Inside Anthropic's S-1
Anthropic filed to go public today. I read the prospectus and charted its financial tables. The filing shows how a company that sells intelligence by the token makes money, where the money goes, and who controls it.
Summary
Anthropic filed its Form S-1 with the SEC this morning. I pulled the financial tables out of the filing and into Hex, then drew the charts below from those numbers. Nothing here is a forecast. It is the company as it describes itself to regulators, with the arithmetic shown.
Three things stood out. The business turned. Revenue was $16.3 billion in the first half of 2026, 14 times the same period a year earlier, and Q2 was the first quarter with an operating profit. The $80.7 billion net loss is almost all paper. It comes from marking convertible notes, mostly Amazon's, to Anthropic's own rising valuation, and it stops when those notes convert at the IPO. And public shareholders will own a company they cannot control. One share held by the seven co-founders carries 50.1% of the vote, and an independent trust picks most of the board.
The rest of this post walks the filing in the order I read it: how Anthropic charges, what it costs to run, where the compute comes from, who funded it, and what the amended filing still has to fill in.
Business model
Anthropic sells one thing in several packages: access to Claude, its family of AI models. Founded in 2021, it had no revenue until it launched commercially in the first quarter of 2023. By the second quarter of 2026 it was booking $11.6 billion a quarter.
The prospectus calls its main pricing model "pay-for-intelligence." Customers pay per token, the unit of text a model reads or writes, at a price set by the model they pick. There are four model classes, each a different trade between capability, cost, and speed: Mythos (which includes Fable), Opus, Sonnet, and Haiku. Customers use the bigger models for coding, agents, and hard reasoning, and route simpler tasks to the cheaper ones. The model chosen and the volume of use drive the bill.
A second, smaller stream is subscriptions to Anthropic's own apps, sold to businesses and individual power users. Terms run 12 months or less. When subscribers use up their plan, they can move to a higher tier or buy more usage by the token. A growing number of enterprises also sign annual or multi-year contracts with minimum spend in exchange for fixed prices. That money still shows up as usage revenue.
The sales motion runs from the bottom up. A developer or an individual tries Claude, brings it to work, and the account grows from one workflow to many. By the time Anthropic's direct sales teams reach a customer, Claude is often already in use. Claude is also sold on Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic keeps the contract with the customer in those deals and pays each cloud provider a fee. That matters for any comparison with OpenAI, which I get to at the end.
Show the numbers
Switch between the three periods and the change is plain. In 2024, Anthropic spent $3.4 billion to earn $386 million, and most of the money came from investors. In the first half of 2026 revenue nearly covered everything: $16.3 billion against $17.0 billion of expenses. The $693 million gap is smaller than the $2.2 billion of stock-based compensation in the same period, which is why Anthropic reports an adjusted operating profit of $1.6 billion.
Revenue
Revenue grew from $41 million in 2023 to $386 million in 2024 and $4.6 billion in 2025. Then it sped up. Q1 2026 revenue was $4.7 billion and Q2 was $11.6 billion, 144% more in one quarter. Year over year, Q2 revenue grew 1,368%.
Show the numbers
Four in five dollars come from usage. Consumption-based revenue was 75% of the total in 2024, 83% in 2025, and 80% in the first half of 2026. Subscriptions grew faster than usage over the past year, rising from 15% to 20% of revenue between the first halves of 2025 and 2026. The U.S. supplies about seven in ten dollars.
Customers
The S-1 does not count customers in total. It counts them by how much they spend over the trailing 12 months.
Net revenue retention of about 900% means that the customers Anthropic had a year earlier now spend roughly nine times as much. Accounts also widen across products. In Q3 2026, more than half of the $100,000-plus customers used both the developer platform and at least one of Anthropic's own apps, up from about 30% in Q4 2025.
The filing includes customer case studies, filed as images rather than text. Each logo below opens that company's page in the filing on EDGAR.
Run-Rate Revenue, the figure Anthropic uses for its annualized pace, is defined in the S-1 but left blank for Oct. 31, 2026. It annualizes 28 days of usage plus current subscriptions, so it swings more than reported revenue. Expect the number in an amended filing.
Costs
Anthropic reports costs two ways. The income statement groups them by what was bought: compute and infrastructure, personnel, sales and partnerships, and other. The notes regroup the same dollars by what they were for: serving customers, research, selling, and administration.
Compute is the largest cost. Anthropic treats it as one pool that serves every job: training new models, research, serving customers, and its own staff's use of Claude. It buys most of that compute 12 to 18 months ahead on multi-year contracts. When demand outruns the plan, it buys short-term capacity at a higher price. Expense starts when capacity comes online, before the revenue it will support. The company warns that margins will swing for that reason.
Show the numbers
The functional view separates serving from building. Cost of revenue, the compute and staff spent serving paying customers, was $2.4 billion in the first half of 2026 against $16.3 billion of revenue. That leaves an 85% gross margin. In 2024 the same measure was negative: serving customers cost more than they paid. Training spend lives in research and development, which was $9.7 billion in the half, or 60% of revenue.
Anthropic makes two efficiency claims. The first is about compute: Q2 2026 revenue grew 1,368% year over year while compute expense grew 300%. Anthropic credits training and serving improvements it calls "compute multipliers," plus software that runs Claude across four chip families. The second is about staff: trailing revenue per employee rose from $1.4 million to more than $6.6 million in a year while headcount grew about 170%. Claude wrote over 90% of code contributions to Anthropic's own codebase in September 2026.
Sales, marketing, and partnerships went the other way, from 6% of revenue in Q1 2025 to 18% in Q2 2026. The functional breakdown says selling costs include channel partner costs, and Anthropic pays each cloud provider a fee on the sales it makes through them. The 2025 total also includes a one-time $1.5 billion charge for the Bartz copyright settlement, booked in Q3 2025.
Net loss
Anthropic lost $80.7 billion in the first half of 2026 on $16.3 billion of revenue. Almost none of it was cash. The operating loss was $693 million, and operations produced $3.9 billion of cash.
The rest is accounting for the company's own rising value. Anthropic owes convertible notes, mainly to Amazon, that turn into shares at fixed low prices. Under GAAP, the company carries those notes at fair value. When Anthropic's valuation rises, the notes are worth more to the holder, and the increase runs through the income statement as a loss. The S-1 attributes the $72.4 billion charge primarily to that effect. A further $8.8 billion of other expense is mostly the recognition and fair value change of note purchase agreements.
The charge tied to the existing notes stops when they convert into stock at the IPO. The prospectus's pro forma columns, which remove it, are still blank in this filing.
Compute
Anthropic trains and serves Claude on four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Its own orchestration software moves work to whichever hardware is best for price and performance. That spreads supplier risk. The S-1 also notes that some of the cloud providers it relies on for compute are investors in Anthropic, customers of it, and competitors at the same time.
The company is moving from renting cloud capacity to a hybrid it calls Anthropic Direct. It leases chips itself and puts them in data centers that partners build and run to its specification. Data centers take longer to build but cost less than the chips inside them, so Anthropic can lock in buildings early and defer the bigger chip spending until demand is clearer.
Non-cancelable commitments at the end of 2025 totaled $54.6 billion. The largest single years are 2027 and 2028. The S-1 says commitments signed through September 2026 run as late as 2045. That total is left blank.
Equity and debt
Anthropic has funded its losses mostly by selling preferred stock. Investors paid $2.57 a share in Series A and $589.01 in Series H, 229 times more. Series H alone carries $65.2 billion of liquidation preference. The preferred stock converts to common stock at the IPO.
The debt is unusual. Anthropic has no drawn bank loans. Its notes come from its compute partners, and they are built to turn into shares. The lenders are also suppliers that Anthropic buys compute from.
Amazon's notes from 2023 and 2024, $6.2 billion of principal, convert into 1.84 billion non-voting Class N shares at $2.09 and $5.05 a share. Those notes sit on the balance sheet at a fair value of $108.6 billion, which is how a few billion dollars of borrowing turns into an $80 billion paper loss. The 2026 agreements with Google, Amazon, and Broadcom add up to $77 billion more that Anthropic can draw. None was drawn at June 30. Broadcom's $42 billion facility carries 9% interest.
A $15 billion secured revolving credit line from a bank group led by Morgan Stanley, signed in September 2026, runs to 2031. Cash and short-term investments were $131.7 billion at June 30, mostly from the $110.4 billion of preferred stock sold in the first half.
Control
Anthropic is a public benefit corporation, so its board must weigh its mission alongside shareholder returns. After the IPO it will have five classes of common stock. Two special classes, each held by a single owner, hold most of the control.
Public investors will own shares but will not run the company. The founders' single share casts a fixed majority of votes, and the trust picks most of the board. Nasdaq will treat Anthropic as a "controlled company," which exempts it from some board independence rules. Each of the seven co-founders also holds roughly 230 to 245 million Class A shares.
Dario Amodei's 2025 compensation was $18.0 million, mostly equity, on a salary of $654,167.
Risks
Risk factors fill about a third of the prospectus text: 87 of them across six categories. The largest group covers the business itself: staying at the frontier, getting enough compute and power, depending on partners who also compete, and the chance that a model causes real-world harm. Others cover copyright and data, regulation and export controls, the tax and accounting effects of the notes, and the governance structure.
Two labs
Google and Meta build frontier models inside companies that earn most of their money elsewhere. Anthropic and OpenAI are the two large independent labs whose main product is the model itself, and both are on the way to public markets. OpenAI filed confidentially in June 2026 and, according to Bloomberg, has pushed its listing to at least 2027.
| Anthropic | OpenAI | |
|---|---|---|
| Filing status | Public S-1, Oct. 26, 2026 | Confidential filing, June 2026 |
| Annualized revenue | $46.2B (Q2 2026 revenue × 4) | About $50B at end of September (Bloomberg) |
| Reported revenue basis | Gross of cloud partner fees | Net of partner share, per press reports |
| Audited financials public | Yes | No |
The Anthropic annualized figure is my arithmetic, not a number from the S-1. OpenAI figures are press reports, not filings.
The two companies also count revenue differently. Anthropic records the full price when Claude is sold through a cloud marketplace and books the cloud provider's cut as an expense. The Financial Times reported that this difference in method explains part of the gap between early estimates of OpenAI's revenue and its own figures. Until OpenAI files publicly, there is no like-for-like comparison.
What to watch
The S-1 describes a company whose economics changed in the past year. Serving customers became profitable at the gross margin level once revenue caught up with the compute already bought. Training remains the large, lumpy cost, and Anthropic says plainly that margins will dip each time a big block of new compute comes online. The net loss is mostly paper and goes away when the notes convert.
What decides the next few years sits outside this filing's numbers: whether Claude stays at the frontier, whether the 2027 and 2028 compute arrives on time and gets used, and how growth slows from 1,368% a year. The S-1 says growth "may decline significantly." The amended filing will add the price range, the share count, and Run-Rate Revenue.