SemiAnalysis Analyzes Anthropic S-1: 90% Incremental Gross Margin and Half-Trillion Compute Commitment

In the past six months, the laziest line used to short AI has been: these frontier labs collect a dollar and spend a dollar ten, running entirely on financing subsidies. After Anthropic's S-1 came out, that claim has flipped for it: sell a dollar, cost a dime.

And this is the metric it wrote into its own prospectus, not something outsiders calculated for it. SemiAnalysis even used EMERGENCY in the title of this episode, and the four hosts opened by calling it "the most important day in the history of capitalism." Doug admitted he made that line up on the spot, it just "felt right." Laugh all you want, but the numbers are brutal.

Sell a dollar, cost a dime

Incremental gross margin 90%, incremental EBIT 40%. Doug's exact words were that 40% EBIT has already hit the ceiling, and that's the level hyperscalers operate at. Underneath there are two more specific line items: inference gross margin 85%, compute gross margin 41%. On a non-GAAP basis it's already above 10%, profitable for two consecutive quarters, and this was achieved while training spending is still ramping up.

I think the analogy Doug used is clearer than any chart. When you sell something, it's not collect a dollar and hand out a dollar ten; it's collect a dollar and spend only a dime. That's gross margin. He said bears have believed for years that there was massive subsidy in this, and now it's clear there isn't. He also casually recalled that many years ago $1 billion in annualized revenue counted as adulthood, and this company's scale now requires a different ruler.

The episode had another detail: on the day the S-1 came out, the well-known short seller who last week was still calling Jensen Huang a fraud didn't post a single tweet. Doug said this might be the most interesting piece of information in the whole thing.

Customers who spent a dollar last year spend nine this year

On net revenue retention, normally a SaaS company that hits 120% will brag about it for a year. Anthropic's customers who spent $1 last year now spend $9, and at the Q1 checkpoint that number was 500%.

It barely has to pay sales expenses for upsell; usage grows on its own, and every extra dollar drops straight to profit. Doug said many software companies spend heavily on customer acquisition, but this company doesn't.

It's more intuitive when you get to individuals. Among Claude's enterprise customers, the top 5% of users spend $18,000 a year. Meanwhile, ordinary employees who don't use Claude Code spend less than $75 a month on an app. That's a twentyfold gap. There are also 250 customers paying over $10 million a year and 1,900 paying over $1 million. Joey joked that we at SemiAnalysis are one of those 1,900, and sooner or later we'll squeeze into the 250.

In the three months ended June 30, revenue rose 1368% year over year, while compute and infrastructure costs rose only 300% over the same period. This scissors gap is the place in the whole document that best demonstrates "economies of scale really exist." Jeremie reminded everyone that it's actually hard to disentangle how much of this is their engineering capability and how much is industry-wide hardware gains like GB being 50 times stronger than Hopper.

Half a trillion in commitments, but only five gigawatts locked

What outsiders see is $500 billion to $600 billion in compute commitments, a number big enough to make you close the page on the spot. Jeremie's breakdown is by unit price: the Nscale deal is priced at $100 billion per gigawatt, so half a trillion only buys 5 gigawatts. And the compute they have online now is still less than 5 gigawatts, while the gigawatts they've signed for are far above that. Between commitments and actual demand, there's a large buffer. Jeremie was looking at this document from the data center side. For the past few weeks the market has been hammering assets nakedly exposed to Anthropic and OpenAI, because no one dares to finance data centers on the credit of a private company.

A horizontal comparison shows how conservative this is. Last year when OpenAI was being criticized, it had $20 billion in annualized revenue and $1.4 trillion in commitments. Anthropic is $80 billion ARR against $500 billion in commitments, with over $100 billion in cash on the books, and including this round of financing and the money Google and Amazon have already committed, it will exceed $200 billion. Broadcom's $42 billion deal is for buying TPUs, and Google's $40 billion comes with milestones, one of which is completing the IPO.

It hasn't gone all-in on long-term contracts. Besides long deals like Nscale, there are also a batch of SpaceX-style on-demand contracts, GB300 at $14 per hour, cancelable at any time, with corresponding locked amounts of only $15 billion per gigawatt. Expensive, gross margin gets compressed, but the balance sheet stays clean. Jeremie's judgment is that data centers will be tighter than chips before the end of next year, and after that it may reverse. And Elon, while shouting that all neoclouds must die, is himself the world's largest neocloud, and right now new neocloud debt is basically unsellable.

On the customer concentration section, the four of them started arguing on the spot

The top five customers account for 25% of revenue, and another 48% comes through cloud vendors' marketplaces, meaning channels like Bedrock and Vertex. Jordan wanted to add these two numbers together; Doug said you can't, because these are two different accounting bases. Joey added that the 48% does not include the top five customers, since those customers wouldn't go around through Bedrock to place orders. The four of them didn't fully converge in those two minutes, and finally landed on the line "you have to see who it's counted against after passing through the channel." Joey also cited an alternative data set: the top 1% of customers contributed 70% to 80% of usage in the panel, more concentrated than the 25% in the prospectus.

What really demonstrates stickiness is another number. In Q3, 20% of revenue still came from models released before April. 80% have already switched to the latest version, but the remaining 20% shows that a batch of customers need to queue through procedures to change models; for a Fortune 1500 company to certify a new model is much slower than for a startup. Add to that the customers who signed prepayment commitments in advance, the same logic as signing a long-term AWS contract, and they are locked in tighter than you'd imagine.

The prospectus says "we won't let you down"

Doug said the craziest line in the entire S-1 is putting "under-promise, over-deliver" directly into a formal document; he thinks there's no more outrageous move than that. The market is still anxious about the Q4 ARR number, but the company has just told you in the prospectus that they never let expectations fall short.

Jordan pushed it one layer further, saying that the claims over the years about cybersecurity ending the world and AGI having a 10% chance of wiping out humanity also count as under-promise; the reality may be worse. This part is funny, and the chill is in it too.

Q4 has a real pitfall. The prospectus says that in the two weeks ending January 2, total customer spending fell about 20% compared with the two weeks before December 19, which translates to about 10% for the month of December. The four of them argued about this: is it because everyone is off work for the Christmas holiday, or is the product itself seasonal? Joey thinks this year will be somewhat better, because Claude Code has already entered daily workflows and is no longer just an app you open during meetings.

Second place is the hardest; the parking lot has only just begun to be paved

When it came to OpenAI, Doug's framing was direct: it is second on every battlefield, losing to Gemini on the consumer side, fighting Claude on the enterprise side, and open-source models are pushing up from below. Second place is the most uncomfortable position in the market, because from now on it can only compete with Anthropic on price. His argument is that the lowest-cost, highest-gross-margin player takes the premium, and everyone else can only compete on price, and a price war will constantly tell customers you're not as good as the others.

To be fair, he also gave the counter-evidence. OpenAI has recently been strong in enterprise sales, signing those established Fortune 1500 companies, and is also offering aggressive discounts targeting Anthropic's customers; Q3 execution was decent. The number Joey added was even more piercing: of 1 billion weekly active users, only 6% pay, and they still have to support all the free users.

Joey was calculating a different set of numbers. Today they account for only 0.1% of Fortune 500 operating budgets, which works out to less than $8 per employee per month. Government customers account for only about 1% of revenue, while across that line are hundreds of millions of Microsoft 365 commercial seats.

The combined ARR of the two frontier labs has already reached $120 billion; Joey said this is equivalent to the stadium not yet being built, and they are still paving the parking lot asphalt and painting the lines. The metaphor is a bit rustic, but it makes the position clear.

What really needs watching is whether this 90% incremental gross margin can hold when customers spread from programmers to ordinary white-collar workers. The people who wrote the prospectus have already put the answer there: they never let you down.

More hardcore podcast content: Podwise