Dolphin Research breaks down Anthropic's prospectus: 85% inference gross margin, trillion-dollar compute bill requires revenue to rise another 4 to 5 times

On October 26, Anthropic finally laid its prospectus on the table. Opening the first page reveals a tale of fire and ice:

On one side, second-quarter revenue of $11.56 billion, up 14-fold year over year, achieving GAAP operating profitability for the first time;

On the other side, compute contracts signed over ten months totaling nearly $1 trillion, equivalent to 22 times its annualized revenue.

In January this year, in "Deep Dive into Minimax and Zhipu: Large Models, Compute Intensity and Financing Endurance—A Decisive Battle?", Dolphin Research asked a question: what kind of business are large models, really? The answer at the time was not optimistic—it is a business that puts a capital-intensive operation "entirely on the income statement," where no matter how fast revenue runs, it cannot catch up with each generation's ever-fiercer training investment.

Nine months later, Anthropic's prospectus has for the first time provided a sample of "catching up." But what is the price of catching up? This time, Dolphin Research, drawing on the prospectus, seriously answers four questions:

1) What does Anthropic sell, and where does the money come from?

2) Is inference actually profitable? Compared with Zhipu and MiniMax, where exactly is the gap?

3) With a trillion-dollar compute bill, how much must revenue grow to bear it?

4) Is the money from Amazon, Google, Broadcom and others investment, or order purchases?

Below is the detailed analysis

01 The core of the toB business: selling Token "intelligence," charged by usage

Anthropic's business model is simple once stripped down: selling "intelligence" by token (the unit of measurement for text processed by a model). Of the $16.3 billion in revenue in the first half of 2026:

a. Usage-based revenue: consumption-type revenue billed by call volume (API (Application Programming Interface), Claude Code, etc.) was $13.1 billion, accounting for 80%;

In Dolphin Research's view, this share only explains two things: a. Pricing is expensive, and there are many KA (Key Account, major customer) clients; KA users generally consume directly; b. Model usage loyalty is still a question mark—usage-based consumption does not bind users to a model, and when a better and cheaper model emerges, users are unwilling to lock themselves into an annual-fee contract over which model they use.

The customer structure indeed reflects this:

KA customers: customers paying more than $100,000 annually surged from about 20 at the end of 2023 to more than 11,900 by the end of September 2026, contributing 74% of first-half revenue; among them, more than 1,900 customers paid over $1 million, and about 250 paid over $10 million.

The good news is that with the launch of Opus, customer penetration has risen and customer concentration is declining; the share of the top five customers fell from about 36% in 1Q26 to about 21% in 3Q26.

02 "Supermodel" vs "Supercloud": Is a showdown inevitable?

Channel game: large models vs cloud platforms—who ultimately has more bargaining power? Here the cloud platforms' distribution capabilities look not to be underestimated:

In Anthropic's rise this year, channel distribution has become increasingly important:

Revenue contributed by the three "checkout counters"—Amazon Bedrock, Google Gemini enterprise platform, and Microsoft Foundry—rose from 11% in 2023 all the way to 48% in the first half of 2026, and 58% of accounts receivable is owed by cloud vendors.

Because Anthropic signs directly with end customers and is "the principal responsible for the main delivery obligation," the company records revenue on a gross basis, and the channel share paid to cloud vendors is booked in sales expenses.

This "toll" is not small: Dolphin Research estimates that channel fees as a share of cloud marketplace revenue jumped from about 16–18% in 2025 to about 35% in the first half of 2026.

Behind this is that the supercloud giants have channel distribution power built on long-term customer accumulation and cloud compute reserves from massive capital investment, while large models, as investment black holes, can only use limited resources for model R&D, lack the energy for customer development, and have even less spare capital to stockpile large amounts of compute.

But once a supermodel like Anthropic has ample funds, it will most likely try to build its own AI cloud Token factory, strengthening its official website channel power and back-end self-built compute.

Figure 1: Channel and customer structure Note: Channel fee rates are estimates. In the first half of 2026, "sales, marketing and partnerships" expenses were $3.168 billion, channel fees increased by about $2.7 billion year over year (the total for this item in the same period last year was only $120 million), corresponding to cloud marketplace revenue of about $7.8 billion (48% × $16.286 billion), a ratio of about 35%; on the same basis in 2025 it was about 16–18%.

03 Inference earns effortlessly, training devours cash: 85% vs 45%, the real economics behind two gross margins

The most eye-catching set of numbers in the prospectus is the scissors gap:

2Q26 revenue rose 1368% year over year, while compute and infrastructure expenses rose only 300% over the same period. Compute expenses as a share of revenue fell from 251% in 1Q25 all the way to 48% in 2Q26 over five quarters, and the operating margin flipped from −249% to +8%, or 19% adjusted.

Among all listed large-model companies, we finally see one whose revenue growth has caught up with model training investment.

Figure 2: The scissors gap between revenue and compute costs

But the compute costs here need to be viewed separately: the bulk is training costs for model training (within R&D expenses), and a smaller part is inference costs incurred in serving users after the model is deployed, which sit in cost of revenue.

Inference revenue generation: In the first half of 2026, the inference cost (cost of revenue) of serving paying customers was only nearly $2.4 billion, with a gross margin of 85%, more than double the 39.5% for all of 2025.

Training devours cash: Of the $9.7 billion in R&D expenses, training costs alone were $6.6 billion, swallowing 40% of revenue in one bite; excluding the hard spending on training and inference, the gross margin is 45%.

English text is being prepared.

Factoring in R&D personnel investment and so on, the final GAAP operating loss was $690 million. Excluding $2.2 billion in stock-based compensation and other items, adjusted operating profit only turned positive to +$1.62 billion.

Figure 3: Income statement by function for the first half of 2026

Walking through this account, the business logic is also simple: looking at the inference gross margin alone is meaningless. The company putting training and inference into "the same compute pool" for unified planning and assessing them on a total-volume basis is the same signal.

This is tantamount to saying that as long as model progress does not converge and no single model has a monopoly, booking these two numbers under different line items is almost self-deception. When doing the math on the model business, it is simple and brutal—revenue growth must outrun the rigid growth of training and inference investment. To command a high valuation, this is the metric that must be watched closely every quarter.

IV. Same model business, where is the gap? Anthropic vs Zhipu vs MiniMax

Dolphin Research puts the already-covered Zhipu and MiniMax here for comparison:

Figure 4: Same-basis comparison of Anthropic, Zhipu, and MiniMax

Gross margin: one in the sky, two on the ground. Anthropic 85.4%; but it books about $2.7 billion in channel fees and compute for free users under sales expenses, and after deducting channel fees it is about 68% (estimated), which is the comparable basis with Chinese companies.

Zhipu overall 26.4%, of which the cloud/API business is 24.6% (in the same period last year it was still −0.4%, turning positive only thanks to a 101% price increase plus scale effects); MiniMax overall 17.9%.

The root of the gap is the pricing power difference between SOTA (State of the Art, industry-best) performance "supermodels" vs. cost-effective "ordinary models."

Compute burden: the two domestic companies are a year and a half behind on compute economics. Anthropic's total compute cost in the first half accounted for 55% of revenue. Zhipu and MiniMax do not break out compute within R&D in their interim reports; Dolphin Research estimates based on "inference compute in cost of revenue + R&D × the compute share disclosed in the prospectus":

Zhipu about 226%, MiniMax about 283%—roughly where Anthropic was in 1Q25 (251%).

Expense structure: the gap is almost entirely in training. The sales expense ratio for all three is 19–23% (if channel fees are excluded, probably 10%+, somewhat lower than Chinese model companies), with little difference; R&D as a share of revenue is 60% for Anthropic, 223% for Zhipu, and 255% for MiniMax; operating margins are −4%, −225%, and −280%.

By comparison, Anthropic has proven that an "intelligence premium" can tame the training cash-devouring beast.

Of course, the comparison here is also somewhat unfair—the Chinese model with the best-looking financial statements should be DeepSeek, and comparing it with Anthropic is a true contest between the two poles of super performance vs. super cost-effectiveness.

05 The soul-searching question: with a trillion-dollar bill, how much must revenue rise to bear it?

At the end of 2025, Anthropic's non-cancelable compute commitments were only $54.6 billion. By the end of June 2026, purchase commitments had become $265.6 billion, plus signed but not yet commenced leases of $321.4 billion; after June, it signed another $245 billion in long-term contracts and $21 billion in short-term contracts.

Add the floating portion not recorded on the balance sheet—$125–150 billion in additional TPU (Tensor Processing Unit, Google's in-house tensor processor) for 2028–2031, and more than $20 billion from AMD. A simple sum gives about $1.01 trillion (upper-bound basis, estimated).

SpaceX AI: at most $120.5 billion (about $3.1 billion per month);

The contracts with Google, Amazon, and Microsoft are all "take-or-pay"—pay whether you use it or not; in the entire bill, the only part with flexibility is SpaceX AI—after 1Q27 it can be canceled with 90 days' notice.

Figure 6: Compute contracts by counterparty

The key question: how much revenue is needed to bear this bill?

Dolphin Research spreads the named contracts on a straight-line basis over their contract terms, giving a peak annualized figure of about $115 billion/year (estimated, not yet including additional TPU, AMD, and short-term contracts), about 5 times current compute spending (2Q26 annualized at about $22.2 billion).

If the compute cost ratio is to stay at the roughly 55% break-even line seen in the first half, annual revenue needs to be about $210 billion; to return to the roughly 45–48% level of 2Q26, it needs about $255 billion. And right now, 2Q26 annualized revenue is $46.2 billion.

Figure 7: Closing conditions of the economic model

In other words, revenue needs to rise another 4–5x. Over the past four quarters revenue rose about 15x, so another 4–5x is not fantasy—but contracts are rigid, revenue is not.

Management itself did not shy away: "As a large amount of new compute comes online, the related costs will be recognized before the revenue they ultimately support, and operating margin and adjusted operating margin may be compressed during these periods… and the more short-term compute is used, the higher the unit cost."

In translation: the 8% operating margin in 2Q26 is very likely the high point before a new wave of concentrated compute deployment.

06 Is big tech's money investment or a deposit?

Anthropic has said of its relationship with the cloud giants: the cloud vendors are "simultaneously our investors, customers, distribution partners, and competitors"; and Broadcom is "simultaneously a supplier and a lessor."

Figure 8: Matrix of Anthropic's relationships with big tech

Taking them one by one:

Amazon: two rounds of $4 billion convertible notes each in 2023 and 2024; the remaining $6.2 billion in notes is all of Anthropic's current debt.

In 2026 it invested another roughly $10 billion in equity, plus a $15 billion facility—released $5 billion at a time "according to compute capacity milestones with Amazon"; stake capped at 12%.

In the other direction, Anthropic committed to buy at least $110 billion of compute from AWS from May 2026 to April 2036; Amazon can even offset its note lending to Anthropic with "money Anthropic owes Amazon."

Google: the early roughly $3.3 billion in notes has been fully converted to equity; in 2026, $20 billion in equity investment, plus another $20 billion facility (drawdown conditions tied to third-party compute agreements Anthropic signs and revenue targets), and up to $10 billion in subscription rights at IPO; stake capped at 14.99%.

In the other direction, the Google Cloud contract is at least $114.5 billion, and in 2028–2031 it will add another $125–150 billion from Google and other TPU providers.

Microsoft: there is equity investment but the amount was not disclosed; the Azure contract is at least $44.9 billion; Claude is also sold through Foundry and Microsoft 365.

Broadcom: the most "ingenious." A 5-year TPU equipment finance lease, with Broadcom collecting about $161.2 billion in rent; at the same time, Broadcom lends Anthropic up to $42 billion in convertible notes, "specifically to pay the quarterly rent on this lease," with 9% payment-in-kind interest, convertible after IPO into shares at a 10% discount to the 30-day average price.

Use Broadcom's money to pay Broadcom's rent, and finally convert the debt into Anthropic stock.

AMD: from 2027 it will supply MI450 Helios, with an initial batch of more than $20 billion; AMD will subscribe to up to $5 billion of non-voting stock as GPU deliveries progress.

Nvidia: no direct contract, but its GPUs come in through a roundabout route via Microsoft, SpaceXAI, Nscale, and Lambda; the latter three contracts total about $206 billion.

Adding up the accounts: Amazon, Google, Broadcom, and AMD together provided or committed about $133.3 billion in funding to Anthropic, in exchange for about $405.7 billion in compute orders—for every $1 invested, roughly $3 in orders comes back.

Figure 9: Supplier funding and compute orders

So is this money investment or a deposit? Dolphin Research's view: both.

This web lets Anthropic lock in compute far beyond its own size with very little cash; it also lets the big tech companies turn their investment in Anthropic into orders for their own clouds and chips.

The benefit to Anthropic: money is no longer scarce—at the end of June, cash plus short-term investments were about $131.7 billion, plus $77 billion in undrawn strategic note facilities and a $15 billion revolving credit line.

But such rigid payment agreements mean that once Anthropic faces a deterioration in competition similar to the third quarter, it will definitely cut prices: locked-in compute capacity that has just been paid for must be fully utilized; if users cannot use it all, it must be sold at a lower price; with capacity locked in now, no one will defend price—penetration rate comes first. Under this logic, valuation pressure on cost-effectiveness models is enormous.

07 A few special reminders + special equity structure

In the prospectus, several points deserve attention:

In the first three quarters of 2026, 5–7% of revenue came from customers who violated usage policies; after enforcement was strengthened in August–September, this portion will decline.

Net revenue retention (NRR) of about 900% is a figure from the customer phase of going from trial to scale-up, and cannot be extrapolated.

In the first half of 2026, the net loss was $80.7 billion, mainly due to fair value changes on convertible notes ($72.4 billion)—valuation rose, so the notes rose with it; this is a non-cash loss, and such losses should not be looked at.

Management said plainly: "Revenue growth will decline, and may decline significantly."

In this version of the prospectus, third-quarter revenue, cash at the end of September, the annual schedule of committed payments, and big tech's shareholding percentages are all blank.

The equity arrangement is very special:

Control:

Two class shares without economic rights—one trust T share, with the right to elect 4 of 7 seats; one founder F share, with a fixed 50.1% voting right.

Big tech has no say:

Amazon and Google hold Class N shares (non-voting); the prospectus discloses no board seats or observer rights for any big tech company; after IPO, each of the two will hold more than 5%.

Philanthropic arrangement:

847 million Class N shares (5% on a fully diluted basis) will be contributed to a philanthropic fund, with up to an additional 2% issuance each time the stock price rises 50%; the founder has committed to donate 80% of his equity.

08 Conclusion: "Cash-devouring beast on the income statement" vs "mortgage slave on the balance sheet"—who laughs last is still hard to say

When Dolphin Research tracked China's two major model companies, it said that large models are a business that puts an entire capital-intensive business "entirely on the income statement"; only when a single training run can be amortized over the long term and the model returns to being a true "balance sheet business" will this business truly stand.

Anthropic's prospectus happens to show half of that path: relying on revenue scale, it made the training cash-devouring beast be caught by the income statement for the first time, with a single-quarter operating profit and 85% inference gross margin, leaving the two domestic companies far behind.

But the hidden cost of scaling up is also clear: it has turned the compute of the next few years into nearly $1 trillion of contracts and leases on the balance sheet, the vast majority take-or-pay.

In the past it was "losses spread across the income statement"; now it is "burden pressing on the balance sheet." If revenue rises another 4–5x, this is a cash cow; if it does not, who ultimately wins under this trillion-dollar bill will depend on who is really the boss at the table in the competition between AI models and cloud vendors.

This article is from the WeChat public account "Dolphin Research" (ID: haituntouyan), author: Dolphin Research, published by 36Kr with authorization.