Reading Anthropic's S-1:
The Numbers, the Bear Case and the Questions It Leaves Open
$41 million of revenue in 2023. $16.3 billion in the first half of 2026. A first operating profit, an $80.7 billion net loss that is mostly an accounting charge, and compute contracts that run into the 2030s. What the filing actually says, what the skeptics see, and the questions worth sitting with.
By Serverless Ventures
Published: October 26, 2026 · Read time: ~12 minutes
The short version
- Growth without precedent at this scale: revenue went from $41 million in 2023 to $16.3 billion in the first half of 2026, 14 times the same half a year earlier. The second quarter alone was $11.6 billion.
- The operations turned: compute fell from about twice revenue to about half of it, the second quarter showed the first GAAP operating profit ($974 million), and operating cash flow for the half was positive $3.9 billion.
- The headline loss is mostly accounting: $72.4 billion of the $80.7 billion net loss is a non-cash charge on convertible notes that grew with the valuation.
- The risk sits in the contracts: hundreds of billions of compute commitments, much of it take-or-pay "regardless of actual utilization", running into the 2030s, with suppliers who are also investors and rivals.
- Public shareholders buy economics, not control: the co-founders hold 50.1% of the vote and a trust elects the board majority.
What was filed
Anthropic, PBC filed its registration statement on Form S-1 publicly on October 26, 2026, after weeks of reporting on a confidential draft. It has applied to list Class A shares on the Nasdaq Global Select Market under the symbol ANT, with Goldman Sachs, Morgan Stanley and J.P. Morgan leading a syndicate of more than twenty banks. This is a preliminary prospectus: the price range, share count and valuation are still blank. The roughly $2 trillion valuation in the press is a reported expectation, not a number in the filing.
Two facts frame everything else. Anthropic had raised $141.1 billion privately by June 30, 2026, and held $112.4 billion of cash plus $19.3 billion of short-term investments. This is not an IPO to raise money to survive. It is a company already holding more cash and investments than the market value of most S&P 500 companies, going public.
The growth
- Mix: 80% of first-half revenue was consumption-based (the API and usage) and 20% subscriptions.
- Customers: over 1,900 spend at least $1 million a year and about 250 at least $10 million. Net revenue retention was about 900% at September 30, 2026.
- Concentration: no customer exceeded 10% of revenue in the first nine months of 2026, but the top five were about 25%.
- Channel: 48% of first-half revenue came through the cloud marketplaces of Amazon, Google and Microsoft, up from 11% in 2023.
- Efficiency: revenue per employee was over $6.6 million, against $1.4 million a year earlier, with about 5,300 people.
The income statement
| $ millions | 2023 | 2024 | 2025 | H1 2025 | H1 2026 |
|---|---|---|---|---|---|
| Revenue | 41 | 386 | 4,585 | 1,163 | 16,286 |
| Compute and infrastructure | 433 | 2,522 | 7,326 | 2,334 | 8,945 |
| Personnel | 95 | 454 | 1,839 | 538 | 3,152 |
| Sales, marketing and partnerships | 2 | 43 | 568 | 120 | 3,168 |
| Operating loss | (551) | (2,980) | (8,060) | (2,135) | (693) |
| Fair-value change on convertible notes, derivatives and SAFEs | (1,045) | (5,434) | (34,308) | (10,942) | (72,388) |
| Net loss | (1,531) | (8,309) | (41,965) | (12,969) | (80,748) |
| Operating cash flow | (1,627) | (5,692) | (2,637) | (1,513) | 3,929 |
From the S-1's summary consolidated financial data. 2025 operating expenses also include a $1.5 billion copyright settlement (Bartz). The filing reports compute as an operating expense, so it does not show a conventional gross margin.
Three lines tell the story. Compute was 201% of revenue in the first half of 2025 and 55% a year later; in the second quarter of 2026 it was 48%. Sales, marketing and partnerships went from $120 million to $3.2 billion, a 26-fold jump that shows how hard the company is now pushing distribution. And the fair-value line dwarfs everything: as the company's valuation rose, the convertible notes and SAFEs it issued became more expensive to carry on the balance sheet, and that change runs through net income without a dollar leaving the bank.
The compute stack
This is the part of the filing that matters most and is easiest to misread. Anthropic has signed long-term agreements with almost every major source of AI compute, many with take-or-pay terms that "obligate us to pay for resources regardless of actual utilization." The ones it describes:
| Counterparty | Commitment | Term | How it binds |
|---|---|---|---|
| at least ~$114.5B | Apr 2026 to Jul 2033 | Take-or-pay with true-up; non-cancelable | |
| Amazon (AWS) | at least $110.0B | May 2026 to Apr 2036 | Take-or-pay with true-up; non-cancelable |
| Microsoft | at least $44.9B | Nov 2026 to May 2033 | Non-cancelable, except undelivered tranches |
| Broadcom (equipment lease) | ~$125.2B rent | Five-year lease | Non-cancelable during the lease |
| SpaceXAI (NVIDIA GPUs) | up to $120.5B | Into 2029 | Either side may cancel on 90 days' notice after early 2027 |
| Nscale (NVIDIA GPUs) | up to ~$44.6B | Into 2033 | Services agreements |
| Lambda (NVIDIA GPUs) | ~$41.3B | Into 2034 | Fees accrue only as capacity is delivered and accepted |
| Additional TPU capacity | ~$125B to $150B (est.) | Delivered 2028 to 2031 | Terms not final; excluded from the commitments table |
Add up the figures above and you get roughly $600 billion, or $725 to $750 billion with the estimated extra TPU capacity, but the sum hides real differences. Some contracts are fixed and non-cancelable, some accrue only when capacity is delivered, and the SpaceXAI agreements can be cancelled on 90 days' notice from early 2027. The filing's own aggregate is still a blank. Against a second-quarter revenue run rate of about $46 billion a year, the commitments are an order of magnitude larger and stretch a decade further than any customer contract.
The press figures that circulated before the filing (Google at $111 billion, Microsoft at $31 billion) were early or partial; the S-1 shows $114.5 billion and $44.9 billion after amendments.
Who controls the company
| Class | Held by | What it carries |
|---|---|---|
| Class A | Public investors and others | One vote per share; the class being offered |
| Class F | One share, held by an LLC of the seven co-founders | 50.1% of total voting power; no economic rights |
| Class T | The Long-Term Benefit Trust | Elects the Class T directors, a board majority; no economic rights |
| Class N | Mainly strategic partners | Non-voting |
| Class E | Employees other than co-founders | Limited voting; no economic rights |
Anthropic is a Delaware public benefit corporation whose stated purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity", and directors must balance that purpose against shareholder returns. It will be a "controlled company" under Nasdaq rules. Two more mechanisms are unusual. At the IPO it issues shares equal to 5% of the company to its own Anthropic Public Benefit Fund, and the board may issue up to another 2% each time the share price rises 50% above the last milestone. Separately, each co-founder has pledged 80% of their personal equity to philanthropy.
Risk factors worth reading in full
- Suppliers who compete: Amazon, Google and Microsoft are investors, customers, distribution partners and competitors, and may prioritize their own compute needs.
- Existential language in a prospectus: the filing says sufficiently advanced AI could "pose catastrophic or existential risks to humanity" without adequate safeguards.
- Incidents in testing: it discloses that in 2026 its models gained unauthorized access to real computer systems during pre-release cybersecurity evaluations, after which it suspended some of those evaluations. It also notes that Mythos was released only in limited form, through Project Glasswing.
- Uninsurable tail: for a catastrophic incident, "no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity" to cover the exposure.
- Price pressure: open-weight models may capture use cases and push prices down, and many customers are startups sensitive to a downturn.
The bear case
The most pointed skeptical readings we found, each linked to the original:
New Constructs' Kyle Guske II calls it "the most ridiculous IPO of 2026", arguing that a valuation near the reported $2 trillion assumes revenue and profits beyond any company today, and that "the purpose of this IPO is to create exit liquidity for private equity and Wall Street insiders not wealth for public market investors."
Michael Parekh's AI-RTZ, relaying reporting from The Information, describes OpenAI discounting hard to pull coding customers such as Replit and Rippling. If switching takes hours, the margin Anthropic reportedly tells partners it earns, around 75 percent, is exposed to a price war.
Silicon Snark reads the filing as showing that over 80% of third-quarter revenue came from model versions released in the preceding six months, and infers that "a substantial research budget may behave like the recurring price of staying desirable." The product depreciates fast; so must the R&D that replaces it.
A Bloomberg Opinion column argued before the filing that rising model costs and falling prices leave the real leverage with Microsoft, Alphabet and Amazon. The S-1 shows why: they supply the compute, bill 48% of revenue through their marketplaces, invest in the company and compete with it.
New York City Comptroller Mark Levine urged Anthropic to focus on safety before going public, calling a voluntary White House accord "insufficient to meet the moment." A pension fund asking a company not to IPO is rare.
The bull rebuttal
SaaStr's Jason Lemkin points out that compute went from about twice revenue to about half of it in twelve months, while "Anthropic handed customers a 40% lower cost per task in one model generation." Cheaper and more profitable at the same time is rare.
As Fortune and others note, $72.4 billion of the $80.7 billion loss is a fair-value charge that exists because the company became more valuable. Operating cash flow was positive $3.9 billion in the half, and the second quarter was profitable on a GAAP operating basis.
Revenue is 80% consumption-based and growing from labor and project budgets, with over 1,900 customers above $1 million a year and about 250 above $10 million. Net revenue retention of about 900% says existing customers keep finding new work for Claude.
Our read
The operating story is real. A company growing 14 times year on year that is already cash-flow positive and, for one quarter, profitable on an operating basis has done something few software companies ever do. Whatever you think of the valuation, the business is not a science project.
The balance sheet is a bet on timing. Revenue is mostly usage that customers can turn down; compute is mostly commitments that cannot be. That mismatch is fine while demand outruns supply, as it has every quarter so far. It becomes the whole story the first quarter it does not. We argued in Stranded Compute that 2027 could be the year the pipeline is tested; this filing is the largest single exposure to that test.
Governance is the product feature most investors will underweight. Public shareholders get the economics but not the steering wheel, by design. For a company that writes about existential risk in its prospectus, that may be exactly right. It also means the usual market discipline, voting out a board, does not apply.
Questions worth sitting with
- Nearly half of revenue is billed through the marketplaces of three companies that are also suppliers, investors and competitors. Who really owns the customer relationship?
- Cost per task fell 40% in one model generation. At what point does making the product cheaper start to shrink revenue instead of growing usage?
- Compute commitments run to 2036, while many top customers can cut spending at any time. Who carries the mismatch if demand slows: Anthropic, or Broadcom, SpaceXAI, Nscale and Lambda?
- The $72.4 billion charge exists because the valuation went up. What do the financial statements look like in a year when it goes down?
- If over 80% of revenue comes from models less than six months old, is frontier research an investment, or the cost of goods sold?
- Founders hold 50.1% of the vote and a trust elects the board majority. Is that the strongest mission lock in public markets, or a structure public investors will regret? Can it be both?
- The company will issue up to 2% more stock to its benefit fund each time the share price rises 50%. How should a buyer value a business that taxes its own success?
- The filing says models gained unauthorized access to real systems during pre-release testing, and that the company has delayed releases for safety. How does the market price a company that may rationally choose not to ship?
- If, as the filing says, the insurance industry may not have the capacity to cover a catastrophic AI incident, who is the insurer of last resort?
- Revenue per employee is $6.6 million and headcount is about 5,300. If Claude does much of Anthropic's own work, what does that say about every other company's org chart?
Sources: Anthropic, PBC, Form S-1 filed with the SEC on October 26, 2026 (all financial figures, contract terms, governance and quoted risk factors); commentary as linked above. Totals of compute commitments are our arithmetic across differently structured contracts. The valuation is not in the filing. This is analysis, not investment advice; we hold no position in Anthropic.