Anthropic's leaked draft S-1: $518B of compute nobody can cancel

The S-1 is not publicly available. I verified this myself. I searched SEC EDGAR directly on 2026-10-04 and found no Anthropic registration statement:

• EDGAR company search for "anthropic" returned "No matching companies."
• URL: https://www.sec.gov/cgi-bin/browse-edgar?company=anthropic&CIK=&type=S-1&dateb=&owner=include&count=40&action=getcompany
• EDGAR full-text search for S-1 / S-1A / 424B4 filings containing "Anthropic" between 2026-06-01 and 2026-10-02 returned only other companies' filings (SpaceX, Gloo Holdings, Oura, QumulusAI) that mention Anthropic as a customer or counterparty. No Anthropic-issued registration statement.

So what is the document? FACT: Anthropic confidentially submitted a draft Form S-1 on 2026-06-01. FACT: Anthropic has not publicly filed it. FACT: Reuters, the Financial Times and The Information obtained and reviewed copies under confidentiality; Fortune described it as a "leaked" draft.

This distinction matters and is the single most important framing correction:

Every financial figure below is therefore company-prepared text inside a confidential SEC submission relayed by journalists who read it. It is not independently audited by me, and it is not publicly verifiable by any third party at the time of writing. TechCrunch noted that scanning the SEC's public database showed no prior IPO prospectus containing "existential risks to humanity", implying this language has no clear public precedent.

--

On the evening of 2026-09-28, Reuters published an exclusive based on a confidential draft Form S-1 that Anthropic had submitted to the SEC on 2026-06-01. Two further Reuters stories on 2026-09-29 unpacked the compute contracts and the safety risk factors.

FACT (as reported from the document):

• Anthropic targets a public offering that could value it above $2 trillion.
• It commits to at least $518 billion of cloud, compute and infrastructure spending over about a decade with six partners, of which about 80% is non-cancelable or payable regardless of usage.
• FY2025: revenue ~$4.6 billion; operating expenses ~$12.65 billion; operating loss $8.06 billion (widened from $2.98 billion in 2024); net loss ~$42 billion; cash, cash equivalents and short-term investments $20.28 billion at 2025-12-31.
• Compute and infrastructure cost $7.33 billion in 2025 — a roughly threefold increase on 2024 and more than half of total operating expenses.
• The prospectus warns that Anthropic's models could pose "catastrophic or existential risks to humanity".
• The offering is expected in November 2026, likely after the 2026-11-03 US midterm elections.

Anthropic declined to comment in every Reuters story.

Before this story, no frontier AI lab had ever let the public see audited-style financial statements in an SEC registration statement. The public picture of frontier-lab economics was: large private valuations, circular vendor/investor funding, and opaque compute spend.

This disclosure changes the conversation in three specific ways:

1. Compute obligations became a disclosed, contractual, non-cancelable liability. For the first time a frontier lab has published the names and amounts of its compute counterparties and the cancellation terms.
2. The gap between revenue growth and unit economics became measurable. $4.6B revenue against $12.65B of operating expense is not a rounding error; it is a business model with a stated, audited-style cost curve.
3. An existential-risk warning entered an SEC-facing document. That converts a philosophical position into a potential legal-liability disclosure for a company seeking public shareholders.

Before (through 2026-09-27)

• Frontier-lab finances known only through private rounds: Anthropic's May 2026 Series H at a reported $965 billion post-money valuation; OpenAI reported around $852B private.
• Compute costs treated as an abstract "capex race" narrative.
• Anthropic had a Public Benefit Corporation structure and a Long-Term Benefit Trust, but no public capital-structure disclosure.
• Safety positions lived in blog posts and essays (e.g. Dario Amodei's "We Must Pace the Frontier").

Change (2026-09-28 / 2026-09-29)

• Confidential S-1 content disclosed by Reuters, FT and The Information.
• A dated, itemised commitment table appears in the press: Google ≥$111.1B; Amazon $110B; Microsoft $31.4B; Broadcom ~$161.2B equipment leases; xAI up to $84.5B; AMD >$20B capacity plus up to $5B of equity.
• Risk factors occupy roughly 80 of 261 pages of the main body — nearly double the 48 pages describing the business. Reuters benchmarks this against SpaceX's June 2026 IPO prospectus: about 38 of 277 pages.

After (implied state)

• Any buyer, partner or journalist can now anchor on concrete figures rather than venture narrative.
• Anthropic's suppliers (Google, Amazon, Microsoft, Broadcom, xAI, AMD) hold obligations that survive an Anthropic demand shortfall.
• Anthropic's own governance model — a "Founder LLC" holding a single Class F share with 50.1% of voting power over key matters — becomes legible and contestable.

The commitment structure, as reported from the prospectus:

Mechanics worth understanding:

• "Pay the difference" is the crux. This is a take-or-pay structure. Anthropic reserves capacity; if it under-uses it, it still owes the delta. So compute is a fixed cost, not a variable cost, for roughly 80% of the committed total.
• The four largest lines sum to ~$413.7B, which is the ~80% figure. The xAI and AMD arrangements are the genuinely flexible portion.
• The stated rationale is scarcity, not demand. Anthropic says future demand for advanced AI systems is likely to exceed supply and will be "limited principally by the availability of compute". The commitment is a bet that demand materialises.
• Counterparty role-conflict is disclosed by the company itself. Google, Amazon and Microsoft are simultaneously investors, customers, cloud providers, distributors and competitors. The filing says their incentives "may not be fully aligned" with Anthropic's, and warns that if third-party compute is "curtailed, repriced, or terminated" results could be adversely affected.
• Net loss vs operating loss. The ~$42B net loss includes roughly $34B of non-cash accounting charge reflecting the increased estimated value of financing instruments that could later convert into Anthropic shares. This is not operating cash burn. The economically meaningful 2025 figure is the $8.06B operating loss.

• It is the first real balance sheet for frontier AI. Enterprises and investors can finally see that Anthropic's 2025 cost base was ~2.7x its revenue, and that compute alone was 58% of opex.
• It reframes "AI economics" from narrative to contract. "Compute is expensive" becomes "this lab has $518B of take-or-pay obligations, 80% of which it cannot cancel."
• It quantifies a supply-chain concentration risk in AI that most buyers never considered: their model provider's suppliers may also be their provider's competitors, and owe nothing to the buyer.
• It puts an existential-risk statement into a liability-creating document in a way that could shape how frontier labs write future registrations.
• It is a systemic valuation event. Per NYT DealBook (Andrew Ross Sorkin), Amazon and Google carry large Anthropic stakes at high valuations on their balance sheets. A discounted or delayed Anthropic listing would "set off a cascading mark-to-market event" and drag OpenAI's implied valuation down, threatening megacap backers including Nvidia. Sorkin linked this directly to "circular financing" fears.

• Independent cost modelling of frontier AI. Teams can now build a top-down cost model for a frontier lab and test unit-economics assumptions against disclosed actuals.
• Contract-term negotiation leverage. Buyers now know that Anthropic's compute is pre-committed and its customers are not. That asymmetry is publicly visible for the first time.
• A new diligence category. "Compute commitment coverage" — reserved capacity versus expected demand — becomes a line item in AI procurement and vendor-risk review.
• A precedent for safety disclosure in capital markets. Expect other labs facing public listings to draft existential-risk risk factors with Anthropic's language as a template.
• A governance template to contest. The Class F / Founder LLC / Long-Term Benefit Trust structure is now a documented, citable design that public-market activists and ESG investors can point at.

1. What exactly is the contractual payment schedule? $518B spread over a decade averages ~$52B/year, but Reuters notes the agreements "do not require even payments." What are the annual peaks?
2. What are the 2026 gross margin and true inference cost per customer?
3. Who are the two customers making up ~25% of 2025 revenue — and are they suppliers?
4. What is the dollar amount behind "adjusted operating profit" in Q2 2026, and what is excluded?
5. Will Anthropic publish the safety compute allocation as a standing disclosure, given it did not in the filing?
6. How is the $34B financing-related accounting charge settled — dilution mechanics and timing?
7. Does the Class F 50.1% voting block face proxy-adviser or index-inclusion consequences?
8. Will the ~$100B+ raise size and the >$2T target survive a market in which Oura just postponed?
9. Does OpenAI's prospectus (expected early 2027) match Anthropic's disclosure depth, and does OpenAI accept Anthropic as the valuation benchmark?
10. How much of the $518B is genuinely fixed if Anthropic's demand model is wrong — i.e. what is the realistic downside case?

FACT — scheduled/mechanical:

• Anthropic must make the registration statement public at least 15 days before the investor roadshow begins (Implicator.ai, on the confidential-submission process).
• Reuters reported the debut is likely to land after the 2026-11-03 US midterm elections.
• OpenAI is expected to list by early 2027, per media reports.
• The 2026-11-03 US midterm elections are a fixed date.

PREDICTION — clearly labelled as such, not fact:

• The public S-1 will land in late 2026 (Q4), with the listing shortly after the midterms. The timeline has already slipped repeatedly: "after Labor Day" → mid-October → November → after the November 3 election. Assuming no further slip is optimistic.
• Gross margin will be disclosed in the public filing, because the SEC review process makes a complete income statement effectively unavoidable. This is the single most valuable missing number.
• At least one of the two ~25%-of-revenue customers will be identified or characterised in the public version, because the concentration disclosure becomes a liability once shares trade.
• Anthropic will either raise above $2T or reframe. Given Oura postponed on 2026-09-29 for market "uncertainty" and AI/chip stocks had sold off, a soft or postponed listing is a live outcome. If it prices below $2T, expect the Sorkin-predicted mark-to-market cascade across Amazon, Google, Nvidia and OpenAI's implied valuation.
• The existential-risk risk-factor language will be copied. It is quotable, defensible and differentiating; other labs listing will face pressure to match it.

The headline number is the wrong headline. "Anthropic wants a $2 trillion valuation" is a target attached to an expectation. The story is the $518 billion: a frontier AI lab has told the SEC, in writing, that it has signed up to spend roughly 80% of that regardless of whether anyone uses the capacity — while telling the same document that its own products may pose catastrophic or existential risks to humanity.

Those two statements are not contradictory, and that is the point. They are the actual shape of frontier AI in 2026: a company that must spend whether demand materialises, on infrastructure it does not own, supplied partly by its own competitors, to serve customers who can walk away, in order to stay ahead in a race whose leaders are publicly warning that the race itself may be dangerous.

Three editorial rules for covering this story:

1. Never write "Anthropic filed with the SEC." It confidentially submitted a draft on 2026-06-01 and has not publicly filed. Until the public S-1 appears, every number in circulation is press-relayed draft text.
2. Never quote the $42B net loss as the cost of running the business. ~$34B of it is a non-cash charge on financing instruments. The operating loss was $8.06B.
3. Never present ">80% non-cancelable" as a scandal in itself. It is a rational response to the company's stated thesis that demand is supply-constrained. The genuine story is the asymmetry: the cost side is contracted and non-cancellable; the revenue side is not.

The durable lesson for any reader: for the first time, the economics of frontier AI are legible, and they do not look like a software business. They look like a capital-intensive infrastructure undertaking with software margins still to be proven — carrying a self-declared, SEC-filed warning that its own success condition may be dangerous.