Anthropic's IPO filing gives AI finance something it has rarely had, a frontier lab that appears to make money. The same document also commits the company to something close to a trillion dollars of compute spending that only works if growth keeps showing up on schedule. The S-1 became public earlier this week, trading on the Nasdaq could start as soon as next month, and the valuation being discussed runs from $1.5 trillion to $2 trillion. Two readings of the numbers have emerged. One sees a disciplined enterprise business with healthy margins. The other sees a company whose fixed costs leave it no choice but to build something extraordinary. Both can be true at once, which is what makes the filing so entertaining.
What the filing shows
Start with the part that pleases Anthropic's fans. These are the headline figures as the filing and the analysts reading it present them:
- Revenue. $16.3 billion in the first half of 2026, fourteen times the same stretch of 2025.
- Reported loss. $80 billion, which turns out to be mostly an accounting artifact (more on that below).
- Adjusted income. $1.62 billion for the half, once that charge and some others, employee compensation among them, are stripped out.
- Retention. Net revenue retention of 900 percent.
- Customer size. About three-quarters of customers in the first half were spending at least $100,000 a year.
- Run rate. SemiAnalysis puts annualized recurring revenue above $70 billion, and the company's current valuation above $1 trillion.
The question of whether AI can earn enough to pay for itself has hung over the industry for a while. For this company, at this moment, the filing says yes.
Joey Brookhart, an analyst at SemiAnalysis, was more enthusiastic still. He called it an awesome S-1 and said he could find little to nitpick in gross margins, compute margins or free cash flow. He pointed out that the semiconductor complex, the neoclouds and compute stocks all rallied again in the two days after the filing. That was a relief of sorts, because the preceding three weeks had been rough. Some investors had been disappointed by the September ARR numbers after the company's investor day.
Brookhart's firm built an outside-in model of Anthropic's finances back in July, when the company was private and the inputs were press reports. The audited numbers now let them explain the unit economics properly. The most striking one is gross margin on inference, which he says was negative 94 percent two years ago and is positive 85 percent today.
An $80 billion loss that isn't quite one
The $80 billion figure drew the most attention in the comment sections, as large red numbers tend to. Brookhart explained that it is a non-cash charge. Anthropic issued warrants and convertible notes, its valuation then rose steeply, and accounting rules require the increase to show up on the income statement. Anyone who has to ask whether this belongs in an assessment of the underlying business can be told that it does not. He compared it to stock-based compensation, which also hits the income statement without any cash leaving the bank.
The interview raised Amazon as the counterparty to that arrangement. Amazon is pleased that Anthropic's valuation has climbed and that Anthropic buys a great deal of AWS. Over time, though, its stake gets diluted, and Anthropic is about to put a large batch of new shares on the market. It does not look like a company that plans to stop raising money.
Brookhart's view is relaxed. By his count the company has burned only about $10 billion of cash. It holds more than $120 billion in cash and marketable securities, and he expects roughly another $100 billion after the IPO. Anthropic rents its compute rather than buying it, so there are no giant capital outlays on its books. The commitments are, as he put it, technically future debt that sits somewhat off the balance sheet. Might it need to raise another $100 billion of debt or equity at some point? Possibly. Can it keep structuring compute deals creatively? Probably. And for an investor who believes the long-term AGI case, he said, dilution is not the worry, because a $10 trillion-plus business can outrun it.
The bill
The reading that worries more starts from the S-1's commitments section. Anthropic leaves the exact purchase and lease totals blank in places, but the commitments and contingencies note gives enough to piece the picture together. The commitments run through 2045.
| Item | Amount | Terms |
|---|---|---|
| Fixed minimum of non-cancelable purchase commitments | $265.6 billion | 1 to 20 years |
| Lease agreements not yet commenced, mostly compute and infrastructure capacity (unaudited figure as of June 30) | $321.4 billion | Not given |
| New compute and infrastructure agreements signed after that date | About $245 billion | 1 to 20 years |
| Transactions not yet finalized (company estimate) | $125 billion to $150 billion | Not given |
| Other compute commitments | Variable, not quantified | Not given |
One analyst's tally of the firm items comes to a minimum of around $850 billion over the next 20 years. That rises to about a trillion dollars if the unfinalized deals come in at the estimated cost. He allows that he may be missing something.
Anthropic gives no payment schedule. Harrison Rolfes of PitchBook models it at roughly $90 billion a year from 2027 to 2030. By his math, holding the current ratio of cost to compute steady would take $188 billion in revenue. "Growth is not optional for this company," he writes. "The contracts assume it."
Brookhart does not dispute the arithmetic of the commitments, but he reads the same facts more calmly. The headline number he works with is $500 billion to $600 billion over seven to ten years, and the interviewer cited another analysis putting the figure above $800 billion over a decade. He argues that today's spending buys fewer gigawatts than the totals suggest, since Anthropic currently pays close to $11 million per megawatt, or $11 billion per gigawatt. At 20 gigawatts the commitments would run well into the trillions, but that is a function of how much compute the company is running. The real variable, in his telling, is revenue relative to compute. The company does not have to reach a trillion dollars of revenue extremely quickly to pay for all this, he said, unless revenue disappears or customers churn.
Why the economics keep improving
Source of the margin gains, according to Brookhart: token throughput. Anthropic runs a diverse fleet of Trainium chips, TPUs and GPUs. On Trainium, throughput is up five to ten times year over year. When hardware gains stack with software that fits the model to the hardware, a chip that served X tokens per second a year ago now serves five or ten times that.
The revenue-per-megawatt figures show what that does:
- Today. Anthropic earns about $65 million of revenue per inference megawatt, up from the mid-teens a year ago.
- Opus 5.5 on Nvidia Blackwell. SemiAnalysis estimates $100 million to $120 million per megawatt.
- Cost side. The company pays a little over $10 million per megawatt in compute cost of goods sold.
The fat margin does not sit in the bank. Only about 30 percent of Anthropic's compute goes to inference that earns revenue, Brookhart said. The rest goes to experimentation, research, training and post-training. The filing's cost lines tell a similar story in dollars. Compute and infrastructure tied to customers (plus a few other items) cost $2.3 billion in the first half, while model development (plus a few other items) cost $9.7 billion.
This is also where the filing gets less sunny. Brendan Burke, an analyst at Futurum Group, noted that the S-1 "does freeze in time a happy period when their initial models skyrocketed in usage and they didn't have enough compute so they are able to charge more for what they had." The third-quarter numbers are not available to check whether the trend held.
Burke also said frontier labs need high-margin products to avoid a race to the bottom on token pricing. Opus 5.5 cut customer costs by 40 percent compared with Opus 5. Anthropic's argument is that lower cost brings more usage, which preserves its margins. If that happens with every new model, tasks get very cheap very quickly. Gil Luria of DA Davidson said this is survivable, since AWS and other cloud providers cut prices every year and still enjoy steady growth and rising margins.
Brookhart said that Anthropic did not expect to be at an 11 percent non-GAAP EBITA margin and would prefer to sit near zero, or below, if it could put the compute into training. He says the company has told investors on its roadshow not to expect the current margin leverage to last.
The treadmill
The interviewer asked what might be the most naive question in AI, and it is a good one. If you have built a technology that could reshape every industry, why not stop training and sell the profitable model into law, medicine and everything else that has yet to adopt it? Apple reuses its iPhone tooling across several generations and keeps selling the older phones. The AI industry deprecates last year's model almost immediately.
Brookhart gave several reasons. The filing shows that 80 percent of revenue in the last three months came from models released in the last six months. Revenue also arrives in bursts after each release, so the company cannot stop training and expect the money to keep flowing. In cybersecurity, if an attacker uses a frontier model then the defender needs one too. In a legal case, would you want your lawyer using a non-frontier model if the other side was not? Demand for frontier intelligence will persist, he said, especially in competitive industries and wherever work moves from labor augmentation to labor replacement.
He also granted the iPhone point partly. Plenty of work, such as summarizing email, needs no frontier model, and open source is picking up those tasks. At many enterprises only the coders are allowed near the frontier models, while everyone else is limited to Sonnet. That says something about what management thinks frontier intelligence is worth. Anthropic can use a Sonnet-class model the way Apple uses the SE, as the cheaper tier.
The lighter-touch strategy would also face a governance problem. Brookhart thinks Anthropic is structured to resist the public-market pressure to stop spending on training, because of voting stock arrangements. He pointed to Mark Zuckerberg's heavy investments as a precedent. The interviewer added that those did not always pay off, and Brookhart agreed.
Competition from every side
The filing names only OpenAI and Google as rivals. Brookhart's account is broader, and it is less comfortable.
In the first half of the year, Anthropic took 90 percent of the industry's net new ARR. In the third quarter, he said, OpenAI probably took the majority. He blames part of OpenAI's earlier slowdown on a botched release of Fable and the lack of zero data retention, which enterprises care about. OpenAI could fix that with Fable 5.5. Open-source token inference grew from perhaps $1 billion of ARR in the first quarter to more than $10 billion in the third.
Google DeepMind and Meta Superintelligence Labs are well funded and are adding training compute at a similar scale. SpaceX is a further wild card, given Elon Musk's statements about 10 net new gigawatts next year, half of it for long-term internal use. The Verge-style gossip about Musk and rivals aside, the other sourced piece of evidence is that Anthropic's own filing has Google as a supplier and competitor at once.
Switching costs in coding are low. SemiAnalysis itself went from 90 or 95 percent of usage on Anthropic's models at the end of the second quarter to 70 percent after OpenAI's ChatGPT 5.6 and Astra arrived. With Opus 5.5 it is back above 90 percent. Claude Opus 5.5 currently tops the Epoch Capabilities Index, but that leaderboard changes frequently.
The revenue is also a pure consumption business. Anthropic's 900 percent net revenue retention is exciting, but it is not the three-year-contract stickiness of ServiceNow or Workday, where gross retention runs 95 to 98 percent. Brookhart noted a 20 percent drop in token spend in the last two weeks of last year compared with the prior two weeks, when people stopped coding for work over the holidays.
Some enterprises are also training and post-training open-weight models themselves and running them on their own servers. Several chief executives are wary about giving their data to OpenAI or Anthropic. Law firms are running AI on-premises, which the interviewer called possibly the funniest turn in history.
There is a distribution problem too. Google can sell AI inside products that businesses already pay for, and Anthropic has no such second revenue stream. One analysis recalls how Microsoft Teams beat Slack not because it was better but because it was bundled with software companies already bought. The best model matters more in cybersecurity than in instant messaging, but distribution is hard to compete against.
Where the next dollar comes from
The present customer base is narrow. Brookhart said the two leading labs together have about $120 billion of ARR, concentrated in the top 1 percent of customers, who spend $50,000 to $100,000 per employee. Most of the Fortune 500 and Global 2000 spend little. Much of Anthropic's revenue is coding, which makes it close to a single-vertical company.
The growth case rests on penetration. Anthropic says half the Fortune 500 are customers, yet that spending is only about 0.1 percent of those customers' operating expenses. Per user, it works out to around $8 a month if operating expenses exclude cost of goods sold, and under $50 a month if they include it. The top 5 percent of Claude Code users spend an average of $18,000 a year, while other users spend well under $75 a month.
Getting from here to the $1 trillion in revenue that Dario Amodei has said he can see requires white-collar work, Brookhart said. AI has to move from augmenting people to saving labor or replacing it in some areas. He cited Microsoft's buyout offers to employees as an early sign. The enterprise budget cycle for the fourth quarter is the first one these labs will go through in any meaningful way, since ARR only began to accelerate at the start of the year.
Anthropic is building bridges in the meantime. It is putting out more vertical products, including Claude for financial services and Claude for legal, and running a cyber program that gives away 25,000 in credits per month to participants. Cyber could become a big product in 2027, with eight-figure contracts at large enterprises that can scale quickly given the number of security incidents. Bio is another candidate. Brookhart described his own world as an example. Claude could not build a good financial model six months ago, and now an analyst wonders how many junior analysts they still need.
The interviewer worried about a less idealistic competitor. It could skip the machine-god project, embed forward-deployed engineers across industries, automate chunks of their business and sell them high-margin inference. A good share of the Fortune 500 has seen no return on AI spending so far, so a firm that is better at deployment could get ahead of Anthropic. Brookhart agreed it is an interesting risk and said it is the reason Anthropic is verticalizing. He listed two others. One is a mismatch of timing and adoption, in which revenue falls short of the pace of the chase for AGI. The other is open-source models matching the frontier in coding and forcing token prices down.
Renting from the people you compete with
AWS, Azure and Google Cloud are suppliers to Anthropic and competitors of it at the same time, Google especially. The S-1 warns that these relationships could turn. The relationship with SpaceXAI could grow precarious as well, since Grok is a rival.
Brookhart said both Anthropic and OpenAI are heading toward rent-to-own arrangements or special purpose vehicles, as some financial sponsors use, rather than outright renting. Google is also selling its chips externally. Anthropic has orders of 3.5 to 4 gigawatts of TPUs, with more in the pipeline. The SPV buys the chips and Anthropic rents them from it.
The cloud providers do well from the relationship. About 48 percent of Anthropic's business comes through Bedrock, Foundry and Vertex, where customers pay through an existing cloud relationship and never talk to Anthropic's sales force. The clouds take a revenue share, which Brookhart put at 16 percent. These token-as-a-service businesses are lifting AI margins at the clouds, he said, and both Azure and AWS had very strong second quarters, AWS in particular on the back of Anthropic's growth on Bedrock.
Amazon, he thinks, is happy to serve many customers. His estimates give it the least training compute of any hyperscaler except possibly Oracle. Microsoft, under its new agreement, can monetize OpenAI models without paying OpenAI anything for them. The interviewer saw Microsoft's job cuts as a glimmer that it might compete again.
The neocloud question and a possible glut
The past few weeks have been tense for neoclouds. Musk predicted a guppy-like extinction event. Brookhart noted that Musk has reasons to say so while promoting his own data-center plans, which include putting them in space. His own view is that CoreWeave, Nebius and anyone backed or backstopped by a major lab, a hyperscaler, Nvidia or Broadcom will be fine. Trouble lies in the longer tail, including ex-crypto miners that are not good at running clusters. Hyperscalers account for 65 percent of capital expenditure and new compute, so 35 percent of the market lacks their balance sheets. Financing is very difficult for those players.
The big four hyperscalers' capital expenditure is growing from $1.2 trillion to $2 trillion, though he said they do not need to raise outside money over the next two years. Data centers will get built despite the political backlash and permitting problems. The shell is cheap and the chips are most of the spending, so if anyone overbuilds, they can leave the buildings chipless.
The worry for 2028 is a glut. As each inference megawatt earns more revenue, the industry needs less compute for the same revenue. Whether there will be enough revenue to justify the ROI on more than $2 trillion of capital expenditure is, in Brookhart's words, a live question.
Recursive self-improvement
The darker reading connects those financial pressures to the company's stated method. The filing says that "more capable models help our researchers discover these improvements faster; in turn, more powerful compute multipliers enable us to train even more capable models and deploy them more efficiently." Claude is already building Claude under human direction.
The risk factors admit a catch. Using AI to develop AI "may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures." One reading is that the business depends on building new models fast, using work the company may not fully understand, so it has a commercial reason to do things that could be risky to everyone. That would explain the unusual language in the same section about catastrophic harm and the development of forms of sentience. The interviewer noted that the S-1 even says Claude might be conscious.
If dramatic recursive self-improvement that outruns human supervision is not possible, that reading says, Anthropic will find it much harder to meet its compute commitments and turn a profit. The author does not claim to know whether it is possible.
Brookhart's version of the same logic is friendlier. As long as inference gross margins stay at 85 percent or more, a company that believes it is heading toward AGI can afford to push more of its compute into training. He expects Anthropic to want a larger training share over time, with the fleet possibly going 80 to 85 percent to training and research. The filing's profit curve shows a good business model, he said, but with a warning attached that margin leverage should not be expected to continue at this level.
The blank pages
The two readings agree on the facts and split on the temperament. One sees a company that has to grow because its contracts assume it. The other sees a company with plenty of cash, rising efficiency and an enormous, barely tapped market. The first view needs recursive self-improvement or something like it. The second needs only that the customers show up.
In the version of the filing that was read, the page-64 totals for purchase commitments and lease agreements are blank, and so is the page-137 figure for what the company expects to spend before December 31, 2030. Anyone modelling the next few years has had to estimate the very numbers the company will eventually have to pay.