made the financing question explicit.

Issuing shares does not automatically destroy value: the cash received can fund investments worth more than their cost. But it changes the bargain. Existing shareholders own a smaller percentage of a better-funded company, and the return on that enlarged capital base—not the popularity of its models—determines whether they benefit.

Repurchases were only US$162m in the quarter, against approximately US$0.8bn a year earlier. Those are company-reported transaction amounts; our stock prices and valuations remain in HKD. Meaningful capital return has given way to construction.

The most revealing post-earnings change is in estimates:

Source: Bloomberg point-in-time estimates for the year ending March 2027, using the same RMB-reporting BABA-US panel. Contributor counts changed: EBITA 27→24, capex 30→26, FCF 14→13. These are consensus-snapshot changes, not matched-analyst revisions.

Analysts came away expecting more operating profit and substantially less cash. That is not confusion; it is recognition that the better business requires a bigger investment.

Management has offered a concrete defence. AI assets are said to pay back in roughly three years, potentially shortening toward 2–2.5 years. On procurement timing, the call was explicit:

“I don’t think we should take the spending for this quarter and multiply it by four.”

That is reasonable. RMB190bn spent against a RMB380bn three-year program does not, by itself, prove the budget has failed. Nor does a shorter hardware payback establish an attractive return after recurring research, replacement equipment, financing and the time value of money.

Our November article began with Alibaba’s failed New Retail investments precisely because coherent architecture is insufficient evidence of good capital allocation. Today’s external revenue and operating margin offer more tangible verification. They do not exempt this investment cycle from the same test.

The relevant question is whether profits eventually grow faster than the capital required to sustain them—not whether Alibaba can keep finding projects on which to spend.

Owning intelligence is not owning the return

One claim from our earlier work needs correction rather than refinement.

We suggested that developers standardizing on Qwen were implicitly standardizing on Alibaba Cloud. AWS has offered Qwen through Bedrock since September 2025. A customer can use Alibaba’s intelligence while paying Amazon to execute it. The assumption was too strong when we made it.

Management’s answer about model-as-a-service economics narrows the thesis further:

“The level of gross margin from those two kinds of models is actually very comparable.”

That refers to hosting proprietary versus third-party models. It supports Alibaba’s ability to monetize a multi-model market; it does not prove models have no value. Comparable serving margins also leave unanswered whether developing Qwen earns an adequate return on its research expenditure.

Alibaba must win workloads through better execution economics, useful integrations or distribution not merely through the availability of its weights.

This is where Second Trust still matters. Qwen can connect recommendations to merchants, fulfilment and after-sales processes. Those relationships could make an agent more useful and transactions more profitable. But 250 million users having their first AI-shopping experience demonstrates reach, not repeat usage or incremental commerce profit. Our March requests for repeat rates, incremental GMV and order economics remain unanswered.

The investment case has two distinct accomplishments to prove: Alibaba must capture a growing share of useful work, and the combined cost of models, applications and capacity must leave sufficient profit for shareholders.

My disagreement with the sceptical reading is not that cash burn is harmless. It is that the 21% incremental cloud margin creates a plausible route to better returns even if growth eventually moderates. A 30–35% growth business with improving cash conversion could be more valuable than a 45% growth business requiring repeated capital injections.

That would also change who can own the stock. A value investor awaiting distributions and a growth investor underwriting productive capacity can interpret the same quarter differently. A durable re-rating requires evidence that satisfies both: profitable expansion followed by cash generation. We should not infer that shareholder transition merely from a rally.

Three futures, with the arithmetic exposed

We retain Kingdom, Grid and Furnace, with conditional operating-case probabilities of 25%, 45% and 30%. Better cloud economics strengthen the upside; greater funding requirements strengthen the downside.

One further correction is necessary: our previous Furnace price band was too forgiving. A downside case must allow weak cash conversion and further dilution to damage value substantially—not assume support near HK$100.

The valuation date is September 2029, using FY30 earnings. All assumptions below are ours, not company guidance.

Revenue starts from Bloomberg’s September 22 FY27 median of RMB1.124tn. Price equals revenue × normalized net margin ÷ diluted shares ÷ assumed RMB0.87 per HK$1 × P/E. Net margins include recurring model/application costs and share-based compensation but exclude investment revaluation gains. Share counts incorporate financing and assumed net repurchases; buybacks are not added again to returns. Dividends are excluded.

The Furnace can grow and still disappoint because continued capital consumption deserves a low multiple. The Grid earns 16 times earnings only if cash conversion becomes credible. The Kingdom requires differentiated returns across cloud and commerce, not simply larger chip shipments.

These are business scenarios, not downside limits. Severe geopolitical disruption could produce losses beyond the Furnace case; the probabilities are conditional on its absence.

What the next evidence must show

We will judge the next chapters against five tests rather than another model leaderboard.

Thresholds are our monitoring judgments, except management’s RMB30bn year-end ARR target. No isolated result establishes or falsifies the whole thesis.

Alipay reduced the distance between a promise and verification by withholding payment until delivery. Equity investment cannot offer that protection: capital must come first. What shareholders can demand is that each round of investment produces more evidence, rather than merely a larger ambition.

Alibaba has finally delivered part of the verification we were waiting for. It has simultaneously asked shareholders to enlarge the deposit. The investment works only when the value of the delivery grows faster than the deposit required to secure it.

Subscribe now

• The core is still funding the experiment: operating profit excluding new AI products grew 19%, while the drag from new AI products increased to RMB10.5 billion. The established businesses remain strong enough to finance the build, but that does not make every yuan of AI spend value-creating.

• Weixin’s moat may be execution, not frontier intelligence: management explicitly said Xiaowei does not require Hunyuan to be state of the art. Identity, permissions, Mini Programs and payments may matter more than owning the best model.

• The investment case now turns on incremental economics: WorkBuddy is beginning to generate real receipts, but Tencent still needs to show that agents create new profitable activity rather than simply rerouting existing transactions while adding compute cost.

We started writing about Tencent around HK$662. By our June article, the reference price was roughly HK$420; Tencent’s September 21 close was HK$430, according to FactSet. A decline of approximately 35% from our first reference price, before dividends, deserves more than another explanation of why the business is interesting.

Our operating thesis has held up better than our investment conclusion. Tencent continues to monetise its existing distribution, while games and advertising remain strong. But we underestimated capital intensity and overstated the protection provided by apparently inexpensive earnings. The original three-year scenarios have not matured; the suggestion that downside was tightly bounded has nevertheless proved inadequate for the journey.

In June, we wrote:

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

That remains the right starting point. Q2 strengthens the case that Tencent can create an agent economy inside Weixin. It also reveals why believing in that opportunity does not require accepting every investment made in its name. The question is now whether Tencent can turn more delegated tasks into enough incremental profit to justify the capital committed.

The Core Is Paying the Bill

Tencent reported RMB204.8 billion of revenue, up 11%, and RMB75.6 billion of non-IFRS operating profit, up 9%. Excluding new AI products, operating profit reached RMB86.1 billion, up 19%. This was not a newly invented Q2 disclosure: our June article already used the same separation to describe two companies inside one filing.

Almost the entire RMB1.7 billion sequential improvement in profit excluding new AI products was absorbed by the increase in AI losses. The distinction is useful, but it is not absolution. Adding losses back demonstrates the strength of the funding business; it does not establish that those losses are discretionary or that financing them creates value.

The established businesses remain encouraging. Domestic games grew 17%, Marketing Services 22%, and VAS gross margin reached 64%, versus 60% a year earlier. But the old AI-dividend argument needs precision: gross-profit growth exceeded revenue growth by about two percentage points, not the seven we previously highlighted. Management also said advertising-funded games contributed approximately two percentage points to Marketing Services growth. Strong advertising does not mean every point of growth came from better AI targeting.

The cash picture requires equal care. Total capex reached approximately RMB52.8 billion, while reported free cash flow was negative RMB13.8 billion. Excluding compute-procurement prepayments, however, management said FCF would have been positive RMB37.6 billion. The RMB88.7 billion sequential decline in net cash also included RMB41.6 billion of dividend payments; it was not simply an AI operating loss.

FCF figures matter. The adjusted number demonstrates substantial cash generation before those procurement prepayments. The reported number records cash shareholders have committed. Removing the prepayments helps explain the quarter, but it does not make the investment free or establish that similar commitments will not recur.

The Model Is Not the Tollbooth

Our earlier articles established why agents might make Weixin more valuable. A model supplies intelligence; a useful agent also needs identity, permissions, service access and settlement. Mini Programs potentially supply the services, Weixin the trusted relationship, and Weixin Pay the transaction capability. Q2 adds Xiaowei, a Weixin-native prototype designed to connect these pieces, with user confirmations still required.

The most consequential clarification came when JPMorgan asked why Tencent needed to spend toward frontier-model capability. Martin Lau answered:

“Weixin’s agent doesn’t really require or depend on Hunyuan’s … SOTA status. Weixin’s design, as we have said a few times, is actually centered around user privacy and focusing on solving all the necessary interactions an agent needs within the Weixin environment and also for cost efficiency.”

State-of-the-art intelligence and Weixin execution are therefore distinct propositions. Hunyuan’s frontier ambitions support more demanding WorkBuddy tasks and a broader token business. Xiaowei is designed around the intelligence necessary to operate effectively within a particular environment. Better models can help both, but winning the frontier race is not management’s stated prerequisite for Weixin’s agent to work.

This changes the capital-allocation argument. Xiaowei could extend a scarce asset Tencent already controls; frontier Hunyuan also competes for a new profit pool. The first opportunity cannot automatically justify the second budget. Weixin’s agent thesis could succeed while the next increment of frontier-model spending still earns an inadequate return.

The potential compounding mechanism remains attractive. Successful delegation encourages users to entrust more tasks to Xiaowei. That activity gives merchants an incentive to make their services agent-accessible. More accessible services make the agent more useful, attracting further delegation. Tencent need not charge a new fee on every task: additional advertising, payments and merchant activity could monetise the expansion through existing businesses. That is the economic possibility behind management’s proposed agent-to-agent transactions.

But the relevant word is “additional.” UBS asked whether agents might merely reroute existing transactions, adding compute costs while removing advertising impressions. Lau answered with a vision of a more useful Weixin, not a quantified economic bridge. A better customer experience can expand a market, but it can also eliminate profitable friction. The quarter establishes neither outcome.

Where the First Receipts May Appear

WorkBuddy offers a different kind of evidence. Management says paying users already generate gross margins comparable with Tencent Cloud, although subsidising free users lowers the product’s overall margin. Tencent also redirected spending toward WorkBuddy after identifying stronger traction, rather than increasing every AI project’s budget indiscriminately.

James Mitchell supplied the most useful commercial disclosure:

“And so similar to games and some of our other businesses, there’s a lengthy time lag between the cash receipts coming to us from the users and those cash receipts translating into reported revenue. But we are seeing a substantial ramp in the cash receipts today, and that will translate into reported revenue growth for Tencent Cloud as we move through the year.”

This is progress beyond retention claims. Customers are paying, and management describes a revenue-recognition lag. But it disclosed neither the magnitude of receipts nor paid cohorts, while paying-user gross margins do not establish profitability after acquiring users and developing models. WorkBuddy has evidence of monetisation, not yet evidence of satisfactory returns on all the capital supporting it.

It also qualifies something our Q4 article overstated. Difficulty selling general-purpose consumer chatbot subscriptions does not mean subscriptions are unavailable throughout Chinese AI. WorkBuddy’s productivity use cases apparently support payment. Tencent’s opportunity is broader than collecting transaction fees, but its different businesses need different tests.

That is why I would not collapse the investment case into “Tencent is China’s Meta.” The useful distinction is between raising yield on existing activity and creating genuinely new paid activity. Nor should WorkBuddy’s SkillHub be presented as proof that Weixin’s Mini Programs have become agent-callable at scale. They are related ambitions, not interchangeable evidence.

The Price of Owning the Future

The analysts on the call were asking economic questions. Bernstein wanted depreciation and payback explained; UBS challenged incremental demand; Bank of America asked whether compute would displace repurchases. This does not establish what every investor believes, but it hardly supports the claim that the market failed to notice Tencent’s 19% ex-new-AI profit growth.

Management’s defence is that compute has alternative uses. Tencent could rent capacity through Cloud rather than devote it to models and proprietary applications. Lau also said some recent orders and prepayments could then have been resold at more than a 30% profit. That is evidence of current scarcity, not a realised gain across the compute estate or a permanent valuation floor.

A profitable fallback reduces the risk of stranded capacity; it does not clear the capital-allocation hurdle. The comparison is not a positive return versus zero. It is risk-adjusted returns from AI versus repurchases and other uses of capital. Management acknowledged the trade-off by saying superior compute opportunities could mean less cash for buybacks. Calling the initial build a lump sum does not identify its peak or eliminate future replacement spending.

Our variant perception is therefore narrower, and stronger, than “the market does not understand.” Weixin’s execution advantage may require less dependence on frontier-model leadership than a generic AI narrative implies. But that makes it more important to evaluate Tencent’s expensive model ambitions separately. The scarce asset may be the environment in which intelligence acts, without every attempt to manufacture more intelligence becoming equally valuable.

Three Futures, Not a Promised Recovery

I retain June’s FY29 scenarios rather than raise the targets because the prototype is more tangible. Product evidence improved; the capital burden increased. These are conditional investment assumptions, not company guidance or consensus forecasts.

Source: June’s FY29 scenario framework, retained rather than newly recalculated. Operating margins include new-AI investment; multiples apply to FY29 earnings. Probabilities are subjective and dividends are excluded.

The bull case requires WorkBuddy to become a meaningful profit pool and Xiaowei to expand economic activity, with capital intensity easing as revenue catches up. The base case allows strong games and advertising to finance slower AI progress, followed by cash-flow recovery. The bear case combines persistent investment with weak monetisation or an outside agent taking control of user intent. Its price range is not a guaranteed floor.

These outcomes require earnings as well as multiples. At midpoint multiples, the midpoint prices imply approximately HK$51, HK$47 and HK$37 of FY29 EPS equivalent, respectively. Strong core growth alone does not deliver those outcomes if depreciation, dilution or continuing AI losses absorb the gains.

Our dashboard must not become easier as the spending rises. June’s ex-new-AI operating-profit growth test remains above 15%, with below 10% signalling deterioration. The 25–30% second-half growth ambition applied to Business Services; low-20s Cloud growth in Q2 is encouraging but is neither the same measure nor the same period. June’s reported quarterly FCF bar remains RMB50 billion, with repeated quarters below RMB40 billion a warning. Ex-prepayment FCF is a separate diagnostic, not a substitute pass.

For the new business, I would treat Q3 AI drag reaching RMB14 billion without quantified commercial progress as an escalation trigger, not proof that an undisclosed management budget has been breached. WorkBuddy needs paid-user, revenue or cohort disclosure; Xiaowei needs successful tasks, repeat delegation and incremental transactions. Those observations would test whether usage is becoming durable economics.

We were right to identify valuable distribution and its potential transformation. We were wrong to treat that insight as sufficient protection against capital-allocation and valuation risk. Q2 strengthens the strategic case, but the appropriate response is not another declaration that vindication is inevitable.

Tencent may own an unusually valuable place for agents to act. The next question is how much shareholders should pay to make those agents intelligent, attract their users and complete their tasks. The tollbooth thesis will earn its valuation when the resulting cash exceeds that cost, not merely when the architecture becomes impossible to ignore.

Subscribe now

• The stock no longer needs a heroic AI thesis: Adobe has grown revenue, earnings and free cash flow over two years while the share price has collapsed nearly 59%. At today’s valuation, stabilization alone can create upside.

• The real risk is managed decline: freemium can be either investment or concession, margins are lower, and Adobe still has not disclosed the conversion, retention and organic ARR evidence needed to tell the difference.

• The investment question has simplified: Adobe no longer must prove AI makes its moat stronger. It must show that existing cash earnings are durable enough for investors to stop shortening the useful life they assign to them.

Adobe’s quarterly revenue has grown 25% in two years, adjusted earnings per share approximately 32%, and free cash flow approximately 24%. Around the corresponding earnings releases, its share price has fallen almost 59%. The company has become larger while investors have become much less willing to pay for the earnings it produces.

Our earlier work tried to resolve that contradiction by asking whether AI would make Adobe more valuable as content became abundant. That remains an important business question. At today’s valuation, however, the stock can work before Adobe establishes everything we once believed about that opportunity. The more immediate question is whether the deterioration investors fear will continue.

There are two investment paths. Adobe’s earning power proves durable enough for pessimism to bottom or reverse, and today’s price is low. Alternatively, new products and repurchases sustain the reported numbers while customer economics continue deteriorating, and Adobe enters managed decline. A temporary share-price bounce would not establish the first; an extended period of poor returns would not, by itself, establish the second.

The structural mechanism remains straightforward. Adobe becomes more valuable as customers depend on its assets, brand rules, permissions and performance history to complete growing volumes of commercial work. That advantage compounds when each project leaves context that improves subsequent projects and retains more spending within Adobe. It breaks when competing tools can use that context equally well or additional consumption fails to compensate for weaker pricing and fewer paid seats.

Q3 delivered $6.76 billion in revenue, up 12.9%, and adjusted EPS of $6.13, up 15.4%, with higher full-year guidance. Those results leave one fundamental question: is Adobe’s earning power durable enough for pessimism to stop worsening, or are improving headline numbers concealing managed decline?

When the stock stops fighting its earnings

In June’s The Abundance Paradox, we acknowledged what our earlier optimism had missed:

“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 correction needs to go further than saying we were right about the fundamentals and wrong about the stock. Pricing power, retention and acquisition costs are fundamentals. We identified a real opportunity but treated greater usage as stronger evidence of economic power than it was.

The same discipline now requires separating a weaker competitive advantage from an inevitably declining business. Adobe could retain substantial earning power without regaining its old dominance. That distinction matters much more near ten times adjusted earnings than it did at a premium valuation.

The record shows a growing business under profitability pressure. Over two years, adjusted net income increased approximately 16.5%, with the lower share count contributing materially to faster EPS growth. Cash generation also increased, so the improvement cannot be dismissed as repurchases alone.

Meanwhile, the shares moved from $586.55 at the September 12, 2024 regular close to $243.08 in September 10, 2026 after-hours trading. Those are different session windows, but the valuation compression is unmistakable. The successful investment path begins when investors stop shortening the useful life they assign to Adobe’s earnings.

That need not require accelerating growth. If the earnings multiple stops falling, modest EPS growth can again translate into share-price appreciation. If confidence also recovers, the multiple can contribute rather than subtract. A valuation floor removes a headwind; the return must still come from earnings growth, actual cash distributions or some recovery in what investors will pay.

The arithmetic illustrates the opportunity. At the FY26 adjusted EPS midpoint of $24.475, a valuation of twelve to fifteen times earnings implies approximately $294–367, versus the dated $243.08 reference. That is 21–51% upside without assuming any earnings growth. It is entirely a rerating sensitivity, not a promised return or a timetable.

When adaptation becomes managed decline

The other path is more difficult to recognize because the company can remain profitable for years. Existing contracts, new products, acquisitions and a declining share count can support reported growth while the economics of attracting and retaining customers weaken.

Freemium sits at the center of that possibility. Asked about postponed pricing, Shantanu Narayen said:

“I’m actually really happy that we didn’t focus on the pricing actions, because that, while it may have provided some short-term relief, would not be as critical as continuing to drive new user adoption.”

Free access can recruit customers who become more valuable over time. It can also become necessary because customers have cheaper alternatives. Both forces can operate together; choosing the right defensive response does not establish that Adobe can preserve its previous profitability.

Managed decline would mean the concessions keep expanding. More usage requires more computing and product investment, but customers resist paying proportionately more. Enterprise growth remains positive without offsetting weaker creative economics. Repurchases support EPS while aggregate profit or economically available cash begins to erode.

That is why the eight-quarter margin record matters. Q3’s approximately 44% adjusted operating margin compares with 46.3% a year earlier. The materials do not isolate the contribution from AI, acquisitions and other spending, so we cannot attribute the entire decline to competitive pressure. We can say that sustaining growth currently absorbs more of each revenue dollar.

A low multiple offers less protection if the earnings being multiplied are themselves too high to sustain. Managed decline should therefore be judged through customer spending, retention and cash generation, not inferred from the share price. Equally, a rally driven by sentiment alone would not demonstrate that those economics had stabilized.

What Q3 tells us about the choice

There is evidence supporting the successful path. AI-first ARR exceeds $650 million, growing more than 150% year over year. Firefly app and credit-pack ARR grew 40% sequentially, and Brand Visibility doubled paying customers. These demonstrate payment as well as adoption, even though their undisclosed bases limit the consolidated conclusion.

The enterprise opportunity also survives. Generating an image becomes easier, but companies still need the correct product, approved claims, rights, branding and evidence of commercial results. More content increases the work surrounding generation. Adobe can benefit if its customer knowledge makes that work more dependable.

Competition remains relevant: Canva offers brand controls and approvals, while Figma connects AI creation with existing design context. Adobe must establish that its accumulated customer relationships deliver enough additional value to retain spending. Greater usefulness does not automatically confer greater pricing power.

The enterprise disclosures are less helpful for identifying momentum. Q2 reported GenStudio ARR growth above 25% and AEP subscription-revenue growth above 30%. Q3 reported above-20% ARR growth for each of three product families. Different metrics and disclosed thresholds prevent a clean deceleration calculation, but the presentation makes progress harder to assess

Anil Chakravarthy explained how management intends to convert engagement into revenue:

“The goal is, of course, to prove value first, and once they get value, we get a lot of signals about the engagement, about the intensity of AI usage, about whether they’re ready to move to, for us to convert them into ARR. We’re constantly calibrating that. What is the best conversion point? What kind of paywall tuning, for example, would we do?”

The missing evidence is what that calibration earns. Adobe has not disclosed comparable conversion and retention cohorts or acquisition payback. Asked for Semrush’s contribution, management emphasized integrated selling; total ARR rose approximately $400 million sequentially, but the organic bridge requires Semrush’s change during the quarter.

Nor can we manufacture the missing answer from AI-first ARR. Above $500 million becoming above $650 million does not establish at least $150 million of additions: $625 million becoming $675 million would satisfy both disclosures while adding only $50 million. The remaining business’s contribution cannot be reliably inferred by subtracting disclosed minimums.

June anticipated transition pressure and benefits during FY27. Q3 continues that bet, with paid-adoption progress and incomplete economic evidence. Better disclosure would help us distinguish the paths; it would not itself make either outcome true.

The evidence that changes the bet

Our June article established six signposts, and they remain useful evidence for this decision. Their job is to tell us whether customer economics are finding a floor.

Anil was named September 3, after Q3 ended August 28. That resolves the leadership vacancy but narrowly misses our literal timing condition. The commercially meaningful question is whether his management makes customer economics more durable and more visible

My lean is toward stabilization rather than managed decline. Revenue, aggregate adjusted profit and cash generation have grown, and customers are paying for new AI products. Those facts do not prove a bottom, but they give me more reason to investigate a valuation recovery than to treat economic decline as established.

That is a different conviction from our September 2025 argument. I no longer need Adobe to prove that AI creates a stronger moat before the stock can work. I need evidence that its existing earning power can endure, allowing the market to stop discounting each year as another step toward erosion.

Because prices can turn before reported growth does, waiting for every signpost to pass may mean paying substantially more. The investment judgment is whether the evidence already favors stabilization strongly enough to accept what remains unknown.

The valuation must still recognize stock compensation and necessary reinvestment. Adobe trades at approximately 13.4 times FY26 GAAP EPS guidance, versus 9.9 times adjusted guidance at the reference price. Repurchases can improve each remaining shareholder’s claim, but their cash cannot also be counted as an additional yield once the EPS benefit is included.

By Q1 FY27, I want a comparable organic bridge or convincing alternative evidence of conversion and retention economics. If neither arrives, I would stop assuming the freemium investment is working and re-underwrite observable earnings and cash. Persistent deterioration in those economics would move my view toward managed decline; silence alone would reduce confidence without proving it.

Over the next three years, the decisive issue is whether pessimism has already outrun the damage. If earnings prove durable and distrust stops deepening, Adobe is cheap enough for the stock to work without recovering its former dominance. If maintaining the business requires continuing concessions and shrinking economic returns, the low multiple is anticipating managed decline. That is the decision this quarter leaves us to make.

Subscribe now

• The funding model has changed materially: customer prepayments with a significant financing component rose to $11.4 billion in Q1, helping fund a massive infrastructure build without making Oracle economically capital-light.

• The operating evidence is increasingly convincing: OCI growth reached 121%, utilization was 97.9%, and even four-year-old GPUs renewed or resold at a 20% premium, arguing against the simplest “dark fiber” comparison.

• But the shareholder question remains unresolved: revenue is accelerating faster than per-share earnings, interest expense and dilution remain meaningful, and Oracle still has to prove that pre-funded growth converts into attractive full-cycle cash returns.

The most revealing exchange on Oracle’s earnings call began with a misunderstanding. An analyst asked about contracts requiring no additional capital expenditure. Co-CEO Clay Magouyrk corrected him:

“Sure. Let me make sure we clarify one thing first, and then I’ll dive into the pieces. I didn’t say, and I don’t think myself nor Hilary said that it doesn’t require additional CapEx. We said it doesn’t require additional cash from Oracle. Right?”

That distinction explains why I am more constructive after Q1, and why the qualification matters. Oracle reported revenue of $19.35 billion, up 30%, and adjusted EPS of $1.92, up 30%. More consequentially, it demonstrated that customers will help finance an increasingly large service obligation.

AI makes Oracle’s ability to turn processors, power and networking into reliable computing more valuable. Repeated delivery can improve execution, build customer confidence and attract further funded commitments, allowing operating expertise to compound as demand grows. That mechanism breaks if customers mainly value temporary scarcity, or if their funding and Oracle’s contractual returns cannot sustain expansion.

Our previous article, The Price of the Moat, accepted the financing mechanism but withheld a verdict on the whole business. I am now moderately bullish: delivery has earned greater confidence, although the connection from computing growth to durable profit and cash per share remains the fundamental question.

The Service Behind the Silicon

A customer can own the processors and still need Oracle. Power, cooling, networking, software and operations must work together before expensive equipment becomes useful computing. A larger fleet can spread engineering knowledge across deployments and make it easier to reassign capacity as contracts expire. Customer-owned hardware therefore helps reveal what Oracle contributes beyond financing the purchase, provided those operating gains outweigh the complexity of expansion.

Management reported delivering 850 megawatts and more than 300,000 GPUs during Q1, with GPU utilization at 97.9%. At Abilene, customer acceptance reportedly took approximately 24 hours. OCI revenue provides the financial counterpart to those operating claims.

OCI grew approximately 28% sequentially, almost exactly meeting our $7.40 billion forecast and exceeding the $7.09 billion consensus in our pre-release supplement. Remaining performance obligations, or contracted revenue still to be recognized, rose from $638 billion to $664 billion, below our $700 billion expectation. Those differently scoped figures cannot calculate a backlog-conversion rate. They do show why delivery increasingly deserves attention alongside commitments.

The intriguing disclosure concerned older processors: capacity renewed or resold during Q1 achieved a 20% premium, with most of those GPUs at least four years old. That challenges immediate-obsolescence arguments. It does not yet establish durable pricing power because the cohort renewed during scarce supply. Oracle’s advantage becomes more convincing if attractive renewal economics survive greater customer choice.

An Advance Is a Promise

Our Q4 article already recognized the $75 billion pool of prepaid and customer-supplied hardware. It also drew a boundary:

“But Q4 also drew the limit of the argument. BYOH changes the return on Oracle’s capital; it does not make Oracle capital-light. If the customer-funded model solved the problem outright, Oracle would not still need another large year of external financing. The bridge is real. It is not free. It reduces the capital burden per contract; it does not eliminate the capital burden of chasing a market this large.”

The latest release makes that distinction measurable. These are quarterly figures, in US$ billions:

Nearly half of Q1’s $23.10 billion operating cash flow came from the identified prepayment contribution. The $11.36 billion is a quarterly flow, not the outstanding customer balance. It reduces external funding needs while paying for services Oracle still owes. It is already included in operating cash flow; subtracting net capex would count its benefit twice.

Despite that contribution, net capital outlay increased approximately 60% sequentially and reported free cash flow became more negative. Collections and construction payments are lumpy, so this does not establish a worsening trend. It does rule out describing Q1 as a demonstrated cash turnaround.

Supplier financing, customer-owned hardware and prepayments also allocate risks differently. A smaller initial funding requirement improves returns only if sufficient economics remain after operating costs, replacement obligations and contractual concessions. Management says most newly added contracts use these arrangements and affect FY28 or later; the release does not reconcile those contracts with this quarter’s prepayment flow.

The counterparty question deserves equal weight. Asked whether funded contracts came from AI labs, chip companies or sovereigns, Magouyrk described customers ranging from startups to major investment-grade companies. Management describes broad participation, without revealing the dollar-weighted distribution.

Several customers can still represent a common economic exposure if their spending depends on the same investors, suppliers and financing conditions. If a chip supplier finances a customer that prepays Oracle, which then buys that supplier’s chips, several transactions depend on the same capital. These disclosures do not establish that such a chain explains the $11.36 billion. Cash already received reduces collection exposure on prepaid services; dependence on future fundraising can still threaten subsequent demand and capacity commitments.

I therefore want to know who owes what, under which terms, and how much demand customers can sustain from their own cash generation. Oracle’s operating success and its customers’ commercial success must eventually reinforce each other. A growing advance balance alone cannot demonstrate that relationship.

The Profit Has to Follow

Management’s margin argument is coherent. Compute has lower gross margins than software but requires less selling and research expenditure relative to revenue. Adjusted operating margin remained around 42% year over year, helped by lower combined selling, research and administrative expenses.

The forward arithmetic is less comfortable. Q2’s revenue midpoint implies approximately $21.20 billion, up 9.6% sequentially, while adjusted EPS of $1.89 would be below Q1’s $1.92. Using that midpoint, full-year guidance requires at least approximately $49.46 billion of second-half revenue and roughly $4.29 of second-half EPS, allowing for share-weighting differences. Revenue is growing faster than the guided earnings accruing to each share.

Hilary Maxson acknowledged the margin pressure in the call:

“Yeah. So we had mentioned already, I mentioned in the Q4 that we would expect a step down in gross margins this year. You can see the EPS guidance that we give, though, so you can see what we might expect in terms of operating margin.”

The EPS comparison cannot isolate operating margins: tax, interest, other income and share count also matter. Interest expense rose 55% year over year versus 31% growth in adjusted operating income, and Oracle completed its $20 billion gross equity issuance. Q1 EPS beat our $1.75 forecast, but the full-year guide rose only five cents to $8.10. Execution deserves greater confidence; extrapolating the quarterly earnings beat would go beyond the evidence.

Enterprise software offers another source of profit. Our earlier thesis suggested computing relationships could deepen Oracle’s database advantage. That connection needs evidence. Multi-cloud database revenue grew 353%, but distributing Oracle databases inside customers’ chosen clouds also demonstrates that the franchise can prosper without Oracle owning every surrounding computing workload.

Co-CEO Mike Sicilia supplied a more direct mechanism for improving the applications business:

“In our Oracle NetSuite applications, we have seen early customers leveraging these AI tools coming down from double-digit months down to single-digit weeks in order to be able to go live in production. We think that does two things. Number one, it helps customers get to value from AI more quickly than ever and certainly at a lower cost. Number two, in some of these very complex industries, there are ramps associated with these go-lives, and it allows us to unlock the ramp and recognize revenue more quickly than we have in the manual implementation piece.”

Faster implementation can lower customers’ costs and accelerate Oracle’s revenue. That is a credible advantage, although SaaS growth of 10% has yet to reflect a comparable acceleration. I would underwrite computing on its own economics and treat additional enterprise profit as something Oracle must demonstrate.

Three Prices, Three Businesses

My variant perception is that Oracle’s operating service may prove more durable than a simple bet on scarce GPUs suggests. Customer-owned hardware makes that service visible. The opposing interpretation is equally clear: customers are securing scarce capacity while Oracle accepts obligations whose returns become less attractive as supply expands.

This tension also explains how different investors may read the quarter. Growth investors have stronger delivery evidence; investors prioritizing predictable cash returns still face negative free cash flow and dilution. A broader shareholder base could emerge as funded contracts become demonstrably durable cash earnings. That is a conditional valuation argument, rather than evidence that a new investor category automatically pays a higher multiple.

Our published Q4 ranges were $100–140 for the bear case, $225–290 for the base and $440–625 for the bull. The updated points remain within or near those ranges; Q1 increases my confidence in the central outcome.

The bear assumes severe margin compression as compute becomes a larger, less profitable business. Revenue of $140 billion at 28% operating margin produces $39.2 billion of operating profit; financing, tax and the common-share adjustment leave $25.46 billion, or $7.72 across 3.30 billion shares. EPS therefore falls slightly below FY27’s $8.10 guide despite substantial revenue growth. Oracle can succeed at building the business while disappointing its shareholders.

The base assumes reliable delivery, continued customer funding and contained financing costs. An 18× multiple does not make those business assumptions conservative: reaching $180 billion means doubling FY27’s revenue guidance floor in two years, a 41.4% annual growth rate. The bull requires operating differentiation to survive greater supply while enterprise profits support stronger margins.

At the dated $164 reference, the base offers approximately 18% annualized appreciation. I find that attractive enough to be moderately bullish, with confidence tied to evidence. Adjusted EPS must still become cash available to shareholders, and the bear price provides no floor.

What Would Change My Mind

October’s Investor Day should clarify project economics, funding obligations and the margin trajectory. I want a reconciliation of opening advances, new funding, services delivered and closing obligations, together with customer concentration and delivery periods. These are requested disclosures, not information management has promised to supply.

At Q2 results, expected December 14, the immediate tests are the $13.16–13.64 billion cloud revenue guide and an explanation of earnings conversion. Through FY27, I will track the ≤$70 billion net-outlay guide alongside dilution and reported free cash flow. Higher investment becomes adverse evidence when it fails to bring stronger economics; a missed disclosure alone leaves the thesis unproven.

Across subsequent quarters, attractive renewals must survive broader equipment cohorts, and faster implementations must produce measurable enterprise growth. Repeated delivery disappointments, weaker contract returns or financing that consumes operating progress would reduce my conviction. Management has given no date for positive consolidated free cash flow.

Last quarter, we accepted that Oracle had found ways to finance the opportunity. This quarter earns more confidence that it can deliver. The next judgment concerns what delivery leaves behind: after fulfilling customers’ prepaid claims and paying the costs of expansion, how much more profit and cash belongs to each existing share? That is what Oracle gets to keep, and what will determine our view from here.

Subscribe now

• The growth transition is finally becoming visible: organic net-new ARR grew 17%, non-seat products exceeded 100% ARR growth, Zero Trust Everywhere reached 950+ enterprises, and Z-Flex TCV surged above $770 million.

• The thesis has become more precise: Zscaler does not need to become the “brain” of AI security. Identity, endpoint and other systems can make decisions upstream; Zscaler can still become increasingly valuable as the privileged place where those decisions are enforced.

• The new uncertainty is economic: machine traffic expands the number of interactions Zscaler can secure, but transactions have risen above 750 billion per day while FCF margins have weakened and CapEx is rising. The question is whether this is temporary equipment spending or structurally greater capital intensity.

“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.”

That was where I left Zscaler three months ago. Q3 had not broken the business: revenue and ARR were still growing 25%, Z-Flex was expanding, non-seat usage was scaling, and margins had reached a record. What broke was the inference I had been drawing from those facts. A valuable architecture did not necessarily mean Zscaler would own the most valuable security decision in an AI world; intelligence could increasingly move upstream toward identity, endpoint, code, browser and agent governance while Zscaler remained indispensable plumbing.

Q4 should be judged against that correction, not used to erase it.

On that basis the quarter delivered something much more useful than another beat: a split verdict. The growth transition we had been waiting for became considerably more tangible. Organic net-new ARR accelerated, non-seat products kept taking share of new business, Z-Flex took another extraordinary step higher, Zero Trust Everywhere adoption jumped, and the sales disruption that haunted Q3 largely disappeared from the numbers.

The cash-flow test, though, did not pass. And management’s FY27 guide still assumes remarkably little of Q4’s acceleration survives.

The first finding makes me more constructive. The second prevents the old bull case from simply returning.

What Did Not Change

Across this series I have used several analogies to describe Zscaler: Salesforce, ServiceNow, and most recently the cardiovascular system. The changing metaphors reflected a real evolution in the thesis, but they also risked making the underlying view appear less stable than it was. The architecture thesis has barely changed. What changed was how much economic dominance I was willing to infer from it.

Zscaler’s original insight was that the corporate network was the wrong thing to trust. Instead of putting a user onto that network and then protecting everything behind it, the Zero Trust Exchange connects a specific entity to a specific application according to policy. The architecture governs an interaction rather than protects a location.

That becomes more relevant as the entity changes from an employee to a workload, branch, device, model or autonomous agent. Zscaler’s 10-K now explicitly describes the Exchange as connecting users, devices, workloads and AI agents to applications, models and data according to identity and business policy.

Q3’s correction was that circulation is not cognition. The cardiovascular system can be essential without being the brain. If the highest-value security decision increasingly occurred in identity, endpoint or code before traffic ever reached Zscaler, then the Exchange could remain extremely durable without becoming the control point I had once assumed.

The burden of proof therefore became economic. Zscaler had to demonstrate that workloads, branches, data, agents and the products surrounding them could become large enough to alter company-level growth before the original user-security business matured.

Q4 began to do exactly that.

The Transition Appears in the Numbers

The headline quarter was excellent. Revenue reached $898 million, ARR $3.77 billion, net-new ARR $246 million and non-GAAP operating margin a record 24.3%. The $246 million net-new ARR result also just cleared the roughly $245 million investor bar we had identified before earnings.

But those are not the figures that change the thesis. These are:

Excluding Red Canary, net-new ARR was $232 million and grew 17%, after growing 7% in FY25 and 10% in the first half of FY26. Zero Trust Everywhere enterprises have moved from more than 350 at FY25-end to more than 700 in Q3 and more than 950 in Q4.

Z-Flex supplies the commercial connection. Q4 TCV exceeded $770 million, up more than 60% sequentially, while FY26 Z-Flex customers increased ARR by nearly 30% on average.

I used to describe this as lock-in; that was too strong. Z-Flex is better understood as commercial pre-commitment. Once a customer has committed a multi-year budget, adding another Zscaler capability no longer requires starting procurement from scratch. The next product is easier to adopt than the first.

That creates what I think is Zscaler’s more defensible economic advantage: economies of scope around an enforcement point. More of a customer’s estate comes under common policy; each additional use case increases the usefulness of consistent enforcement; Z-Flex reduces the friction required to expand that footprint.

This is why the 30% non-seat figure matters more to me than the revenue beat. An employee-seat business grows with headcount. A security system priced around workloads, branches, data volumes and machine activity can grow with the number of interactions being secured.

AI matters because it radically increases those interactions.

Upstream Was Too Binary

Jay Chaudhry made the distinction unusually clearly:

“While identity solutions answer ‘who’ is requesting access, our in-line platform determines ‘what’ that user or agent should be allowed to do.”

Q3 framed the emerging security world as anticipation versus interception. That now looks too binary.

Symmetry’s Access Graph maps relationships among entities, applications and data. Data Security and Security for AI add context about sensitive information, models, prompts and permissions. The Zero Trust Exchange applies policy to the actual interaction. Agentic SecOps, combining Zscaler telemetry with Red Canary’s operating expertise, is intended to investigate and remediate after something happens.

The ambition increasingly looks like:

understand → decide → enforce → observe → remediate.

That does not mean Zscaler suddenly becomes the brain. Microsoft begins with identity, endpoint and distribution; CrowdStrike with endpoint telemetry and response; Palo Alto with network-security incumbency and consolidation. Each has a legitimate claim on an important part of the decision chain.

Zscaler’s more interesting proposition is that intelligence gathered elsewhere becomes more useful when attached to an unusually powerful place to turn that intelligence into action.

In other words, Zscaler may not need to own every upstream security decision. It needs to remain the privileged place where those decisions are enforced.

Q4 makes that possibility materially more credible.

The Bill We Do Not Yet Understand

There is, however, a potentially uncomfortable consequence to the same transition.

As Zscaler secures more workloads, branches and machine interactions, the Exchange itself must carry more traffic and make more real-time decisions. Zscaler now processes more than 750 billion transactions per day, compared with roughly 500 billion in earlier periods.

At the same time, the cash-flow economics have weakened. FY26 FCF margin fell to 23%, and FY27 guidance remains only 23–23.5%. CapEx is expected to reach the low teens as a percentage of revenue.

It is tempting to make those facts one elegant story: machine traffic expands, Zscaler needs more equipment to inspect it, and the new growth model is inherently more capital intensive.

Q4 does not prove that.

Management attributes much of the increase to higher memory, storage and processor prices, tighter availability, and purchases that were accelerated when equipment became available. There are therefore at least three things inside current CapEx: component inflation, purchasing timing, and potentially a structurally larger equipment requirement as non-human traffic grows.

The distinction matters enormously.

Our old bull case effectively wanted the revenue opportunity of traffic-based security with the capital intensity of seat-based SaaS. That may have been too generous. But assuming FY27’s unusual equipment spend is the permanent economics of the business would be equally premature.

This is now a second fundamental question alongside growth: what does machine-scale enforcement cost once component prices and purchase timing normalize?

We do not know yet.

The Model Still Does Not Believe the Quarter

There is an even cleaner contradiction.

FY27 ARR guidance is $4.396–4.426 billion. At the midpoint, subtracting Q4’s $3.771 billion ending ARR implies about $640 million of FY27 net-new ARR. Organic FY26 ARR increased by roughly $615 million.

In other words, management’s FY27 guide implies only about:

4% growth in annual net-new ARR.

Q4 organic net-new ARR grew 17%.

Oppenheimer challenged Kevin Rubin on exactly this calculation. Rubin did not dispute it. He instead pointed to the sales-leadership transition, the ramp of Agentic SecOps, stable 115% NRR, investment in new-logo coverage and AI tailwinds. He also said Zscaler had not specifically called out the contribution from Security for AI in guidance.

Both numbers cannot describe the normalized future.

Either Q4 was an unusually strong fiscal-year-end quarter and the FY27 model is approximately right, or management has built considerably more conservatism into the annual guide than current demand warrants.

That is also why Zscaler has not yet recovered like several cybersecurity peers after the AI scare. The argument that AI itself disrupts Zscaler looks increasingly weak. What remains unresolved is whether Zscaler’s strategic relevance translates into faster financial growth.

CrowdStrike and Palo Alto have already persuaded growth investors of that connection. Zscaler has connected its story to leading indicators while still formally guiding like a 16–17% company.

The market is entitled to wait.

There are disclosure reasons to wait too. Security for AI bookings increased more than 50% sequentially and pipeline 75%, but management still does not provide a clean Security for AI ARR figure. Z-Flex TCV is spectacular, but longer contracts are not the same as recurring-revenue acceleration. RPO growth has fallen from 35% in Q1 to 27% in Q4, only modestly ahead of revenue, meaning our old “speedometer versus engine” argument has lost much of its informational advantage.

The mechanism increasingly looks real. The model still assumes it barely matters. One of them must move.

Three Prices, Three Zscalers

For a company at this stage, I would value the scenarios primarily on EV/revenue, with growth, operating margin and FCF margin determining what multiple is deserved. FCF is an important quality check; it should not be allowed to answer an unresolved capital-intensity question by itself.

Our estimates. Price ranges use FY29 revenue implied by the stated CAGR, approximately 180 million diluted shares and current net cash as a conservative capital-structure assumption.

The bear case does not require Zscaler to lose relevance. The user-security core simply matures faster than non-seat security can compensate, while higher capital intensity prevents a premium valuation.

The base case is more interesting because it requires no return to hypergrowth. Zscaler compounds revenue around 17–18%, improves FCF toward the mid-20s, and proves that workloads, branches, data and agents can sustain the business as seat growth matures. Once the Q3 uncertainty has been resolved favorably, I think a 7.5–8.0x terminal revenue multiple is reasonable; a business still growing high-teens with mid-20s cash margins would deserve more than the valuation of a permanently decelerating SASE vendor.

The bull case requires actual company-level reacceleration. Security for AI and Agentic SecOps need to become meaningful recurring businesses, non-seat adoption needs to move materially beyond 30%, and the incremental equipment burden has to prove scalable enough for FCF margins to approach the high-20s.

Using midpoint outcomes gives a probability-weighted three-year value around $245, versus roughly $178 today. The important change from Q3 is not that the bull case deserves to be inflated again. It is that Q4 makes the pure maturity case less likely.

That is a healthier reason for fair value to rise.

The Trial Continues

The variant perception has narrowed considerably over the last year. It is no longer “the market sees software; we see security infrastructure.” That was too easy.

The market now sees a strategically relevant company whose growth is settling around 16–17% and whose cash economics may be becoming somewhat more capital intensive. Management wants investors to see a deliberately cautious guide sitting on top of accelerating organic net-new ARR, record sales productivity, rapidly growing non-seat security and AI products that are still early.

I think the market is right to withhold the old premium multiple. I also think it is increasingly difficult to look at 17% organic net-new ARR growth, >100% non-seat ARR growth, 950+ Zero Trust Everywhere enterprises and $770 million of Z-Flex TCV and conclude that nothing underneath the headline growth rate is changing.

Q3 correctly raised the burden of proof. Q4 has begun to meet it.

The remaining tests are clearer now. Organic net-new ARR needs to stay in the mid-teens rather than collapse back toward the ~4% annual growth embedded in guidance. Non-seat contribution needs to move toward 35%. Z-Flex must convert commitment into ARR. Security for AI needs to graduate from bookings and pipeline to recurring revenue. Agentic SecOps needs to become financially visible in the second half.

And the newest test may be the most important one we were not asking three months ago: whether FY27’s elevated equipment spending normalizes as prices and purchase timing normalize, or whether carrying machine-scale traffic really does require more capital than our old model assumed.

The architectural thesis did not change. Q3 changed what I was willing to infer from it. Q4 changes the evidence.

The transition is increasingly visible in the operating data. What we still do not know is the normalized cost of running it.

That is what FY27 now must answer.

Subscribe now

• The prediction arrived: product-revenue growth accelerated from 29% to 37%, and management says AI contributed roughly half of that seven-point improvement.

• CoCo is becoming more than an interface: by shortening migrations, reducing implementation friction and helping customers build faster, it can pull future Snowflake consumption into the present.

• The moat is still unproven: growth has accelerated, but retention has not yet reaccelerated, AI gross margins are lower, and Snowflake still needs to show that faster creation leads to deeper customer dependence rather than just more expensive consumption.

In December 2025, after Snowflake reported product-revenue growth of 29%, I argued that the most important number was not revenue but the 800-basis-point gap between revenue growth and RPO growth. Enterprises were committing faster than they could deploy (Snowflake 3QFY26 Earnings: The 800 Basis Points Nobody Noticed) I ended with a prediction: “The ‘slowing growth’ narrative has about two quarters left to live.”

The stock fell 8%, then kept falling. By April it had reached $118, down 56% from year-end. The market’s argument was no longer that Snowflake was experiencing a temporary optimization cycle. It was that agentic AI would make the company structurally less necessary. If agents could reach into enterprise systems, retrieve data and act on it directly, why would anyone need a centralized data layer standing in the middle?

That was the point of maximum discomfort because the architecture made sense before the consumption evidence did. RPO accelerated, enterprises signed record nine-figure commitments, and Snowflake’s governed context became more valuable as models became interchangeable. Yet net revenue retention remained stuck near 125%, product growth had slowed, and management could not show how much AI was contributing. We were not arguing merely that the stock was cheap. We were arguing that the market had the direction of causality wrong: agents would increase the need for governed data and reduce the difficulty of creating workloads on Snowflake.

The evidence then arrived. Product-revenue growth accelerated to 34% in Q1 and 37% in Q2. Full-year guidance rose by $180 million last quarter and another $230 million this quarter. Asked how much of the seven-point acceleration since Q4 came from AI products, CEO Sridhar Ramaswamy answered: “I would roughly call it even.” Faster migrations and other core workloads supplied the balance. The two-quarter prediction turned out to be almost literal.

Last quarter, I wrote that CoCo was not the product; it was the interface (Snowflake 1Q27 Earnings: Second Interface Collapse). That description now looks incomplete. An interface helps a user operate a product. An engine changes what the product can produce. By reducing the time and expertise required to migrate data, build pipelines and deploy applications, CoCo is beginning to create demand for the underlying business.

This matters because Snowflake sells consumption, not access. Direct revenue from running CoCo is useful; the workload CoCo leaves behind may be far more valuable. The architecture has begun to appear in the income statement. We were right about the mechanism when the market was pricing the opposite. That does not mean Snowflake has already won the agentic enterprise.

The Big Fundamental Question

Can Snowflake turn its governed enterprise context into a self-reinforcing engine in which AI creates workloads faster, those workloads deepen core consumption, and the resulting usage strengthens Snowflake faster than commoditization and inference costs weaken it?

Powerful models have become broadly available, shifting the enterprise bottleneck from model capability to the ability to connect models safely to governed data and turn an idea into a working application. Snowflake becomes more valuable if CoCo and CoWork compress that process, because each new workload creates direct AI usage, recurring core consumption and deeper organizational dependence. The mechanism breaks if generic agents make workloads portable, retention fails to improve, or lower AI margins consume too much of the value being created.

The Interface Became a Demand Engine

Snowflake’s original achievement was a form of interface collapse. By separating storage from compute and automating much of the operational complexity of a traditional data warehouse, Snowflake reduced the expertise required to put data to work. Customers could begin with one workload and expand as consumption justified itself.

CoCo attempts a second collapse. A customer can use conversational language to write code, debug queries, build pipelines and migrate legacy systems. This is often described as productivity, but that word is too small. The strategically important effect is a reduction in the activation energy required to consume Snowflake.

There are three layers to the economics. First, customers consume CoCo, CoWork and other AI services directly. Second, those products induce core consumption by helping customers build things that run on Snowflake. Third, they accelerate consumption by pulling future workloads into the present.

Management offered a useful example. A large network-equipment manufacturer is completing a Teradata migration in less than three quarters; Snowflake estimates that the same project would previously have required two to three years. The economic value is not confined to the cost of the coding agent. Snowflake begins earning from the migrated estate perhaps eighteen months earlier than it otherwise would have.

Ramaswamy disclosed another metric that may matter more than the AI account count: time to 80% of purchased consumption. Snowflake measures how quickly each new-customer cohort reaches 80% of its contracted capacity, and management says the newest cohorts have improved “very, very visibly.” This is the mechanism our earlier writing anticipated. AI is not merely adding a new workload category; it is shortening the distance between a commercial commitment and actual consumption.

Snowflake did not disclose the magnitude of that improvement. It should. If time to 80% continues falling, it would link AI adoption, implementation velocity and revenue conversion in one number. If it stalls, the interface may be popular without being economically transformative.

Seven Points, One Engine

The quarter itself was unusually strong. Product revenue reached $1.49 billion, up 37%, while total revenue increased 35% to $1.55 billion. Snowflake added 692 net new customers, up 32%, reaching 14,554. Deployed customer use cases rose 89%; use cases won per account executive increased 43%.

CoWork expanded to 5,800 accounts, up nearly 11% sequentially. CoCo surpassed 9,100 accounts after adding more than 2,000 during the quarter. Management raised FY27 product-revenue guidance by $230 million to $6.07 billion, or 36% growth, and guided Q3 to 37–38%. This was not simply a quarterly beat flowing through the annual number. Snowflake embedded a higher observed consumption curve into the second half.

The historical thesis ledger now looks different:

Source: Snowflake Q2 FY27 earnings release, supplemental materials and earnings call.

There are reasons to retain some restraint. Net revenue retention was 126%, healthy but unchanged from Q1 rather than breaking higher. Total RPO growth slowed to 30%, though the 54% expected to convert over the next twelve months implies roughly $4.86 billion of current RPO and a much stronger near-term conversion signal. Snowflake also refused to disclose the consumption uplift of AI adopters relative to comparable non-adopters.

Management said the effect was “pretty noticeable” across cohorts, then supported the claim with anecdotes about new use cases. Those anecdotes establish breadth. They do not establish lifetime value. Q2 proves acceleration; it does not yet establish its duration or terminal level.

Management Is Selling the Destination

Management’s chosen sentence was unambiguous: “The agentic enterprise runs on Snowflake.” The company says AI brings new data and workloads onto Snowflake, its first-party products attract new users, and those users consume more across the core. Because Snofwflake already sits amid enterprise data, permissions, business definitions, models and workflows, management believes it can become the governed control plane connecting intelligence to action.

Our story now reconciles with management’s on the mechanism. We agree that direct AI revenue understates the opportunity. We agree that models become more useful when they operate against governed enterprise context. We agree that reducing the difficulty of migration and application development can accelerate the existing business.

The disagreement concerns how much of the destination has already been earned.

Our earlier articles sometimes moved too quickly from valuable to defensible. Data sharing creates ecosystem benefits, but it is not a telephone network in which every participant benefits from every new connection. Governance matters, but Databricks, Microsoft, Google and AWS can also build catalogs, semantic layers and governed model access. Snowflake possesses a strong starting position, not a permanent claim on the category.

The deeper potential advantage is the engine itself. Enterprise data and permissions already reside in Snowflake. CoCo understands the environment in which the work must run. Customers can create and deploy workloads faster. Those workloads generate consumption and new operating patterns, which Snowflake can use to make the next project easier.

If this engine produces faster time to value than competing approaches, Snowflake becomes progressively harder to displace. If generic agents can reproduce the same outcome across open formats and different systems, CoCo becomes an effective feature rather than a category-defining advantage.

Management has now supplied evidence for the mechanism we anticipated. It has not yet supplied evidence for the empire it is building around that mechanism.

One further observation is worth adding, because it will matter in the months ahead. Databricks was not named by management or by any of the nine sell-side analysts on the call — a silence that could reflect strategic positioning, analyst preference for the acceleration narrative, or simple absence. Whatever it reflects, the Databricks IPO, whenever it arrives, will produce the first public comparable that is not a consumption-software peer. That event will force sell-side reference classes to update in one direction or another. Until then, templates will default to the nearest available comp, which is Datadog.

The Wrong Reference Class?

There is a subordinate question worth naming, because it may determine how much of the destination Snowflake earns actually gets captured in the stock over the next four to six quarters.

At $118 in April, Snowflake was priced as an impaired data warehouse. At $376 today, it is increasingly priced as an AI-beneficiary consumption business — the sell-side reference class of Datadog, MongoDB and Confluent, adjusted for scale and growth. That reference class produces a fair-value ceiling in the low-to-mid $400s through FY28 on the raised revenue base, which is roughly where sell-side price targets will cluster over the next several weeks.

But if Snowflake genuinely becomes what management is claiming — a governed control plane for enterprise agent workloads — the correct reference class is not consumption software. It is a category that does not currently have a clean public comparable, and whose historical analogs (transaction networks, benchmark providers, certain regulated data services) have earned meaningfully higher multiples on the strength of amortization economics and structural switching costs. The gap between the two reference classes is roughly 400–500 basis points of terminal EV/revenue multiple, or approximately $100–175 per share on FY28 estimates.

This is not, on the current evidence, a case for a $700 stock. It is an observation that Snowflake’s valuation ceiling depends on which reference class the market eventually settles on, and that the market cannot settle on the higher class until measurable evidence forces the update — meaningful governance-related consumption disclosure, retention improvement among agent-heavy customers, or a public comparable that breaks the current templates.

Q2 strengthened the possibility of that reclassification. It did not complete it. A new reference class should be the consequence of measurable retention, governance economics and operating leverage — not a substitute for that evidence.

Faster Creation Has a Cost

The second unresolved question is who captures the value.

Non-GAAP operating margin reached 15.3%, more than four points higher than a year ago, and Snowflake raised its FY27 operating-margin guidance from 13.5% to 14.5%. Year-to-date headcount additions fell sharply once the Observe acquisition is excluded. Management also described material internal savings from using CoCo and CoWork across marketing, finance and sales. There is real operating leverage here.

Product gross margin tells the other side of the story. It fell to 74.7%, and Snowflake reduced the full-year outlook from 75% to 74%. CFO Brian Robins explained the sequence plainly: build a great product, drive adoption and revenue, and then work on the margin implications. That is rational product management, but it also confirms that AI economics come after adoption in the current priority order.

The non-GAAP presentation deserves context. Snowflake still reported a $263 million GAAP operating loss, while stock compensation was approximately $456 million, nearly 30% of revenue. This does not invalidate improving operating leverage, but shareholders are funding part of it through dilution.

The relevant question is not whether gross margin declines by one percentage point. It is whether Snowflake is sacrificing one point of margin to create several additional points of durable growth, or replacing high-margin core growth with lower-margin AI activity. A lower percentage margin can create more value if incremental gross-profit dollars compound faster and customer dependence deepens. Management has shown the growth; it has not shown that complete economic bridge.

This is also where the current variant perception lies. The market now understands that Snowflake benefits from AI. What it may still misunderstand is that the decisive variable is neither AI account adoption nor a standalone AI revenue figure. It is the time required to turn enterprise intent into recurring consumption — and the gross profit Snowflake retains from doing so.

Three Prices, Three Snowflakes

At roughly $375–380 after the results, the stock no longer prices Snowflake as an AI victim. The three-year outcome depends on what kind of company this engine creates.

Scenario estimates are our assumptions, anchored to FY27 product-revenue guidance of $6.07 billion.

In the bull case, CoCo and CoWork become the preferred interfaces for enterprise data and agent development. Retention rises above 130%, revenue compounds near 30%, and operating margins reach the high 20s. Snowflake becomes the governed action layer management is selling.

In the base case, AI makes Snowflake a better data business without creating a new category. Growth settles in the low-to-mid 20s, retention remains around 125–128%, and gross margins remain below historical levels. This is an excellent company, but not a mispriced one.

In the bear case, open formats and generic agents reduce differentiation. AI expands with weaker economics, retention falls below 120%, and workloads move between competing systems with limited friction.

The probability-weighted midpoint is approximately $370, close to the post-earnings price. Conviction in the business has increased; the asymmetry in the stock has disappeared. I would hold rather than chase here, become interested around $325–335, and require higher retention and stable gross margins before underwriting prices above $420.

The Question After the Question

A year ago, the central question was whether AI would disintermediate Snowflake. Last quarter, it became whether AI could accelerate Snowflake. Q2 has substantially answered both: AI is making Snowflake more useful and driving the core business faster.

Six signposts will determine how powerful the engine becomes. Product growth should remain above 35% through FY27 rather than quickly reverting. Net revenue retention needs to move toward 128% and eventually 130%; continued stability at 125–126% would limit the compounding thesis. CoCo should exceed 12,000 accounts by year-end with evidence of repeat production usage, while CoWork must graduate from account adoption to large departmental deployments. Product gross margin should hold near 74%; two quarters below 73.5% would suggest that AI is becoming structurally more expensive. Most important, Snowflake should disclose either the improvement in time to 80% consumption or the consumption uplift of comparable AI-adopting cohorts.

The thesis has advanced. Snowflake no longer needs investors to believe that AI might someday help its business; management has begun to quantify the contribution. What was mispriced at $118 was direction; what is now at issue at $376 is destination. The burden of proof has moved upward from adoption to durability, from acceleration to customer dependence, and from revenue to retained economics.

The interface has become the engine. The next two quarters will tell us whether it also becomes the moat.

Subscribe now

• The scale is no longer in doubt: Broadcom now sees AI semiconductor revenue rising from roughly $58 billion this year to $115 billion in FY27 and $230 billion in FY28, driven by six XPU customers and extraordinary deployment visibility.

• But Broadcom’s model is fundamentally different from NVIDIA’s: NVIDIA integrates the AI factory around itself; Broadcom gives hyperscalers the tools to integrate it themselves. That makes customers more capable and potentially less dependent on any individual supplier.

• Networking is the key test of economic power: if Broadcom can expand Ethernet content across both custom-silicon and GPU clusters, scale may translate into greater value capture. If not, Broadcom may become enormous while the hyperscalers retain most of the strategic leverage.

The easiest mistake in investing is to confuse becoming larger with becoming more powerful. Broadcom’s fiscal third-quarter report made that distinction impossible to ignore: management described an AI semiconductor business expanding from roughly $58 billion this year to $115 billion in fiscal 2027 and $230 billion in fiscal 2028 yet left unanswered whether Broadcom’s share of the resulting economics will improve at all.

We were right that Broadcom would become one of the indispensable industrial companies of the AI buildout. We were less right about why. Broadcom is not integrating the AI factory around itself, as NVIDIA does; it is giving Google, Meta, OpenAI, Anthropic and other hyperscalers the tools to integrate the factory themselves.

That makes Broadcom’s opportunity larger than we previously imagined, but its power less certain. Revenue reached $29.6 billion, adjusted earnings were $3.32 a share, and AI semiconductor revenue rose 221% to $16.7 billion, beating Bloomberg consensus by 0.5%, 2.8% and 4.8%. Broadcom guided fourth-quarter revenue to $34.8 billion and AI revenue to $21.7 billion, but the shares ended our frozen after-hours window near $364, down 0.8%, because the guide cleared consensus without clearing the buy-side whisper. The quarter proved the volume. It did not yet prove the value capture.

Bigger Is Not More Powerful

Our original Broadcom thesis was that AI would increasingly be governed by the “physics economy”: advanced-node wafers, HBM, packaging yields, optics, power, delivery schedules and cash commitments rather than announcements about models and partnerships. That thesis has held up. We then stretched it too far by describing Broadcom as the prime contractor for the AI rack. Last quarter we corrected the metaphor: Broadcom does not own the rack but makes other companies’ systems deployable. Q3 supports that narrower and more durable formulation.

The evidence on execution is formidable. Broadcom now serves six XPU customers; Q3 XPU shipments grew more than 3.5 times year over year, AI networking grew more than 2.5 times, and management says supply has been secured for the $115 billion and $230 billion forecasts. Google is receiving two TPU generations, OpenAI’s first accelerator is shipping, Meta begins production shipments in Q4, and Anthropic is expected to become Broadcom’s largest XPU customer in fiscal 2027. These are no longer isolated design wins. Broadcom is operating several of the world’s most demanding semiconductor programmes simultaneously, although Hock Tan was careful to distinguish visible demand from deployed capacity:

“It will be 30 GW between the customers we have, the six customers we have. Question is, will all 30 come into production within these two fiscal years? We are giving you a judged number of $115 billion of chips to these guys in fiscal 2027, another $230 billion in 2028, which would add up, I know, to about $350 billion. Another way of saying is, we believe with a pretty high degree of confidence, we will ship $350 billion of AI semiconductors to these customers in the next two years. That’s the best way to look at it. It doesn’t necessarily mean all 30 GW need to have been deployed within that period.”

Thirty gigawatts is a demand envelope; $350 billion is management’s judgment about what can be funded, built, supplied and commissioned. If that forecast and Broadcom’s $20 billion–$30 billion of content per deployed gigawatt use comparable definitions, the arithmetic implies 11.7–17.5GW operating during those two years, or 39%–58% of the envelope. That is our inference, not guidance. It makes site readiness, financing and execution as important as chip demand.

The more consequential comparison is with NVIDIA. NVIDIA integrates the AI factory around itself: GPU, CPU, NVLink, networking, systems, algorithms and CUDA. Its stated revenue opportunity rises from roughly $18 billion per gigawatt with Hopper to $25 billion with Blackwell and $40 billion with Vera Rubin because each generation adds more NVIDIA-controlled content. NVIDIA uses scale to capture more of the system.

Broadcom enables the opposite architecture. The customer owns the workload, software, model and distribution, then uses custom silicon to internalize economics that NVIDIA would otherwise capture. Tan stated the proposition directly:

“The lesson here is when you co-develop a chip that is optimized for your particular LLM workloads, you will outperform any GPU. Like GPUs, Jalapeño demonstrates that it can also run other frontier models, and you can do all this at half the cost of a GPU.”

This is a management claim about selected workloads, not an independently demonstrated universal advantage. If it holds, the first-order benefit accrues to the customer: greater volume amortizes the design, lower token costs stimulate usage, and higher usage justifies more custom capacity. NVIDIA makes customers more dependent on NVIDIA; Broadcom helps customers become less dependent on any merchant compute supplier, potentially including Broadcom.

The strongest bull counterargument is what I would call the embedded-partner hypothesis. Each successful generation places Broadcom’s engineers, SerDes IP, packaging knowledge, interconnect design and supply commitments more deeply inside the customer’s roadmap. The hyperscaler retains strategic control, but changing industrial partners becomes progressively more expensive. Broadcom can also reuse design methods and supply knowledge across otherwise bespoke programmes. That could turn scale into lower development costs, stronger access to scarce components and a higher probability of winning the next generation.

The financial evidence is not yet conclusive. Q3 consolidated gross margin fell 210 basis points sequentially to 75%, and semiconductor gross margin was corrected on the call to approximately 67%, as memory-rich XPUs became a larger part of the mix. Operating margin still reached 67.9% because revenue grew far faster than operating expense. That is excellent execution, but the distinction matters: operating margin measures leverage during the ramp, while gross margin helps reveal how incremental value is divided among Broadcom, memory, foundry, packaging and the customer. We were right about scale and wrong about control. Broadcom is becoming essential to the AI factory, but the hyperscaler still owns the architecture, workload and distribution.

Hock Borrowed NVIDIA’s Grammar, Not Its Economics

Management made three revealing substitutions on the call. The horizon moved from Q4 to fiscal 2027–2028, gross margin was displaced by operating margin, and quarterly chip revenue became dollars per gigawatt. Each substitution is economically defensible. Each also directs attention away from a present weakness: the Q4 guide did not clear the whisper, product mix is compressing gross margin, and Broadcom’s content per gigawatt is expected to remain broadly stable while NVIDIA’s rises.

The timing was striking. Seven days after NVIDIA broke its quarter-only convention by presenting a supply-constrained fiscal 2028 outlook in gigawatts and revenue per gigawatt, Broadcom used almost the same grammar to describe its future. Broadcom’s customer roadmaps and supply commitments clearly predated NVIDIA’s call; what appears borrowed is the presentation architecture, not the demand. Hock borrowed NVIDIA’s language for describing the future and used it to describe the opposite system customer-owned rather than supplier-integrated.

Management’s story is that Broadcom knows the customers, workloads, chip roadmaps, supply requirements and deployment schedules well enough to forecast two years of extraordinary growth. I believe it. Our story is that this visibility establishes Broadcom as the preferred execution partner but does not establish increasing bargaining power or a rising share of customer economics. The stories reconcile on demand and execution. They diverge on value capture.

This reporting system will naturally appeal more to long-duration growth investors than to quarterly beat traders, although deliberate investor-base curation remains an inference. Investors should not accept management’s choice between gross and operating margin. Both matter, and the tension between them is part of the story rather than an inconvenience to be explained away.

The Network Is the Test

Networking is the best observable test of whether greater deployment can make Broadcom more powerful because it can spread across bespoke XPUs and third-party GPUs. XPUs accounted for 73% of Q3 AI revenue, leaving roughly 27% for networking, even though networking itself grew more than 2.5 times. Our previous requirement that networking remain above 40% of AI revenue was wrong: an exploding XPU denominator makes mix an unreliable measure of network strength.

The better measures are absolute networking revenue, networking dollars per deployed gigawatt, attachment within Broadcom-designed XPU clusters, and penetration of GPU clusters. Charlie Kawwas said Tomahawk 6 is deployed by “pretty much all” hyperscalers building XPUs with Broadcom and also by customers not using Broadcom XPUs; Tomahawk 6 and Tomahawk Ultra are appearing in both XPU clusters and some GPU clusters. If open Ethernet wins more of the scale-up domain while Tomahawk and Jericho retain scale-out leadership, the network can turn customer independence into Broadcom leverage.

The NVIDIA comparison still requires discipline. NVIDIA’s $40 billion-per-GW figure already includes NVIDIA networking and other system content, so Broadcom Ethernet in a GPU cluster takes a share of that wallet rather than adding another $20 billion–$30 billion on top. The contest is over who captures a finite AI-factory budget. Networking is where Broadcom can participate whether the compute is custom or merchant, making it the clearest bridge between revenue scale and economic power.

Certainty Is the Product

Frontier laboratories want more compute than their cash flows and credit profiles can independently finance. Broadcom’s financing vehicle bridges that mismatch with third-party capital; the first $35 billion tranche supports a 1GW Anthropic deployment, while Broadcom may provide residual-value guarantees. Silicon is the deliverable, but certainty is the product: certainty that the design will work, supply will arrive, the site will be financed, and successive generations will match the customer’s model roadmap.

That service can deepen relationships and pull revenue forward, but it changes the risk. Asked whether the roughly $29 billion of maximum exposure referenced from the previous filing indicated future backstops, management declined to provide a ceiling:

“We don’t have anything to announce today on residual value guarantees or backstop. There’s nothing new to add. The numbers that you outlined around what we’ve already done remain true. As I mentioned in my prepared remarks, we’re going to evaluate any strategic financing really on a deal-by-deal basis. We expect that any that we do in the future, they’re going to have unique features, and they’re going to be tailored specifically to the lab and to the investor needs. I can’t give you an overarching look at what’s the max and what each one’s going to look like.”

NVIDIA argues that its financed systems are fungible and redeployable, and its structure pays it on the hardware plus a share of rental revenue above a floor. Broadcom’s customer-specific XPUs are likely less fungible, while no comparable participation in application or rental upside has been disclosed. Financing does not prove weak pricing power; it proves that desired deployment exceeds some customers’ independent financing capacity. It creates value only if Broadcom is adequately compensated for deployment and credit risk.

Broadcom can absorb some risk. Q3 free cash flow was $13.7 billion, or 46% of revenue, and VMware’s 94% gross margin and 84% operating margin provide substantial ballast. Yet free cash flow missed consensus by 6%, receivables absorbed $2.86 billion of cash during the quarter, and Q4 capital expenditure is expected to rise from $532 million to $1.4 billion. The physics economy must eventually collect the cash it manufactures.

Three Prices, Three Worlds

At the frozen after-hours price of approximately $364, the bear case carries a 20% probability. Fiscal 2026–2029 revenue grows at 22%–25% annually, operating margin falls toward 56%–59%, fiscal 2029 adjusted EPS reaches $21–$24, and a 12–14 times multiple produces approximately $280–$330. This is the world in which Broadcom achieves volume without power: sites slip, customers preserve alternatives, networking capture disappoints, and financing obligations grow faster than cash flow.

The base case carries a 55% probability. Revenue compounds at 41%–46%, operating margin settles near 62%–64%, adjusted EPS reaches $35–$38, and a 17–19 times multiple yields approximately $630–$720. Broadcom remains the preferred industrial partner for customer-owned AI, the fiscal 2027 forecast largely converts, networking broadens, and financing stays contained. The company becomes indispensable without acquiring NVIDIA-like control.

The bull case carries a 25% probability. Revenue grows at 53%–57%, operating margin holds at 65%–67%, adjusted EPS reaches $43–$47, and a 21–23 times multiple supports $900–$1,050. Networking and reusable co-design knowledge turn greater scale into higher value capture, while financing proves both profitable and limited. A probability-weighted value near $675 implies an annualized three-year return of approximately 23%–24%, including dividends, but most valuation variance comes from how much of the $230 billion forecast is deployed, recognized, collected and achieved without underpriced guarantees.

The expected return is attractive, but another increment of conviction must be earned through networking growth across XPU and GPU clusters, fiscal 2027 revenue converting into cash, and maximum financing exposure becoming measurable. I would reduce the position if customer savings coincide with declining Broadcom gross economics or if receivables, deployment delays and guarantees grow faster than free cash flow.

Necessary Is Not Yet Powerful

Q3 increases my confidence that Broadcom will become much larger, while increasing only modestly my confidence that it will become structurally more powerful. The quarter confirmed the physics thesis, demonstrated rare execution across six programmes, and advanced Ethernet as the reusable layer across fragmented compute architectures. It also showed why the prime-contractor thesis was too generous: the customer integrates the system, owns the workload, captures the cost advantage and deliberately preserves alternatives.

Two observations would strengthen the thesis. Absolute networking revenue and networking content per deployed gigawatt must rise across both XPU and GPU clusters, and the fiscal 2027 ramp must convert into cash without a comparable increase in guarantees or working capital. Two developments would weaken it: falling semiconductor gross economics despite operating leverage, or hyperscalers adding alternative design partners without impairing their roadmaps.

The fiscal 2026 10-K and the next two deployment updates should resolve more than the next quarterly beat. Broadcom has become the company that makes hyperscaler independence possible. The next question is whether it can make itself indispensable without making its customers dependent.

Subscribe now

• The strategic thesis strengthened: Credo is moving beyond selling individual interconnect components toward understanding the health of the entire link, with PILOT turning diagnostics and failure prediction into a potentially system-level advantage.

• But technical importance is not yet economic ownership: the first major DustPhotonics photonic-chip wins did not include Credo DSPs, showing that customers may value Credo’s technology while still preserving modular supplier choice.

• The cost of the ambition is already visible: inventory, receivables and R&D are rising well ahead of the proof that optical attachment, margins and cash conversion will resemble AEC economics. The next test is whether broader technical relevance turns into broader wallet share.

A customer recently asked Credo whether its software could detect the slightest electrostatic-discharge damage to an optical transceiver. Not enough damage to stop the transceiver from working, but enough to create a latent defect that might eventually bring down a connection. Credo found a way to detect it.

“Now when we’ve got customers lighting up racks, they can determine immediately if a transceiver was mishandled and needs to be replaced.” — Bill Brennan, CEO

The striking part of this story is not that Credo’s technology worked. It is that the customer asked Credo to understand a failure before it occurred. A company customers call when its component fails is a supplier; a company customers call when the connection might fail is becoming something more important.

That distinction captures what management is selling. Credo does not want to be understood as an AEC company adding a collection of optical products. It wants to own enough of the connection from SerDes and DSPs to photonic chips, transceivers, firmware and diagnostics to understand how the entire link behaves. The more of the link Credo can see, the better it should become at identifying failure modes, improving its products and winning more of the customer’s next design.

The big fundamental question is whether that creates a compounding system or merely a wider product catalogue. The mechanism strengthens if every additional product gives Credo more visibility, that visibility produces better diagnostics and designs, and customers respond by adopting more Credo components. It breaks if customers value Credo’s engineering while continuing to assemble the link from whichever individual components they prefer.

Q1 supplied the best evidence yet that the first half of this mechanism is real. It left the second half unresolved.

The Test We Set Was Incomplete

Our previous article ended with a test:

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

The logic was straightforward. AEC demonstrated that customers would pay Credo for a complete, reliable copper connection. If Credo could move into optics without sacrificing its economics, its advantage had travelled beyond the product that created the company.

Q1 appeared to pass. Revenue reached $479 million, up 115% year over year, while non-GAAP gross margin held at 68%. Credo reported record optical DSP revenue, recognized its first silicon-photonics revenue following the DustPhotonics acquisition and began production shipments of ZeroFlap optical transceivers. Management continues to expect more than $600 million of optical revenue in FY2027, with DSPs, photonic chips and ZeroFlap each contributing more than $100 million.

The problem is that consolidated gross margin cannot answer a product-level question. AEC remains Credo’s largest business, and management did not disclose optical revenue, optical mix or optical gross margin. The established business can therefore preserve the consolidated number while the economics of the emerging business remain invisible.

We chose the right question but an insufficient metric. Credo has now proved that its technology can travel into optics. It has not proved that AEC-like economic ownership travels with it.

The Layer Between Products

Management’s story begins with heterogeneity. Future AI systems will not use one architecture, protocol or physical medium. Copper will remain appropriate for shorter reaches; optics will dominate longer ones; different scale-out and scale-up connections will demand different combinations of power, latency, reach and serviceability.

“Future AI systems will combine optical and copper interconnects across different reaches, protocols, and topologies.” — Bill Brennan

That world rewards product breadth, but breadth alone is not the thesis. Broadcom, Marvell and several optical suppliers can also provide pieces of the connection. Credo’s claim is that its products are joined by a common capability: understanding whether a link is healthy, why it is degrading and what should happen before it fails.

PILOT is the layer that potentially turns a group of products into a system. Brennan compared existing link monitoring to a green light when a connection works and a red light after it fails. Credo is trying to add the yellow “check engine” light in between. Its hardware continuously measures indicators such as eye height, signal-to-noise ratio and error patterns; its software looks for deterioration and can help remove a vulnerable processor or transceiver before a link flap disrupts the cluster.

This matters because AI clusters contain tens or hundreds of thousands of interdependent connections. Management says link instability can cost more than 10% of processor utilization, while slow cluster commissioning can leave enormously expensive equipment idle for weeks. A reliable connection is therefore valuable not because the cable itself is expensive, but because everything connected to it is.

The ESD example turns that argument from a presentation into customer behaviour. Credo is learning about defects beyond the normal responsibility of a component supplier. The accumulated failure cases should influence future DSPs, photonic chips, firmware and qualification methods. More deployments can create better products, which can make the next design easier to win.

This is not an automatic data network effect. One additional link does not mechanically improve every other link, and customers may limit how operational data is shared. The advantage compounds only if learning changes product performance and product performance increases attachment. PILOT is strategically important when diagnosis repeatedly becomes design; otherwise it remains an excellent feature bundled with hardware.

Importance Is Not Ownership

The ESD request and the DustPhotonics wins are mirror images. In the first, a customer asked Credo to understand the behaviour of the entire connection. In the second, customers selected one Credo component while declining another. Q1 therefore contained both the strongest evidence for integration and the clearest warning against assuming attachment.

The DustPhotonics wins provide the quarter’s cleanest counterargument. Asked directly whether the first two major photonic-chip design wins included a Credo DSP, Brennan answered:

“These first two major design wins that we are talking about do not include the DSP.”

That is simultaneously good and incomplete news. The wins validate the acquired photonic technology independently, which suggests Credo purchased a genuinely competitive product rather than a capability that only works inside its own bundle. They also show customers retaining control over component selection. Credo has won a place in the optical module without winning the integrated DSP-and-photonic-chip combination that management says should improve performance, power, yield and cost.

The modular challenge described in our previous article therefore remains. Customers may prefer an architecture in which validated components can be combined across suppliers. Credo believes that owning more of the stack will allow it to innovate faster and deliver better system outcomes. The customer will decide how much integration is worth and how much supplier dependence it is willing to accept.

This is the distinction the next phase of the thesis must preserve: Credo may own the customer’s problem before it owns the customer’s bill of materials. Technical influence often precedes commercial capture, but the second does not inevitably follow the first.

The Cost of Becoming a System

The financial statements show the cost of this ambition earlier than the product disclosures show its reward. Revenue grew 10% sequentially, but inventory increased 25% to $313 million and receivables increased 24% to $289 million. Free cash flow was $83 million against non-GAAP net income of $236 million, a conversion rate of 35%. Non-GAAP operating expenses reached $95 million, above management’s prior range, as Credo increased research and development across a much broader portfolio.

It would be easy to treat this as an earnings-quality complaint. That misses the strategic issue. Moving from a dominant AEC product into DSPs, photonic chips, complete optical transceivers, near-package optics and new scale-up connections requires more designs, qualifications, supply commitments and inventory configurations. Credo must incur those costs before the new products reach volume.

Management is explicit about the bet. Brennan said the company was “leaning in from a supply-chain standpoint” and playing the long game. The full-year arithmetic shows the scale of that commitment. More than 85% growth on FY2026 revenue of $1.335 billion implies at least $2.47 billion in FY2027. Q1 revenue plus the midpoint of Q2 guidance produces approximately $1.01 billion in the first half, leaving more than $1.46 billion for the second half—or over $730 million per quarter, 45% above the first-half quarterly average.

Inventory is not proof that this demand exists. It is proof that management is willing to finance its forecast. If inventory converts into the promised optical ramp while gross margin remains in the high 60s and cash conversion recovers, Q1 will look like the necessary investment behind a larger company. If product attachment remains limited and inventory continues outgrowing revenue, Credo will have broadened its technical capabilities faster than its economic ownership.

The balance sheet is therefore not a side issue. It is where the cost of management’s strategy has become measurable.

What Changed in Our View

Before Q1, we believed AEC gave Credo a wedge into the customer and that optics would determine whether its reliability advantage could travel. We expected gross margin to reveal the answer.

The quarter strengthened the strategic thesis. Optical participation is no longer theoretical. PILOT has moved from a diagnostic claim to a real customer workflow. Credo’s involvement increasingly extends beyond whether its own component functions, and management is committing capital as though the second-half ramps are real.

The quarter also made the economic thesis more demanding. Consolidated gross margin cannot isolate optical economics. The first photonic-chip wins do not attach Credo DSPs. PILOT has no disclosed pricing, attachment or separately observable contribution. Working capital and operating expenses are rising before investors can see how much of the broader connection Credo will capture.

I think Credo is building a genuine system-level advantage, and Q1 increased my confidence in that conclusion. What I am not yet prepared to assume is that customers will buy the entire system simply because they increasingly depend on Credo’s understanding of it. The probability that Credo becomes strategically important across copper and optics has increased; the amount of that importance it will monetize remains the central uncertainty.

Three Futures

The scenarios should be understood as three outcomes for the same mechanism, not three predictions about the semiconductor cycle.

The share-count assumptions reflect different outcomes for dilution and Credo’s ability to offset it through future cash generation. The ranges expose the operating assumptions that matter; they are not predictions of the path the shares take between now and then.

In the bear case, customers select Credo products but assemble the link modularly; optical margins dilute AEC economics, development and inventory remain elevated, and breadth does not produce attachment. In the base case, Credo owns the portions of the link where reliability is most valuable, while component and complete-system sales coexist. In the bull case, telemetry improves product design, better products deepen customer dependence, and that dependence produces greater attachment across DSPs, photonic chips and transceivers.

The signposts follow directly. By Q3, revenue growth should overtake inventory growth and cash conversion should recover above 60%. Optical revenue should exceed $600 million for FY2027 without consolidated gross margin falling below 67%. Over the next year, Credo should show that photonic-chip wins attach its DSPs, and that PILOT influences repeatable product adoption rather than isolated customer engagements. Failure on those measures would not mean the products are weak; it would mean the proposed compounding mechanism is.

The Company Customers Call

Credo did not cause the customer’s ESD problem. That was precisely why the incident mattered. The company was valuable not because its component worked, but because it could explain why the connection might eventually fail.

That is a more ambitious role than selling a cable, DSP or photonic chip. It gives Credo the opportunity to learn across products and become embedded in how customers design, qualify and operate their networks. It also requires Credo to carry more products, inventory and development expense before that opportunity becomes revenue.

The previous chapter of the Credo story asked whether its advantage could travel from copper into light. Q1 suggests that its technical relevance can travel. What remains unanswered is whether product attachment, margins and cash returns will travel with it.

Credo is learning to understand more of the link. The next question is how much of the link it will own.

General Disclaimer: The information presented in this communication reflects the views of the author and does not necessarily represent the views of any other individual or organization. It is provided for informational purposes only and should not be construed as investment advice, a recommendation, an offer to sell, or a solicitation to buy any securities or financial products.

While the information is believed to be obtained from reliable sources, its accuracy, completeness, or timeliness cannot be guaranteed. No representation or warranty, express or implied, is made regarding the fairness or reliability of the information presented. Any opinions or estimates are subject to change without notice.

Past performance is not a reliable indicator of future performance. All investments carry risk, including the potential loss of principal. This communication does not consider the specific investment objectives, financial situation, or particular needs of any individual.

The author and any associated parties disclaim any liability for any direct or consequential loss arising from the use of this material and undertake no obligation to update or revise it.

Subscribe now

• The strategic thesis strengthened: Credo is moving beyond selling individual interconnect components toward understanding the health of the entire link, with PILOT turning diagnostics and failure prediction into a potentially system-level advantage.

• But technical importance is not yet economic ownership: the first major DustPhotonics photonic-chip wins did not include Credo DSPs, showing that customers may value Credo’s technology while still preserving modular supplier choice.

• The cost of the ambition is already visible: inventory, receivables and R&D are rising well ahead of the proof that optical attachment, margins and cash conversion will resemble AEC economics. The next test is whether broader technical relevance turns into broader wallet share.

A customer recently asked Credo whether its software could detect the slightest electrostatic-discharge damage to an optical transceiver. Not enough damage to stop the transceiver from working, but enough to create a latent defect that might eventually bring down a connection. Credo found a way to detect it.

“Now when we’ve got customers lighting up racks, they can determine immediately if a transceiver was mishandled and needs to be replaced.” — Bill Brennan, CEO

The striking part of this story is not that Credo’s technology worked. It is that the customer asked Credo to understand a failure before it occurred. A company customers call when its component fails is a supplier; a company customers call when the connection might fail is becoming something more important.

That distinction captures what management is selling. Credo does not want to be understood as an AEC company adding a collection of optical products. It wants to own enough of the connection from SerDes and DSPs to photonic chips, transceivers, firmware and diagnostics to understand how the entire link behaves. The more of the link Credo can see, the better it should become at identifying failure modes, improving its products and winning more of the customer’s next design.

The big fundamental question is whether that creates a compounding system or merely a wider product catalogue. The mechanism strengthens if every additional product gives Credo more visibility, that visibility produces better diagnostics and designs, and customers respond by adopting more Credo components. It breaks if customers value Credo’s engineering while continuing to assemble the link from whichever individual components they prefer.

Q1 supplied the best evidence yet that the first half of this mechanism is real. It left the second half unresolved.

The Test We Set Was Incomplete

Our previous article ended with a test:

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

The logic was straightforward. AEC demonstrated that customers would pay Credo for a complete, reliable copper connection. If Credo could move into optics without sacrificing its economics, its advantage had travelled beyond the product that created the company.

Q1 appeared to pass. Revenue reached $479 million, up 115% year over year, while non-GAAP gross margin held at 68%. Credo reported record optical DSP revenue, recognized its first silicon-photonics revenue following the DustPhotonics acquisition and began production shipments of ZeroFlap optical transceivers. Management continues to expect more than $600 million of optical revenue in FY2027, with DSPs, photonic chips and ZeroFlap each contributing more than $100 million.

The problem is that consolidated gross margin cannot answer a product-level question. AEC remains Credo’s largest business, and management did not disclose optical revenue, optical mix or optical gross margin. The established business can therefore preserve the consolidated number while the economics of the emerging business remain invisible.

We chose the right question but an insufficient metric. Credo has now proved that its technology can travel into optics. It has not proved that AEC-like economic ownership travels with it.

The Layer Between Products

Management’s story begins with heterogeneity. Future AI systems will not use one architecture, protocol or physical medium. Copper will remain appropriate for shorter reaches; optics will dominate longer ones; different scale-out and scale-up connections will demand different combinations of power, latency, reach and serviceability.

“Future AI systems will combine optical and copper interconnects across different reaches, protocols, and topologies.” — Bill Brennan

That world rewards product breadth, but breadth alone is not the thesis. Broadcom, Marvell and several optical suppliers can also provide pieces of the connection. Credo’s claim is that its products are joined by a common capability: understanding whether a link is healthy, why it is degrading and what should happen before it fails.

PILOT is the layer that potentially turns a group of products into a system. Brennan compared existing link monitoring to a green light when a connection works and a red light after it fails. Credo is trying to add the yellow “check engine” light in between. Its hardware continuously measures indicators such as eye height, signal-to-noise ratio and error patterns; its software looks for deterioration and can help remove a vulnerable processor or transceiver before a link flap disrupts the cluster.

This matters because AI clusters contain tens or hundreds of thousands of interdependent connections. Management says link instability can cost more than 10% of processor utilization, while slow cluster commissioning can leave enormously expensive equipment idle for weeks. A reliable connection is therefore valuable not because the cable itself is expensive, but because everything connected to it is.

The ESD example turns that argument from a presentation into customer behaviour. Credo is learning about defects beyond the normal responsibility of a component supplier. The accumulated failure cases should influence future DSPs, photonic chips, firmware and qualification methods. More deployments can create better products, which can make the next design easier to win.

This is not an automatic data network effect. One additional link does not mechanically improve every other link, and customers may limit how operational data is shared. The advantage compounds only if learning changes product performance and product performance increases attachment. PILOT is strategically important when diagnosis repeatedly becomes design; otherwise it remains an excellent feature bundled with hardware.

Importance Is Not Ownership

The ESD request and the DustPhotonics wins are mirror images. In the first, a customer asked Credo to understand the behaviour of the entire connection. In the second, customers selected one Credo component while declining another. Q1 therefore contained both the strongest evidence for integration and the clearest warning against assuming attachment.

The DustPhotonics wins provide the quarter’s cleanest counterargument. Asked directly whether the first two major photonic-chip design wins included a Credo DSP, Brennan answered:

“These first two major design wins that we are talking about do not include the DSP.”

That is simultaneously good and incomplete news. The wins validate the acquired photonic technology independently, which suggests Credo purchased a genuinely competitive product rather than a capability that only works inside its own bundle. They also show customers retaining control over component selection. Credo has won a place in the optical module without winning the integrated DSP-and-photonic-chip combination that management says should improve performance, power, yield and cost.

The modular challenge described in our previous article therefore remains. Customers may prefer an architecture in which validated components can be combined across suppliers. Credo believes that owning more of the stack will allow it to innovate faster and deliver better system outcomes. The customer will decide how much integration is worth and how much supplier dependence it is willing to accept.

This is the distinction the next phase of the thesis must preserve: Credo may own the customer’s problem before it owns the customer’s bill of materials. Technical influence often precedes commercial capture, but the second does not inevitably follow the first.

The Cost of Becoming a System

The financial statements show the cost of this ambition earlier than the product disclosures show its reward. Revenue grew 10% sequentially, but inventory increased 25% to $313 million and receivables increased 24% to $289 million. Free cash flow was $83 million against non-GAAP net income of $236 million, a conversion rate of 35%. Non-GAAP operating expenses reached $95 million, above management’s prior range, as Credo increased research and development across a much broader portfolio.

It would be easy to treat this as an earnings-quality complaint. That misses the strategic issue. Moving from a dominant AEC product into DSPs, photonic chips, complete optical transceivers, near-package optics and new scale-up connections requires more designs, qualifications, supply commitments and inventory configurations. Credo must incur those costs before the new products reach volume.

Management is explicit about the bet. Brennan said the company was “leaning in from a supply-chain standpoint” and playing the long game. The full-year arithmetic shows the scale of that commitment. More than 85% growth on FY2026 revenue of $1.335 billion implies at least $2.47 billion in FY2027. Q1 revenue plus the midpoint of Q2 guidance produces approximately $1.01 billion in the first half, leaving more than $1.46 billion for the second half—or over $730 million per quarter, 45% above the first-half quarterly average.

Inventory is not proof that this demand exists. It is proof that management is willing to finance its forecast. If inventory converts into the promised optical ramp while gross margin remains in the high 60s and cash conversion recovers, Q1 will look like the necessary investment behind a larger company. If product attachment remains limited and inventory continues outgrowing revenue, Credo will have broadened its technical capabilities faster than its economic ownership.

The balance sheet is therefore not a side issue. It is where the cost of management’s strategy has become measurable.

What Changed in Our View

Before Q1, we believed AEC gave Credo a wedge into the customer and that optics would determine whether its reliability advantage could travel. We expected gross margin to reveal the answer.

The quarter strengthened the strategic thesis. Optical participation is no longer theoretical. PILOT has moved from a diagnostic claim to a real customer workflow. Credo’s involvement increasingly extends beyond whether its own component functions, and management is committing capital as though the second-half ramps are real.

The quarter also made the economic thesis more demanding. Consolidated gross margin cannot isolate optical economics. The first photonic-chip wins do not attach Credo DSPs. PILOT has no disclosed pricing, attachment or separately observable contribution. Working capital and operating expenses are rising before investors can see how much of the broader connection Credo will capture.

I think Credo is building a genuine system-level advantage, and Q1 increased my confidence in that conclusion. What I am not yet prepared to assume is that customers will buy the entire system simply because they increasingly depend on Credo’s understanding of it. The probability that Credo becomes strategically important across copper and optics has increased; the amount of that importance it will monetize remains the central uncertainty.

Three Futures

The scenarios should be understood as three outcomes for the same mechanism, not three predictions about the semiconductor cycle.

The share-count assumptions reflect different outcomes for dilution and Credo’s ability to offset it through future cash generation. The ranges expose the operating assumptions that matter; they are not predictions of the path the shares take between now and then.

In the bear case, customers select Credo products but assemble the link modularly; optical margins dilute AEC economics, development and inventory remain elevated, and breadth does not produce attachment. In the base case, Credo owns the portions of the link where reliability is most valuable, while component and complete-system sales coexist. In the bull case, telemetry improves product design, better products deepen customer dependence, and that dependence produces greater attachment across DSPs, photonic chips and transceivers.

The signposts follow directly. By Q3, revenue growth should overtake inventory growth and cash conversion should recover above 60%. Optical revenue should exceed $600 million for FY2027 without consolidated gross margin falling below 67%. Over the next year, Credo should show that photonic-chip wins attach its DSPs, and that PILOT influences repeatable product adoption rather than isolated customer engagements. Failure on those measures would not mean the products are weak; it would mean the proposed compounding mechanism is.

The Company Customers Call

Credo did not cause the customer’s ESD problem. That was precisely why the incident mattered. The company was valuable not because its component worked, but because it could explain why the connection might eventually fail.

That is a more ambitious role than selling a cable, DSP or photonic chip. It gives Credo the opportunity to learn across products and become embedded in how customers design, qualify and operate their networks. It also requires Credo to carry more products, inventory and development expense before that opportunity becomes revenue.

The previous chapter of the Credo story asked whether its advantage could travel from copper into light. Q1 suggests that its technical relevance can travel. What remains unanswered is whether product attachment, margins and cash returns will travel with it.

Credo is learning to understand more of the link. The next question is how much of the link it will own.

General Disclaimer: The information presented in this communication reflects the views of the author and does not necessarily represent the views of any other individual or organization. It is provided for informational purposes only and should not be construed as investment advice, a recommendation, an offer to sell, or a solicitation to buy any securities or financial products.

While the information is believed to be obtained from reliable sources, its accuracy, completeness, or timeliness cannot be guaranteed. No representation or warranty, express or implied, is made regarding the fairness or reliability of the information presented. Any opinions or estimates are subject to change without notice.

Past performance is not a reliable indicator of future performance. All investments carry risk, including the potential loss of principal. This communication does not consider the specific investment objectives, financial situation, or particular needs of any individual.

The author and any associated parties disclaim any liability for any direct or consequential loss arising from the use of this material and undertake no obligation to update or revise it.

Subscribe now

• Platformization is beginning to look like a genuine capability: CyberArk and Chronosphere are benefiting from Palo Alto’s distribution, cross-selling and integration, while XSIAM and Prisma AIRS add further evidence that the customer loop is strengthening.

• But acquisitions have erased the organic baseline: headline NGS ARR and revenue growth increasingly mix acquired and internally generated growth, making it harder to know how much economic value Palo Alto is creating versus purchasing.

• The decisive test is now per-share economics: gross margins are slipping, stock compensation remains substantial and dilution is absorbing a meaningful part of earnings growth. Palo Alto needs conventional free cash flow per diluted share to compound before the customer loop can truly become a shareholder loop.

“M&A is not a strategy. M&A is a consequence of stuff that we do from a product development perspective.”

— Nikesh Arora, Q4 FY2026 earnings call

Palo Alto Networks ended fiscal 2026 with what management described as a record finish. Revenue grew 34% to $3.41 billion, Next-Generation Security ARR rose 63% to $9.1 billion, remaining performance obligations reached $21.2 billion, and the company added almost $1 billion of net new NGS ARR in one quarter. The shares fell roughly 7%, continuing a pattern in which Palo Alto wins the operating quarter while investors question the economics underneath it.

This is the fifth article in our Palo Alto series, and I think the disagreement has finally become clear. Management is selling a customer loop: AI makes attacks faster and identities more numerous, fragmented security becomes untenable, and each additional Palo Alto control point supplies more data and makes the remaining products easier to sell. The stock depends on a different loop, in which revenue becomes gross profit, gross profit becomes cash, and cash grows faster than the diluted share count.

Q4 is the strongest evidence yet that the customer loop is working. It is not yet evidence that the shareholder loop has closed. The big fundamental question is therefore: can Palo Alto convert a strengthening customer loop into a strengthening shareholder loop before Microsoft commoditizes the value its breadth creates?

The Customer Loop Starts to Close

Our first articles asked whether Palo Alto could move from gate to graph: from guarding a perimeter to observing identity, network, endpoint, cloud and application activity as one connected system. CyberArk and Chronosphere then turned this into the assembly test. Palo Alto was trying to coordinate products with different architectures, buyers and sales motions before AI made machine-speed control compulsory.

The previous article ended with the question:

“Can assembly become architecture before AI makes real-time control mandatory?”

Q4 does not prove that the architecture is complete, but it changes the burden of proof. CyberArk produced more than 400 shared leads and more than 200 new identity logos from Palo Alto’s installed base. Deals worth at least $5 million increased 50%, CyberArk’s margin improved by more than 1,000 basis points in two quarters, and integration synergies were running three to six months ahead of plan. Palo Alto is not merely absorbing the asset; its distribution is changing the asset’s growth and cost structure.

Chronosphere supplies similar evidence. Observability ARR has passed $500 million after more than doubling since the acquisition closed, while XSIAM relationships generated about half of its net new logos in Q4. XSIAM itself ended the year above $700 million of ARR, growing 70%, and Prisma AIRS crossed $100 million within four quarters of general availability. Platformizations reached roughly 2,500 after a record 220 net additions, with net retention above 120% for the platformized cohort.

These numbers reveal the mechanism management wants investors to see. More control points create more telemetry; more telemetry improves detection and automated response; better outcomes encourage consolidation; consolidation gives Palo Alto another surface through which to distribute the next product. In that sense, integration is becoming a capability rather than a sequence of transactions.

Commercial integration is not the same as technical unity. Management acknowledged that Chronosphere still needs about six months of development before competing broadly for conventional enterprise workloads, while customer transformations can take one to three years. Q4 shows that Palo Alto can distribute acquired products faster than we expected; it does not yet show that their data models, policies and workflows operate as one architecture.

Success Has Erased the Baseline

The awkward consequence of this strategy is that better integration produces worse visibility. Reported NGS ARR growth of 63%, revenue growth of 34% and RPO growth of 34% include the acquired identity and observability businesses. Chronosphere’s net new ARR also contained a nine-figure migration by one large language-model customer, a contribution management said will diminish after the first quarter of fiscal 2027.

The new category disclosures help, but do not solve the problem. Network and AI Security generated $8.35 billion of fiscal 2026 revenue and grew 17%, although that includes roughly $85 million from an acquired certificate-management business. Cortex grew 25% to $1.92 billion but now contains Chronosphere, while Idira is almost entirely acquired CyberArk revenue. Investors still cannot construct a clean bridge from the pre-acquisition Palo Alto business to the company that exists today.

This is not an accounting complaint. It is the central analytical problem. The better Palo Alto becomes at integration, the harder it becomes to distinguish value the company created from value it purchased. The headline numbers prove scale and commercial momentum; they do not establish the organic growth rate on which a premium valuation must eventually rest.

The balance sheet offers less cause for alarm than the absolute figures suggest. Approximate year-end DSO improved from about 106 days to 97 days as receivables grew more slowly than revenue, while financing receivables declined from roughly $1.72 billion to $1.54 billion. Revenue quality remains difficult to interpret because of acquisitions and customer financing, but the observable year-over-year trend did not deteriorate.

The Missing Premium

Microsoft was a central risk in our earlier work, and it should remain one. We wrote after Q1:

“Microsoft is security as a feature of the productivity suite you already own. It does not need to be the best. It needs to be good enough and already installed.”

Microsoft does not need to displace Palo Alto at the largest enterprises to matter. Entra, Defender, Sentinel and Azure Monitor give it enough breadth to cap pricing, particularly where a CFO values bundling more than a security team values technical superiority. Palo Alto can win consolidation and still fail to capture all the economics of the value it creates.

That is why Q4 gross margin deserves more attention than we initially gave it. Gross margin declined 100 basis points to 74.8%, while the full-year figure fell 60 basis points to 75.8%. Management expects cloud-hosting costs to grow faster than revenue in fiscal 2027 and also cited hardware component inflation. Non-GAAP operating margin is guided to improve only 30 basis points to 29.5%, while adjusted free cash flow margin slips from 38.4% to 38% before the company attempts to reach 40% in fiscal 2028.

The bearish interpretation is that the enlarged suite wins through breadth but cannot command a premium because Microsoft anchors the price. The benign interpretation is that faster-growing SaaS and acquired products have not reached gross-margin maturity, after which scale and integration synergies restore leverage. Gross margin cannot distinguish those explanations by itself; pricing pressure must also appear in renewal economics, discounting, competitive win rates and organic growth.

Management emphasized AI demand, large transactions and the $20 billion FY2030 NGS ARR target. Analysts received that destination constructively, but their questions kept returning to CyberArk penetration, Chronosphere’s durability, AI-driven firewall demand and the conversion of cross-selling into revenue. The market’s skepticism is therefore broader than one margin line: it is asking whether expansion in corporate scale will become superior economics per share.

A Better Company Must Become a Better Share

Fiscal 2026 generated $4.11 billion of conventional free cash flow and $4.41 billion of company-adjusted free cash flow. The difference matters because the adjusted figure adds back acquisition payments and selected capital expenditures, while neither measure deducts the $1.71 billion of stock compensation, equal to 14.9% of revenue. Acquired-intangible amortization added another $638 million to the gap between GAAP and non-GAAP earnings.

Dilution turns this from an accounting debate into an ownership issue. Full-year non-GAAP diluted shares increased 8.8% to 764 million, and the Q4 count increased 17.7% to 832 million. At the midpoint of fiscal 2027 guidance, non-GAAP net income rises roughly 20%, but EPS rises only about 9% because the guided share count reaches approximately 846 million.

Gross margin is the leading warning; conventional free cash flow per diluted share is the final test. It captures whether revenue growth survives hosting costs, integration spending, stock issuance and recurring capital requirements before reaching the owner. If that number compounds, the customer loop is becoming a shareholder loop. If it stalls while NGS ARR races toward $20 billion, Palo Alto will have built a more important company without necessarily building a better investment.

Three Prices, Three Worlds

At $355.21, Palo Alto trades near 85 times the midpoint of fiscal 2027 non-GAAP EPS guidance, while guided adjusted free cash flow implies a forward yield below 2%. The following scenarios value fiscal 2029 conventional free cash flow rather than the more flattering adjusted measure, and the arithmetic incorporates dilution explicitly.

The probability-weighted value is approximately $382, implying only a low-single-digit annual return from the current price. The decisive assumption is not whether revenue grows 19% rather than 17% next year; it is whether Palo Alto can sustain a high conventional cash margin and premium multiple after acquired growth annualizes. Existing holders can justify holding for the strategic optionality, but the expected return does not support adding at $355. A price of $320–$330 would merit reassessment, while a genuine margin of safety appears closer to $290–$300 unless the evidence improves first.

Two Loops, One Stock

Q4 makes us more constructive about Palo Alto’s organizational capability. The company is integrating acquisitions faster, distributing them more effectively and using AI demand to make its breadth increasingly relevant. The assembly test has moved in Palo Alto’s favor, although technical unification remains unfinished.

The quarter does not prove that Palo Alto can retain the economic premium created by that breadth. Microsoft can constrain pricing, cloud costs can absorb scale benefits, acquisitions can erase the organic baseline, and dilution can divert corporate growth away from each share. The thesis now turns on whether gross margin stabilizes and conventional free cash flow per diluted share begins compounding as acquired growth annualizes.

Perhaps Arora is right that M&A is not a strategy. The strategy is to make fragmented security untenable, while the capability is to build or buy each missing control point and distribute it through an expanding customer system. Q4 shows that the customer system is beginning to compound; at $355, the stock already assumes that shareholders will receive the same benefit, which is why the right action remains to hold rather than add.

Subscribe now

My first reaction is that the market traded the right question and reached too clean an answer. The quarter was considerably better than the stock reaction suggests, but the metric Rubrik introduced to explain that strength was also unusually convenient. That makes the disclosure change not Glasswing, Agent Cloud or even the headline beat the most important development in the print.

Our Rubrik articles have followed a changing question. We began by asking whether the company was building something larger than backup. Q3 FY26 described the janitor becoming the bodyguard. Q4 called Rubrik’s time-indexed record of the enterprise “state infrastructure.” Q1 FY27 asked whether that record could extend from data into identity and autonomous agents.

The Question is now more precise: Can Rubrik turn its time-indexed recovery record into the default control plane for data, identity and agentic actions, so that every additional protected surface increase adoption, recovery confidence and switching costs faster than migration benefits disappear and native alternatives commoditize recovery?

The Scoreboard Graduated or Moved

Rubrik ended Q2 with $1.661 billion of subscription ARR, up 33%, while net-new subscription ARR increased 35% to approximately $96 million. Cloud ARR reached $1.480 billion, up 39%, but calculated net-new Cloud ARR was approximately $89 million, down 4%. The new metric adjusted net-new Cloud ARR, excluding migrations by existing maintenance and non-cloud customers was $86 million, up 20%.

Both measures are valid. Migrations moved existing economics between Rubrik categories without creating new demand, so removing them produces a cleaner organic measure. Yet companies rarely introduce adjusted metrics when the old measure is becoming less flattering. Rubrik introduced this one when the reported measure turned negative and management simultaneously redirected investors toward total subscription ARR.

That redirection can be an economic graduation. Cloud now represents 89% of subscription ARR, and private or sovereign deployments need not be strategically inferior if they carry similar recurring economics. It can also become an escape hatch through which weakening cloud demand disappears into a broader number. The correct response is to accept the adjustment while requiring Rubrik to publish reported Cloud ARR, adjusted Cloud ARR and the migration bridge until the last legacy migration has passed.

Before extending the thesis, we should score the promises embedded in the previous article:

Sources: prior Rubrik Q1 FY27 article; Rubrik Q2 FY27 release; Q2 earnings-call transcript.

The scorecard says the recovery engine strengthened while the proposed second and third engines remained unproven. Our Q1 conclusion still holds:

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.

Q2 advanced the last sentence only slightly. That is meaningful progress for a product, but not enough evidence for a valuation category.

The cash-flow story is better than the quarterly margin suggests, but less simple than blaming Flex commissions. Q2 FCF rose 14% to $65.7 million while the margin fell four points. Across the first half, FCF increased 53% to $139.3 million and margin expanded to 17.1% from 15.4%. Cash invested in deferred commissions rose 40% to $69.1 million, but amortization also rose to $59.5 million; the deferred-commission asset increased only 3.6% from January. There is no disclosed evidence tying that increase specifically to Flex, and amortization does not reverse the original cash outflow. The raised $328 million FY27 FCF midpoint argues against structural deterioration. The more durable earnings-quality concern is $101 million of Q2 stock compensation 23.6% of revenue which separates corporate cash generation from per-share economics.

What Actually Compounds

Bipul Sinha describes Rubrik’s products as creating “complementary network effects.” I think that claim is too generous. One company adopting Identity does not improve another company’s recovery. What Rubrik may possess is customer-level data gravity: each protected workload, SaaS application and identity system makes the historical record more complete; a more complete record makes recovery more useful; broader recovery encourages another product purchase; and every added integration makes displacement more disruptive.

The economics are compatible with this mechanism. Net retention remains above 119%, customers spending more than $1 million grew 57%, and subscription ARR contribution margin reached 14%, up 460 basis points. Those facts prove expansion and operating leverage. They do not prove that customers adopting Identity or Agent Cloud expand faster, renew better or cost less to serve than recovery-only customers. The missing evidence is product attachment and cohort NRR.

Q2 supplied three different kinds of evidence that should not be confused. The largest international Identity deal shows enterprise-scale demand, but the missing ARR update prevents us from measuring the second engine. Project Hourglass gives Agent Cloud potential distribution through five global integrators, but a go-to-market alliance is not converted pipeline. Project Glasswing validates Rubrik’s relevance to emerging AI-security research, but Anthropic describes a broad program spanning vulnerability detection, testing and patching with numerous security partners; it does not validate Rubrik’s commercial moat or recovery architecture.

Agent Rewind remains the most interesting and least settled extension. Restoring a file or directory to a known state is bounded. Reversing an agent is not. An email may already have been sent, a payment executed, a permission propagated or an external system changed. Rubrik’s historical record can establish what happened and help orchestrate recovery, but complete reversal requires authority across systems it does not control. The architecture is coherent; the product boundary is wider than the metaphor suggests.

A Complement Is Not Comparable

The after-hours decline was not a verdict that Rubrik’s business deteriorated. It was a repricing of trajectory after a strong run into the print. NNARR fell from $115 million in Q4 to $103 million in Q1 and $96 million in Q2. Management’s full-year ARR guide implies roughly $101 million in Q3 and $121 million in Q4. CFO Kiran Choudary answered the concern by reframing the annual comparison:

“The net new ARR growth rate implied by the new guidance … went from 7% to 14%.”

That is true, but the market focused on how much acceleration is still required. Management has raised the destination; the burden of reaching it has moved into Q4.

CrowdStrike is useful for understanding the job, not valuing the stock. Detection at machine speed and recovery at machine speed are complements, which is why both companies can participate in Glasswing without being substitutes. That does not mean the market is mistakenly denying Rubrik a CrowdStrike multiple. CrowdStrike has greater scale, cash flow, installed-base evidence and demonstrated product attachment. Rubrik should earn its own valuation through duration and cohort economics, not inherit one from a neighboring category.

One execution risk deserves attention, although it is not a new Q2 revelation. Both Q1 and Q2 releases identify Rubrik’s ability to manage a leadership transition in its global revenue organization; Q1 explicitly referenced organizational changes and their effect on growth. Management has not explained the transition on either call. That silence matters because the H2 guide requires reacceleration, but the repeated language means this is a carried risk rather than a newly added legal warning.

Valuing What Is Proven

At approximately $96 after hours, Rubrik’s equity value is about $21.9 billion using the FY27 diluted-share guide of 228 million. Because that share count includes the potential dilution from converting Rubrik’s notes, subtracting $1.75 billion of cash and investments produces an enterprise value of roughly $20.1 billion, or 11.9 times FY27 revenue guidance. That valuation already assumes growth lasts.

The midpoint probability-weighted value is approximately $118, implying only about 7% annualized over three years from the post-print price. The bear case is not $30 because even a slower Rubrik would produce meaningful recurring cash and retain net cash. The previous $135–160 base range was too generous because it combined high terminal revenue multiples with insufficiently explicit dilution and cash assumptions. At approximately $85, the same probability-weighted value offers an 11% annualized return; near $96, the stock requires new evidence rather than another narrative extension.

What Changed in Our View

Q2 increased my conviction that Rubrik owns a valuable recovery franchise with improving economics. It did not increase my conviction by the same amount that Identity and Agent Cloud form a self-reinforcing second act. The variant perception is therefore narrower than before: the market still underestimates the strategic value of a trusted historical record, while bulls overestimate how much of the wider control plane has already been commercialized.

The next quarter has four tests. NNARR at or above $100 million keeps the annual guide credible; below $95 million leaves an uncomfortable Q4 burden. Rubrik should preserve the adjusted-cloud bridge and keep organic growth near 20%. By Q4, Identity needs renewed quantification and Agent Cloud needs either 30-plus paying customers or meaningful ARR. During FY28, stock compensation must fall below 20% of revenue and management must show that the revenue-organization transition has not weakened productivity.

Rubrik has proven that enterprises will pay for a trusted record of their past. It has not yet proven that this record gives it the right to govern their future. Existing holders can remain patient, but new money should wait for approximately $85 or for product-level disclosure that proves the compounding loop. The recovery engine cleared the bar; the valuation already assumes that the rest of the map eventually will.

Subscribe now

• Q2 looked structural, not merely strong: net-new ARR reached $333 million, up 51%, while revenue, operating income and free cash flow all accelerated.

• The business model may be improving: Falcon acts as technical distribution for new products, while Falcon Flex increasingly acts as commercial distribution, letting customers add capabilities without restarting deployment or procurement.

• The valuation depends on postponing maturity: CrowdStrike does not need to grow 25% forever. It needs new products like AIDR and SIEM to become meaningful before older growth vectors fade, continually pushing the visible maturity point further out.

“Inflection has become acceleration.” — George Kurtz, CrowdStrike Q2 FY27 earnings call.

There are earnings beats, and then there are quarters that change the question investors should be asking.

CrowdStrike’s fiscal second quarter was the latter. Net-new ARR reached $333 million, up 51% year over year, versus roughly $286 million embedded in published expectations. Revenue increased 26% to $1.47 billion, ending ARR reached $5.84 billion, non-GAAP operating income rose 46% to $372 million, and free cash flow increased 33% to $377 million. CrowdStrike also lifted its FY27 net-new ARR growth expectation to 34%, 1,150 basis points above where management began the year.

Those numbers explain why the stock went up. They do not explain why the quarter matters.

Before earnings, the important question was whether AI, Falcon Flex and CrowdStrike’s broader product portfolio were changing the shape of the business or merely giving management better stories to tell around an already excellent cybersecurity franchise. The market did not need another routine beat. It needed evidence that something structural had changed.

Q2 provided the first convincing evidence.

The Question is now this: can CrowdStrike turn the proliferation of AI agents into a compounding security cycle in which its existing endpoint distribution makes new products easier to deploy, Falcon Flex makes them easier to buy, and each additional workload broadens the customer relationship and strengthens CrowdStrike’s data advantage?

If so, the importance of Q2 is not that growth accelerated once. It is that CrowdStrike may have developed a mechanism capable of repeatedly postponing maturity.

The Agent Is Already There

A reasonable concern about AI is that it should make traditional endpoint security less strategically important. Intelligence increasingly lives in frontier models and cloud infrastructure; perhaps security therefore shifts toward OpenAI, Anthropic, Microsoft, Google and AWS rather than remaining with companies built around protecting PCs and servers.

The better way to think about AI security, though, is as several overlapping control layers. Model providers will secure models and inference. Hyperscalers will secure infrastructure and cloud workloads. But agents eventually must act: they access identities, files, credentials, applications and enterprise data. CrowdStrike does not have to monopolize AI security. It needs the layer where AI activity becomes observable and enforceable to remain sufficiently important that more AI makes Falcon more valuable.

The Q2 evidence supports that argument. CrowdStrike said that, within a sample of customers, Claude usage on endpoints had increased more than 400% in recent months and custom-agent usage more than 100%. Meanwhile the endpoint business accelerated for a fourth consecutive quarter.

This matters because Falcon is already installed. CrowdStrike already receives telemetry from the customer environment and already has an enforcement point. A new product does not necessarily require another agent, another deployment project or another technical architecture.

That was why one exchange in the earnings call mattered more than most of the AI rhetoric. Raymond James analyst Adam Tindle asked whether AI Detection and Response was genuinely incremental to endpoint security. George Kurtz answered:

“It is separate, and incremental. It’s another module, and it’s priced separately.”

AIDR therefore sits on the same Falcon agent but creates another entitlement and another revenue stream. The technical footprint already exists; CrowdStrike can monetize another problem without recreating distribution from scratch. AIDR ending ARR nearly tripled sequentially in Q2.

The same mechanism is increasingly apparent in SIEM. Next-generation SIEM exceeded $695 million of ARR and grew 60%, but the architectural advantage matters more than the number. CrowdStrike’s own endpoint telemetry is already in Falcon, which means existing customers can activate SIEM without beginning from an empty data environment. Kurtz described the product as unusually quick to deploy precisely because CrowdStrike already owns the first-party data.

Cloud Security is above $905 million of ARR, Identity above $585 million, and endpoint itself is still accelerating.

The Falcon agent, then, is not simply a product. It is distribution.

That solves one half of the problem.

Falcon Flex Solves the Other Half

Technical distribution does not eliminate enterprise software’s commercial friction. A customer can already trust a vendor and still spend months navigating budget cycles, procurement, legal review and contract negotiations before adding another capability.

Falcon Flex increasingly appears to attack this second problem.

Instead of repeatedly buying discrete modules, customers commit a larger pool of spending to CrowdStrike and consume additional Falcon products over time. CrowdStrike consequently gains something valuable: a commercial channel through which its next product can be monetized without restarting the purchasing process from zero.

Q2 was the first quarter in which the scale of that mechanism became difficult to dismiss. CrowdStrike added more than 935 Flex accounts, more than during the previous three quarters combined. Its ten largest deals were all Flex transactions. Flex new-logo ARR represented 34% of total Q2 net-new ARR, while ARR associated with Flex customers exceeded $2.29 billion and grew 101%.

The more important evidence is what customers do afterward. Standard subscription customers moving into Flex increase ARR by more than 40% on average. Customers completing their first ReFlex subsequently expand ARR by another 25%, while customers that have ReFlexed at least twice are running 53% above their original Flex commitment. The first ReFlex occurs after roughly eight months.

That eight-month figure matters because it addresses the obvious bear case.

Perhaps Flex merely pulls demand forward. Customers sign larger commitments today, CrowdStrike records more ARR, and future expansion disappears because customers eventually must consume what they already bought.

ReFlex is the test. If customers return after only eight months and voluntarily increase commitments again, then consumption appears to be catching contractual capacity rather than lagging it.

The loop becomes visible: Falcon is already deployed; deployment produces telemetry; telemetry supports new products; those products are easy to activate technically; Flex makes them easier to consume commercially; customers use more modules and ReFlex into larger commitments; the larger relationship creates more workloads, more data and more opportunities for the next product.

AIDR may become a very large product. Flex may matter more because it can make every future product larger faster.

This is the potentially important change in CrowdStrike’s business model. The Falcon agent lowers deployment friction. Falcon Flex lowers purchasing friction. AI increases the number of problems CrowdStrike can sell through both.

What We Underestimated

Our pre-earnings framework correctly identified net-new ARR as the metric that mattered, and roughly $300 million as the threshold at which the quarter would begin to look like genuine acceleration rather than another conventional beat. What it underestimated badly was the magnitude: our central expectation was closer to the high-$290 millions, not $333 million.

That mistake is useful. We anchored too heavily on the previous quarter’s beat pattern and not enough on the possibility that the demand regime itself had changed. Q2 did not simply clear a difficult expectation. It forced the expectation framework upward.

That is why the next question is no longer whether CrowdStrike can produce a $300 million quarter.

It is whether CrowdStrike can make $333 million look ordinary.

Management formally guided Q3 net-new ARR to $343–347 million and said it entered the quarter with record pipeline. After a quarter in which the company beat ARR guidance by more than $45 million, though, the formal guide is unlikely to remain the actual market hurdle.

Before Q2, the regime was proving the acceleration. After Q2, it is prove the compounding.

The Moving Horizon

This is also the only sensible way to think about CrowdStrike’s extraordinary valuation.

CrowdStrike has spent most of its public life trading at multiples that look unreasonable by conventional software standards. Simply assuming those multiples must eventually revert to something “normal” has repeatedly been a poor investment framework.

The more useful question is what allows an abnormal multiple to persist.

CrowdStrike does not need investors to believe it will grow 25% forever. It needs investors to keep postponing the year in which they expect growth to become ordinary.

Suppose investors once imagined CrowdStrike fading from the mid-20s toward 20%, then into the teens. A quarter like Q2 changes more than the current year. It introduces the possibility that the whole deceleration curve begins later because another growth vector has appeared before the old one fully matured.

Kurtz made the boldest possible version of this claim when JPMorgan asked about AI-security durability:

“I think the AIDR business can be bigger than the EDR business, just given the pure number of agents which each person will have.”

That should not be treated as a forecast. CrowdStrike has not even disclosed absolute AIDR ARR, and “nearly tripled” tells us little about ultimate scale without the denominator.

But it states management’s duration bet explicitly.

CrowdStrike was built on EDR. If a newly emerging product can someday rival the business that built the company, then the maturity point moves dramatically further away. If SIEM scales toward another billion-dollar-plus franchise while that happens, it moves again. If Flex lets subsequent products monetize faster, it can move again after that.

This is why Q2’s valuation significance is larger than its revenue beat.

The multiple survives not because CrowdStrike grows into it, but because CrowdStrike continually postpones the moment when investors feel compelled to normalize it.

That is also what makes the stock dangerous.

The multiple probably does not collapse because investors suddenly notice that it is high. Everyone already knows it is high. It collapses when investors can finally see maturity.

If AIDR remains small, the horizon moves closer. If customers take 15 months rather than eight to ReFlex, the horizon moves closer. If SIEM decelerates before another product becomes material, the horizon moves closer. If endpoint growth turns down and FY28 or FY29 estimates stop moving upward, investors will eventually have to answer the question CrowdStrike’s execution has repeatedly allowed them to defer:

What is a mature CrowdStrike worth?

That is when an extraordinary multiple becomes a source of extraordinary downside.

Q2 Did Not Make CrowdStrike Cheap

Q2 did not make CrowdStrike cheap. It made the extraordinary valuation more intelligible.

The quarter gave us strong evidence that Falcon is technical distribution, Falcon Flex is becoming commercial distribution, and AI is creating incremental workloads that can be sold through both. AIDR is separately monetized. Flex customers are increasing commitments. ReFlex is happening quickly. Endpoint is reaccelerating. SIEM, cloud and identity are scaling. Growth is occurring alongside operating leverage.

What Q2 did not prove is that these mechanisms will continue producing new growth vectors before existing ones mature.

That is now the investment case.

The Question and the valuation question are ultimately the same question viewed from opposite sides. The business question is whether CrowdStrike can repeatedly turn new security problems into products that become easier to deploy and easier to buy as the customer relationship deepens. The valuation question is whether it can do so frequently enough that visible maturity always remains several years away.

Before Q2, CrowdStrike needed to prove that AI could create another acceleration.

After Q2, the harder and considerably more valuable question is whether the Falcon agent and Falcon Flex can keep turning each new security problem into another reason that maturity remains just beyond the horizon.

If they can, the multiple can continue looking irrational for years.

If they cannot, the multiple will be the first thing investors discover was rational only while the horizon kept moving.

Subscribe now

• The interface may no longer be the moat: as Claude, ChatGPT, Slack and custom apps increasingly own the user experience, Salesforce’s value shifts toward governed context, permissions, semantics and trusted execution.

• AI may create a second user rather than replace the first: human usage remains resilient while machine access through MCP is scaling rapidly, creating a new consumption layer alongside seats.

• Q2 finally connected the AI story to the core business: cRPO accelerated to 14% constant currency and production/refill activity strengthened, but Salesforce still needs to prove implementation intensity, and AI economics can scale without eroding margins.

The most interesting Salesforce customer right now may be Anthropic.

Claude is precisely the sort of product that is supposed to threaten Salesforce. If an intelligent model can answer questions, build interfaces, write software, and take actions, why should a salesperson need to open an expensive CRM application at all? The strongest version of the “SaaS apocalypse” argument is not that software disappears. It is that intelligence moves into the model, the interface moves into the chatbot, and the application sitting between the employee and the data loses the user.

And yet Salesforce disclosed something peculiar on its first-quarter earnings call. Chief Revenue Officer Miguel Milano described what was happening at Anthropic:

“Anthropic is one of our biggest users of CRM, of Sales Cloud, and obviously Slack. Their usage through Q1 has exploded fivefold because now they’re using Sales Cloud from a headless perspective… Sales Cloud has become more prominent and more strategic for them than ever because of Headless.”

Anthropic was using less of the Salesforce interface and more Salesforce.

I think that paradox explains both why Salesforce’s second quarter mattered and why my own view of the company has changed.

When Less Salesforce Means More Salesforce

Last September, I argued that Salesforce had fallen into a complexity trap. The company that displaced Siebel by removing much of the pain from enterprise CRM had spent two decades accumulating Sales Cloud, Service Cloud, Marketing Cloud, Tableau, MuleSoft, Slack, industry clouds, Data 360, and eventually Informatica. Each acquisition increased what Salesforce could do while adding another set of schemas, permissions, workflows, interfaces, and integrations that had to work together.

AI appeared likely to expose that fragmentation. An autonomous agent cannot casually work around conflicting definitions of “customer,” uncertain permissions, or disconnected workflows in the way an experienced employee can. The agent needs context it can trust.

That concern has not disappeared. What I underestimated was the possibility that the user itself would change.

Complexity has two costs. One is the cost of constructing and integrating the system. The other is the cost imposed on the person who has to navigate it. Salesforce has not yet demonstrated that AI eliminates the first. Large enterprise deployments still require implementation work, customer data remains scattered across systems, and management still does not disclose the deployment intensity or AI unit economics that would let investors determine whether the underlying architecture is becoming materially cheaper.

AI can attack the second cost much faster.

A salesperson confronting Salesforce’s accumulated architecture sees objects, fields, dashboards, approval flows, permissions, and applications. An agent can see something else: who owns the account, what a field means, which records matter, what happened previously, which business rules apply, and which actions are permitted.

What looks like complexity to a human can look like context to an agent.

What the Agent Sees

This distinction matters because large language models are making interfaces abundant. Claude can provide one. ChatGPT can provide another. Slack can become another. A company can increasingly build its own.

What does not become abundant nearly as quickly is enterprise state.

A model can reason that a discount probably makes sense. It cannot independently know that a particular quote requires two approvers, that the employee requesting it lacks authorization, that the customer has a contractual restriction, or which system must record the final action. Those are problems of identity, semantics, permissions, governance, workflow, and deterministic execution.

Salesforce has spent 25 years accumulating precisely those things.

Benioff described the strategic change unusually directly on the Q2 call:

“People don’t have to come to Salesforce to get work done. Salesforce comes to them in Claude, in Slack, in ChatGPT, in Teams, wherever they want to work. It is a shift from software as the interface to software powering every interface.”

That is more consequential than another Agentforce feature. Salesforce is explicitly accepting that somebody else may own the interface.

This also changes how I think about our previous concern around federation. I had treated Salesforce’s embrace of zero-copy access to data residing in Snowflake, Databricks, SAP, and elsewhere as evidence that Salesforce had ceded control of the most valuable layer. That now seems too categorical. An agent may care less about where the bytes physically reside than whether Salesforce can explain what those bytes mean, whether they can be trusted, who can access them, and what can safely be done with them.

The moat may be moving from owning the interface to owning governed context and action.

There is an additional irony. Salesforce spent years building its own AI capabilities through Einstein and Agentforce. Now one of its most strategically interesting propositions is that Claude can become the interface to Salesforce. That could be interpreted as defeat. I increasingly think the opposite is true. If Salesforce needs to own the best model for this thesis to work, the thesis is fragile. A genuinely valuable governance and execution layer should benefit whether Claude wins, OpenAI wins, Gemini wins, or model intelligence becomes increasingly commoditized.

The Second User

The architectural change matters to investors only if it changes the economics.

Salesforce laid out the mechanism in Q1. Existing users could migrate to premium editions. AI could make Salesforce economical for additional users. Customer-facing agents could consume Flex Credits. Headless could eventually monetize interactions initiated by third-party agents outside the Salesforce application itself.

The old Salesforce monetized people.

The emerging Salesforce may monetize people and machines.

Salesforce’s Q2 presentation says human application usage remained strong while weekly calls to its MCP server increased approximately sixfold. It is a management-provided chart rather than a formally disclosed usage series, so the human side should not be over-interpreted. Still, the direction is important: machine activity appears to be developing alongside the existing human franchise rather than simply replacing it.

The Q2 transcript gives us a better economic indicator. When Morgan Stanley’s Elizabeth Porter asked how much Agentforce activity was moving beyond pilots, Milano answered:

“50% of the bookings came from customers refilling the tank. So they consume, they use the Flex Credits, they want more… We added 2,000 paying customers into production. That is 70% more quarter on quarter.”

This is more useful than the headline Agentforce ARR number.

ARR proves Salesforce persuaded someone to buy an AI product. Production proves the implementation escaped procurement. Consumption proves customers used it. Refill proves they used enough of it to pay Salesforce again.

The refill itself is not new; similar behavior was visible in Q1. What matters is that it persisted while production deployments and Agentic Work Units scaled rapidly.

And the machine user does not yet appear to be replacing the human one. Sales and Service seats are still growing, while management says penetration of its premium editions remains remarkably low:

“Only 5% of the knowledge workers that use sales and service have upgraded to the higher-end editions. We get a 60% to 80% premium.”

That suggests a considerably more attractive economic possibility than the original SaaS bear case:

existing seats + premium upgrades + machine consumption.

The threat was that AI would reduce Salesforce’s number of users.

The opportunity is that AI creates an entirely new class of Salesforce user.

When Agentforce Became Salesforce

None of this would matter if it remained confined to Agentforce statistics.

Morgan Stanley’s Keith Weiss made exactly that point on the Q1 call. Agentforce usage looked terrific, but cRPO had merely met expectations; Tableau and Commerce were weak; where was the broader business acceleration?

It was the right question, and it echoed the questions we had been asking for nearly a year.

In December, Salesforce was increasing sales capacity by more than 20% while guiding to single-digit growth. I argued that either management had abandoned the discipline investors had spent years demanding or it could see demand that had not yet reached reported revenue.

In February, Agentforce ARR, premium bookings, consumption, and large deals all improved, but aggregate organic growth did not. The distinction I made then was between progress and proof.

Q1 explained the mechanism. Q2 is where the mechanism began transmitting into Salesforce.

Revenue was $11.35 billion, although $456 million came from Informatica, while reported non-GAAP EPS of $5.90 included $2.53 per share of strategic investment gains. Neither explains why the quarter mattered. cRPO does.

Current remaining performance obligations grew 14% in constant currency, up from 13% in Q1, while Salesforce guided Q3 to another approximately 14%, excluding Contentful and Fin. Management also said net-new annual order value growth was the strongest in four years.

Milano then made the accountability test unusually explicit:

“We are very confident. Definitely, we are committed to the H2 revenue re-acceleration. That’s already math. Q3 is math. Q4 is nearly math.”

Three months earlier, investors were being asked to trust a forecast. By Q2, management was describing the next quarter as arithmetic. That does not prove the acceleration will persist, but it tells us why this quarter is different: more of the Agentforce story has migrated from usage into signed contracts and therefore future recognized revenue.

Q2 matters not because Agentforce became more impressive. It matters because Agentforce finally began becoming Salesforce.

The Anchor

The new engine still must pull a very large company.

Under Salesforce’s legacy disclosure, Sales grew 9% in constant currency in Q2, Service 5%, Marketing and Commerce declined 1%, and Integration and Analytics declined 6%. Platform, Slack and Other grew 43%.

That changes the strongest bear case. It is no longer that Agentforce is fake. It is that Agentforce works but never becomes large enough to bend Salesforce.

The complexity question also remains open. PenFed described Salesforce engineers and architects working directly alongside its team, while UCLA said its first customer-facing Agentforce use case took eight months to deploy.

Our December capacity call therefore looks increasingly prescient, as does the focus on refills. The Informatica semantic-layer hypothesis has gained credibility, but I would not declare it proven: we still cannot isolate Informatica as the reason consumption or deployment economics are improving.

Until Salesforce shows declining implementation intensity, strong AI gross margins, and sustained organic revenue acceleration, the Complexity Inversion remains a thesis rather than a conclusion.

The Price of Being Right

At roughly the post-earnings price, I would now assign a 15% probability to a bear case in which Agentforce succeeds but largely offsets legacy deterioration. Revenue grows 6–8% annually through FY30, non-GAAP operating margins settle around 30–32%, and earnings of roughly $16–18 support a stock around $210–250.

My 55% base case assumes the second user becomes economically meaningful but develops progressively. Revenue grows around 10–12%, cRPO stays in the low-to-mid teens, premium upgrades and consumption remain additive to seats, and margins hold around 34–36%. FY30 EPS around $22–23 at 17–18 times earnings implies roughly $375–415.

The 30% bull case is the Complexity Inversion fully expressed. Third-party agents dramatically expand Salesforce activity; premium penetration moves far beyond today’s 5%; Headless turns frontier models into distribution rather than competitors; implementation becomes easier; and Salesforce sustains 13–15% revenue growth with 36–38% margins. EPS around $25–27 at 20–21 times supports roughly $500–570.

The prices matter less than what each world requires. Q3 organic revenue now needs to visibly accelerate because management has made the promise falsifiable. cRPO needs to stay around 14% or better. Agentforce Apps, currently growing 8% in constant currency, needs to move toward double digits. Refills must persist as usage scales. Human seats need to remain resilient. And eventually Salesforce needs to disclose enough about implementation intensity and AI economics to demonstrate that the complexity underneath the abstraction is genuinely scalable.

No Salesforce

Twenty-seven years ago, Salesforce’s breakthrough was captured in two words: “No Software.”

The software did not actually disappear. Customer records still had to be stored, permissions enforced, workflows maintained, servers operated, and upgrades shipped. Salesforce simply hid enough of that complexity behind the browser that customers no longer had to think about it.

AI may force Salesforce to repeat the trick on itself.

Claude can replace the screen. Slack can replace the screen. A company’s own agent can replace the screen. What remains is the customer state, the permissions around it, the semantics that explain it, and the trusted system authorized to act on it.

The bear case assumes losing the screen means losing the customer.

Q2 offers the first financial evidence of a more interesting possibility: Salesforce may lose the screen and gain a second user.

Salesforce’s first revolution was “No Software.” Its second may be “No Salesforce” — at least not where the user can see it.

The irony is that disappearing from the screen may be how Salesforce becomes more important underneath it.

Subscribe now

• The unit of analysis is shifting from the GPU to the AI factory: NVIDIA says its revenue opportunity per gigawatt has risen from roughly $18 billion with Hopper to $25 billion with Blackwell and $40 billion with Vera Rubin.

• Custom silicon is not the whole threat: hyperscalers may take accelerator share while NVIDIA captures more of the CPU, networking, systems, software and surrounding infrastructure.

• The counterweight is increasingly financial: receivables, guarantees, investments and extended terms are rising. The key test is whether NVIDIA support grows slower than the economically productive infrastructure it enables.

Two different NVIDIA earnings reports appeared on Wednesday evening, and for the first hour the market could not decide which one mattered.

The first was the familiar NVIDIA print. Revenue was $96.2 billion, up 106% year-over-year and nearly $4 billion above consensus. Data Center revenue reached $89 billion, up 117%. Adjusted EPS was $2.22 versus $2.09 expected, and Q3 revenue guidance of $108 billion was about $4 billion above the Street.

The second print was less comfortable. Gross margin, still 75% in Q2, is guided to 74% next quarter; management expects memory costs to push it down to 71–72% in Q4 before recovering to 72–73% in fiscal 2028. Accounts receivable reached $63.1 billion and DSO increased from 45 to 60 days because NVIDIA extended payment terms on large multi-quarter purchases. Free cash flow was only $21.3 billion against nearly $54 billion of adjusted net income, while future supply and capacity commitments more than doubled to $279 billion.

The stock initially traded the second report and fell. Then CFO Colette Kress said:

“Customers’ forecasts point to our growth doubling next year. However…we expect to grow approximately 70% as we are supply-constrained.”

The stock reversed to finish the call roughly 4% higher.

That reversal captures the NVIDIA debate better than another discussion of whether $96 billion was a beat. Investors know AI demand is extraordinary; a company this large does not double revenue by accident. What they are trying to determine is what should be netted against the increasingly visible costs of sustaining that growth: custom silicon, falling margins, working capital and NVIDIA’s expanding financial involvement in its own ecosystem.

The answer, I think, begins with a different number.

The Chip Isn’t Enough

Kress said:

“Since Hopper, our revenue opportunity has grown from roughly $18 billion per gigawatt to $25 billion with Blackwell, to $40 billion with Vera Rubin.”

This is the most useful way to understand what NVIDIA has been building.

Power is increasingly the fixed resource in AI. NVIDIA cannot manufacture new gigawatts, but it can extract more intelligence from each one while increasing the amount of NVIDIA technology inside it. Hopper was principally about the accelerator. Blackwell expanded the rack. Rubin increasingly means Vera CPU, Rubin GPUs, NVLink, Ethernet or InfiniBand networking, systems, software and Groq LPUs.

The margin reset looks rather different through this lens. Blackwell at $25 billion of NVIDIA revenue per gigawatt and 75% gross margin implies about $18.8 billion of gross profit per gigawatt. Rubin at $40 billion and a 72.5% margin would imply roughly $29 billion. Even if mid-70s margins do not return, NVIDIA can potentially increase gross-profit capture from every scarce gigawatt by more than 50%.

This is also Jensen Huang’s answer to custom silicon. NVIDIA’s largest customers have powerful incentives to design their own chips: when compute is one of your largest expenses, NVIDIA’s gross margin is your potential savings. Amazon has Trainium, Google has TPUs, Meta has MTIA, and model companies are pursuing silicon of their own.

Huang did not argue that these efforts will fail. Instead he emphasized what NVIDIA sells that a specialized accelerator does not:

“One platform, fungible for every model and workload. Durable for the entire life cycle of AI.”

He made the contrast more explicitly in Q&A: many XPUs are optimized for a particular service or cloud, while NVIDIA is trying to build an AI-factory architecture that can run across clouds, models and stages of the AI lifecycle. That distinction matters because a multi-billion-dollar factory is not merely buying today’s inference workload; it is buying an asset that needs to remain useful when the workload changes.

Fungibility therefore has economic value. AWS can expand Trainium while simultaneously committing to two million additional NVIDIA GPUs and adopting Vera CPUs. Custom silicon may take accelerator share without taking an equivalent share of NVIDIA’s economics if NVIDIA captures more of the networking, CPU, software and system surrounding the accelerator.

That is the netting question in its first form: what should GPU share loss be netted against?

The morning after earnings supplied another clue. Reuters reported, citing The Information, that NVIDIA had agreed to acquire Hugging Face for $12.9 billion; NVIDIA and Hugging Face declined to comment. Hugging Face matters less because NVIDIA needs another model than because it sits between models and developers, hosting the open-model ecosystem and helping developers discover and deploy it.

This move fits a pattern that is becoming difficult to ignore. Networking threatened to become the constraint on GPU clusters, so NVIDIA bought Mellanox. CPU orchestration became more important, so Grace became Vera. Specialized decode created an architectural opening, so NVIDIA incorporated Groq. AI clouds lacked capacity, so NVIDIA helped develop NeoClouds. Model distribution is becoming strategically important, and NVIDIA may now be buying Hugging Face.

Each move looks adjacent on its own. Together they reveal the strategy: whenever a layer either constrains NVIDIA’s growth or offers customers a route around NVIDIA, the company moves sufficiently far into that layer to keep the larger system centered on NVIDIA.

NVIDIA’s customers are integrating downward to escape NVIDIA. NVIDIA is integrating outward and upward to make escaping NVIDIA matter less.

Capital Is a Component

The most unusual extension of this strategy is not Hugging Face. It is finance.

Kress said the frontier AI labs have enormous compute demand but are “growing faster than what their balance sheets and credit profiles can support.” This is an important distinction. NVIDIA is not describing weak demand; it is describing demand that cannot yet finance itself at the rate customers would like to grow.

NVIDIA’s response is strikingly similar to its response to every other bottleneck: remove it.

The company has invested nearly $50 billion in frontier labs and has brought Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR together around financing platforms intended to mobilize more than $500 billion of third-party capital. For some NeoCloud projects NVIDIA can provide a minimum revenue or take-or-pay commitment that helps independent lenders finance the facility, while NVIDIA participates in rental revenue above the floor.

Financing, in other words, has become another component of the AI factory.

The logic is compelling. NVIDIA can use the extraordinary profits from today’s installed base to accelerate tomorrow’s installed base, keep that capacity on NVIDIA architecture and potentially earn a recurring share of the revenue generated by the equipment it sold.

Fungibility matters here as well. A system capable of running many models and workloads for many potential customers ought to be easier to re-lease if the original customer fails. Better redeployability should support higher residual values; higher residual values reduce lender risk; lower risk lowers the customer’s cost of capital. More NVIDIA installations then deepen the secondary market that makes the next NVIDIA deployment still easier to finance.

CUDA could become a financing advantage as well as a software advantage.

This is where the argument becomes uncomfortable. NVIDIA generated nearly $54 billion of adjusted earnings this quarter but only $21 billion of free cash flow. Some of that gap is perfectly explainable by taxes, Rubin inventory and the extraordinary scale of the ramp. The more interesting number is 60-day DSO because management explicitly linked it to extended customer terms at the same time it acknowledged that parts of the AI ecosystem are growing faster than their financial capacity.

None of this proves the demand is circular. Demand can be completely real and still arrive faster than customers can finance it. The question is whether NVIDIA is providing temporary scaffolding around a young economic system or whether progressively more scaffolding is required to keep reported demand growing.

In The AI Handoff, we described this progression as Scarcity → Sponsorship → Conversion → Self-funding. Self-funding was never meant to imply that NVIDIA would stop investing; it meant that customer-generated cash and independent credit would increasingly finance each subsequent infrastructure cycle.

Q2 suggests the architecture is moving faster than the handoff.

The Netting Question

Last quarter I wrote that “a flywheel compounds; a circle spins.” My expectation was that ACIE would help distinguish the two: broadening non-hyperscaler demand combined with moderating NVIDIA support would suggest the ecosystem was increasingly standing on its own.

Instead, ACIE grew 25% sequentially and 138% year-over-year while NVIDIA’s financial involvement increased as well. There was some messiness: the $40.3 billion ACIE result missed consensus even as Hyperscale drove the Data Center beat but the broader demand signal is difficult to dismiss. The operating flywheel and financial scaffolding are not alternatives; at least for now, the latter is helping the former move faster.

My variant perception is therefore not that custom silicon does not matter, or that financing risk is imaginary. It is that the market is too focused on NVIDIA’s share of accelerators and not focused enough on NVIDIA’s expanding economic claim on each AI factory.

The $18 billion-to-$40 billion progression per gigawatt is the evidence. NVIDIA can lose something in GPUs, something in gross margin and something in financing support, yet still create substantially more economic value per scarce unit of infrastructure.

The real question is what must be netted against that expansion. If NVIDIA’s customers gradually become self-funding while NVIDIA continues increasing content per gigawatt, the company’s outward integration can more than offset accelerator-share erosion. If support requirements, receivables and guarantees rise as quickly as revenue while custom silicon takes the most profitable workloads, then orchestration starts looking more like subsidy.

Hugging Face adds a second risk. The orchestrator benefits from neutrality: NVIDIA is valuable because OpenAI, Anthropic, Meta, sovereigns and every cloud can all use it. Hugging Face is valuable for a similar reason. NVIDIA ownership could make the route from an open model to NVIDIA deployment much easier, but it could also make competing hardware vendors and model companies more interested in building an alternative. The more layers NVIDIA controls, the more valuable the architecture becomes and the greater the incentive for everyone else to route around it.

That tension defines the three-year range.

The bear case is not simply slower AI growth. One major customer or NeoCloud runs into financing difficulty, NVIDIA’s guarantees or equity exposure become economically relevant, DSO remains elevated, custom silicon captures standardized workloads and gross margin settles below 70%. AI can still succeed while NVIDIA’s equity does not.

The base case assumes management is broadly right about FY28: revenue grows around 70%, Rubin continues increasing NVIDIA content per gigawatt, ACIE broadens the customer base and hyperscalers use custom silicon without escaping the rest of the NVIDIA stack. Margins stabilize around 72–73%, cash conversion normalizes and financial support grows more slowly than the infrastructure it enables.

The bull case does not require NVIDIA to eliminate custom silicon. It requires NVIDIA systems to become measurably better assets: more fungible, easier to re-lease and consequently cheaper to finance. If the architecture earns a residual-value and cost-of-capital premium, the technical flywheel gains a financial counterpart.

There are four numbers I would watch to distinguish these outcomes. Free-cash-flow conversion should recover toward 70% or better of adjusted earnings; persistent conversion below 50% would make the financing concern structural. DSO should move back below roughly 55 days rather than remain above 60. Gross margins need to bottom around the promised 71–72% and recover thereafter. Finally, ACIE and independently financed capacity need to grow faster than the guarantees, equity and extended terms NVIDIA uses to support them.

The last measure is the one NVIDIA does not yet give us, and ultimately the one I care most about: how much NVIDIA support is required for each additional dollar of economically productive AI infrastructure?

NVIDIA’s customers will keep integrating downward because the economics are simply too important for them not to. NVIDIA does not need to stop them; it needs to keep moving the boundary of competition so that replacing the NVIDIA chip means replacing an ever-smaller portion of what NVIDIA provides.

Q2 was the first quarter to put a compelling number on that strategy: $40 billion per gigawatt.

The next few years will tell us what needs to be netted against it.

Subscribe now