SemiAnalysis on Anthropic’s S-⁠1: 90% Incremental Margins and a Compute Bill Smaller Than It Looks

Margins: inference pays

The panel opens on profitability: 85% inference gross margins, 41% compute margins and two profitable quarters already. O’Laughlin says the incremental numbers are the real story, about 90% gross margin and 40% EBITDA on new revenue, which he calls “as good as it gets” and at least on par with the hyperscalers. In his telling, a dollar of revenue costs about ten cents to serve, which answers the bears who argued usage was heavily subsidized.

Retention does the rest. A customer that spent $1 a year ago now spends about $9, versus 500% net dollar retention in Q1, and because the growth comes from usage, little sales spend sits behind it. About 250 customers spend more than $10 million on a trailing 12-month basis and roughly 1,900 spend more than $1 million, a group the hosts say includes SemiAnalysis itself.

The compute bill is smaller than it looks

Eliahou Ontiveros, the team’s data center analyst, says the market had been anxious about Anthropic because nobody can finance a data center on its credit, and it had been punishing names with direct Anthropic or OpenAI exposure, on the data center side and the GPU side, in recent weeks. The S-1 answers that: more than $100 billion of cash on the books, over $200 billion once the IPO proceeds and promised funding from Google and others arrive, and very little cash burned so far.

On the $500–600 billion of commitments, he uses Anthropic’s deal with Nscale, about $100 billion per gigawatt, as the yardstick: half a trillion dollars buys only about 5 GW. Anthropic runs on under 5 GW today and has contracted materially more, mixing long-term deals like Nscale with short ones from hyperscalers and SpaceX that can be canceled at almost any time. He contrasts last year’s scare over OpenAI, then at about $20 billion annualized against $1.4 trillion of commitments, with Anthropic at roughly $80 billion of ARR against half a trillion.

New money adds to the cushion: a $40 billion Google commitment with milestones that include going public, and a $42 billion Broadcom arrangement that lets Anthropic buy many more TPUs. He says most of 2027’s compute is now de-risked and expects annual revenue in the multiple hundreds of billions next year, though whether revenue can accelerate enough to use all of it is still the question.

Efficiency, and why multi-chip matters

Brookhart points to the S-1’s chart of inference revenue per megawatt across the whole fleet, up from about $16 million last year to more than $60 million, with a small dip in Q3 as new SpaceX compute came online and more went to safety work. The filing’s own line: revenue grew 1,368% year over year in the three months ended June 30, while compute and infrastructure expenses grew 300%.

Is that Anthropic or just better chips? SemiAnalysis’ own benchmarks show Nvidia’s GB systems around 50 times better than Hopper. Nanos credits both: Anthropic runs Google TPUs, Amazon Trainium and Nvidia and AMD GPUs, and the lab with the best model and engineers writes the software that gets the most out of each new generation. The panel adds that Anthropic was first to push hard on multiple chip suppliers while OpenAI stayed an Nvidia-first shop, and that models now handle low-level systems work well enough that the choice of chip matters less than it used to.

OpenAI as number two

Is OpenAI cooked? The panel says no, but argues that being number two across the board is the worst place to be: with Anthropic the low-cost, high-margin producer, OpenAI ends up competing on price rather than charging a premium, much as the number two in semiconductors ends up with less share and worse margins. In AI, they argue, buyers want the best in breed.

There is pushback. One panelist sees room for both labs and still finds OpenAI models more reliable for data analysis. OpenAI is said to sell better into large enterprises, has been discounting against Anthropic accounts, took enterprise more seriously from July and had a strong Q3, but its margins look nothing like Anthropic’s while it carries a huge free consumer base.

Neoclouds, SpaceX and the debt market

The panel reads Elon Musk’s and Chamath Palihapitiya’s recent calls for the end of the neoclouds as self-interested: SpaceX is itself the largest neocloud, and with what the hosts describe as a liquidity run on neocloud debt, rivals can’t place new financing. If they can’t fund their buildouts, investors line up for SpaceX vehicles, its cost of capital falls, and it can sign compute deals at about $50 million per megawatt per year.

The team also sees a possible inversion, with data centers rather than chips the constraint today and the reverse by the end of next year, as Musk talks of building 10 to 20 gigawatts himself. Eliahou Ontiveros explains why short SpaceX contracts suit Anthropic’s books: at about $50 million per megawatt per year on 90-day terms, the commitment is roughly $15 billion per gigawatt, against about $100 billion per gigawatt for a six-year deal like Nscale’s, which he expects to get pricier. It costs more at the margin, but it is light on commitments.

Room to grow, and the market’s reaction

Brookhart sees a long runway in ordinary enterprise work. The top 5% of Claude Enterprise users spend about $18,000 a year while non-Claude Code users spend under $75 a month, and Anthropic is at about 0.1% of Fortune 500 operating budgets, under $8 a month per employee, heading into the first traditional Q4 enterprise IT budgeting cycle for these tools. Government is about 1% of revenue in each period in the filing, and the panel sizes that opportunity against hundreds of millions of Microsoft 365 commercial seats, though many enterprises still restrict some job types to Sonnet rather than frontier models.

On the market’s reaction, O’Laughlin says AI build-out names, especially data centers, were the day’s winners, but the move was weaker than this year’s rallies, and that “the AI trade is exhausted.” The panel still thinks it is early: Brookhart borrows O’Laughlin’s line that the industry is still paving the parking lot and the ballpark hasn’t been built, with the two frontier labs at about $120 billion of combined ARR.

Underpromise, overdeliver

O’Laughlin’s pick for the most striking part of the filing is its stance on expectations. The company says it never wants to overextend itself and calls underpromise and overdeliver a core principle, which he reads as a promise never to miss what investors expect, an extraordinary thing to put in an S-1 when many are nervous about Q4. The panel notes the irony of a company that warns at length about AI risks casting itself as the one underpromising.