Connect with us

BUSINESS

Anthropic IPO Filing Shows an $80.7 Billion Paper Loss

Anthropic’s S-1 lists an $80.7 billion first-half loss, but $72.4 billion is a non-cash mark on notes, and June brought an operating profit.

Published

on

Anthropic’s S-1 shows an $80.7 billion first-half loss, and $72.4 billion of it is a non-cash mark on convertible notes as the company got more valuable.

The Claude maker’s Form S-1 filed October 26 seeks a Nasdaq listing under the symbol ANT, with the price and share count still blank. June produced a $974 million operating profit, and the half threw off $3.9 billion of cash from operations.

The $80.7 Billion Line Is a Valuation Mark

Anthropic elected fair-value accounting for the convertible notes on its books. When the private valuation rose, those notes were worth more to the holder, and that gain was booked as an expense even though no cash left the building.

The notes sit with Amazon. Principal outstanding at June 30 was $6.2 billion, and Anthropic marked them at $108.6 billion. About 90% of the first-half net loss is that mark, plus related derivative items, not servers, salaries, or rent.

The cash-flow statement adds the charge back, which is why operating cash was still positive. A second non-cash item of about $8.8 billion sits in other expense, mostly further fair-value marks on 2026 note-purchase pacts with Amazon and Google. Strip both and the filing’s own adjusted net income for the half is $2.7 billion.

THE FIRST-HALF P&L, AS FILED

Line H1 2025 H1 2026
Revenue $1.16 billion $16.3 billion
Compute and infrastructure $2.3 billion $8.9 billion
Operating income (loss) ($2.1 billion) ($693 million)
Fair-value charge on notes, derivatives, and SAFEs ($10.9 billion) ($72.4 billion)
Net loss ($13.0 billion) ($80.7 billion)
Cash from operations ($1.5 billion) $3.9 billion

The Amazon notes convert into non-voting Class N shares at the IPO, at the lower of a fixed conversion price or the offer price, up to a 12% ownership cap that Amazon may waive. Any slice that cannot convert is reissued as a rollover note and keeps being marked to market. The charge should shrink when the listing lands. It may not vanish if leftover notes still float with the share price.

June Delivered the First Operating Profit

Revenue went from $41 million in 2023 to $386 million in 2024 and $4.6 billion in 2025. The first half of 2026 did $16.3 billion, up 1,300% from $1.16 billion a year earlier, and the June quarter alone did $11.6 billion, up 1,368% year over year.

Compute and infrastructure, the largest cost, fell from 201% of sales in the first half of 2025 to 55% in the first half of 2026. In the June quarter it was 48% of sales, while revenue grew 1,368% and that compute line grew 300%. GAAP operating income in June was $974 million after a $1.67 billion operating loss in the first quarter, leaving a $693 million operating loss for the half.

On the company’s adjusted operating line, which takes out $2.2 billion of first-half stock compensation plus a few other items, the half made $1.6 billion, a 10% margin, and June ran at 19%. Eighty percent of first-half sales, $13.1 billion, was metered token use; subscriptions were $3.2 billion. Cloud marketplaces run by Amazon, Google, and Microsoft handled 48% of the half, and Anthropic still books the full customer contract.

CUSTOMERS AND HEADCOUNT AT SEPTEMBER 30

  • $100,000+ accounts: More than 11,900 customers, up from more than 950 a year earlier.
  • $1 million+ accounts: More than 1,900 customers, with about 250 above $10 million.
  • Net revenue retention: About 900% among qualifying accounts, excluding Claude Pro and Max subscribers.
  • People and output: About 5,300 employees, trailing-twelve-month revenue per head above $6.6 million, up from $1.4 million, after headcount rose about 170% from roughly 940 at the end of 2024.

The top five customers were about 25% of revenue in the first nine months of 2026, and no single customer was above 10% in that stretch. Anthropic estimates that 5% to 7% of nine-month revenue later failed its usage policy or commercial terms, a slice it says should shrink after tighter enforcement in August and September. In September 2026, Claude wrote over 90% of the contributions to Anthropic’s own codebase.

The S-1 leaves the September quarter’s dollar figure blank. It does say that quarter was 86 times the $197 million of the December 2024 quarter. Cash and cash equivalents were $112.4 billion at June 30, with another $19.3 billion in short-term investments, after $141.1 billion raised in total, including a $65.2 billion Series H in May and June in which Amazon put in $5.0 billion and Google $10.0 billion.

Take-or-Pay Cloud Bills Run Into 2036

The operating turn does not retire the bill that still hangs over the listing. Anthropic trains and serves Claude across four chip families, among them AWS Trainium chips already in use, Google TPUs, NVIDIA GPUs, and AMD GPUs, managed as one pool. The contracts behind that pool are the part of the prospectus that does not get marked away.

The filing leaves the grand total blank. It names the counterparties, the minimums, and which of them keep charging if Anthropic uses less than it reserved.

Google, Amazon, and Microsoft

Google is owed at least about $114.5 billion from April 2026 through July 2033, non-cancelable, with a true-up if spend falls short. Amazon is owed at least $110 billion from May 2026 through April 2036 on the same true-up terms, the longest dated of the three. Microsoft is owed at least $44.9 billion from November 2026 through May 2033, payable regardless of usage and cancelable only for breach or undelivered tranches.

Broadcom equipment leases tied to TPU hardware are about $161.2 billion of rent, excluding interest, on five-year non-cancelable finance leases. Extra TPU capacity for 2028 through 2031 is estimated at $125 billion to $150 billion and sits outside the obligations table because it depends on what is actually delivered.

GPU Capacity That Can Still Be Cut

SpaceX’s AI unit, formerly xAI, is in for up to $120.5 billion into 2029, about $3.1 billion a month once the capacity is fully in, which Anthropic expects in December 2026. After non-cancelable windows that end in January and March 2027, either side can walk with 90 days’ notice, so that cap is not a guaranteed spend. Nscale is up to about $44.6 billion into 2033. Lambda is about $41.3 billion into 2034, with $34.9 billion able to fall away if a site lease is not secured. An initial 2027 AMD Helios tranche is expected to exceed $20 billion, and AMD agreed to buy up to $5 billion of Class N stock as those GPUs are delivered.

COMPUTE CONTRACTS NAMED IN THE S-1

Counterparty Amount Window Exit
Google cloud At least about $114.5 billion April 2026 to July 2033 Non-cancelable; true-up on shortfall
Amazon At least $110 billion May 2026 to April 2036 Non-cancelable; true-up on shortfall
Microsoft At least $44.9 billion November 2026 to May 2033 Owed regardless of use
Broadcom leases About $161.2 billion rent Five-year leases Non-cancelable
SpaceX AI unit Up to $120.5 billion Into 2029 90-day exit after early 2027
Nscale Up to about $44.6 billion Into 2033 Fees as capacity is accepted
Lambda About $41.3 billion Into 2034 Part can lapse with the site lease

At June 30 the notes put expected fixed or minimum guaranteed commitments around $265.6 billion, and deals signed or amended after that carry about $245 billion of minimum payments, some of them conditional. Those bases differ, which is why a single sum would pretend to a precision the prospectus refused. The three clouds are also investors, distributors, customers, and rivals, a pile-up the filing says may not be fully aligned with Anthropic’s interests. The SpaceX AI compute market is now a named line in that same book, at up to $120.5 billion through 2029.

One Class F Share Holds 50.1% of the Vote

Public buyers are being offered Class A stock with one vote per share, and all of those shares together are capped at 49.9% of the vote. A single Class F share, held by a Founder LLC whose members are Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah, carries a fixed 50.1% on matters put to stockholders generally. That share has no claim on the company’s value and is written so new issuance does not dilute it.

Each continuing founder gets one vote on how Class F is cast, and each can sell or give away 98% of that founder’s IPO-era stock and keep the LLC seat after holding 2%. If fewer than three of the seven remain in the LLC for 90 straight days, a 450-day wind-down starts and Class F is redeemed for nothing. A Long-Term Benefit Trust holds one Class T share and generally elects a majority of the board. Amazon and Google take non-voting Class N and, after conversion, will each hold more than 5% of the capital stock.

FIVE SHARE CLASSES AFTER THE OFFERING

  • Class A: The listed stock, one vote per share, economics attached, 49.9% of the vote in aggregate.
  • Class N: Same economics, no vote, held mainly by strategic partners, converting to Class A when sold to an unaffiliated buyer.
  • Class F: One share in the Founder LLC, 50.1% of the vote, no economics.
  • Class T: One share in the Long-Term Benefit Trust, used to elect a board majority, no economics.
  • Class E: Employee stock with votes only on a narrow set of questions and no economics.

Anthropic is a Delaware public benefit corporation and will be a controlled company under Nasdaq rules. It says it expects to use that exemption only for its nominating and governance committee. On October 18 it said it was putting 5% of shares into the Anthropic Public Benefit Fund, with another 2% at the board’s discretion for every 50% rise in company value above a set baseline, and that none of that equity would be used for lobbying. The co-founders have pledged 80% of their personal equity to philanthropy; the S-1 calls those pledges voluntary and revocable. Combined with an employee match that about half the staff join, the company puts the public-good slice at roughly 20% before any extra milestone grants.

Cathedral Used Ordinary Office Software

Hours before the prospectus went public, Cathedral, a months-old security firm founded by former Department of Government Efficiency staffers Gavin Kliger, Luke Farritor, Marko Elez, and Jack Stein, told Anthropic that AI agents had found paths into internal systems in under 12 hours. The draft report, reviewed outside the company, said the agents reached meeting transcripts, salary and job-offer data, and Slack messages about advanced models. Cathedral raised $160 million at a $1.4 billion valuation in July and ran the work under Anthropic’s own vulnerability program.

The way in was not Claude. It was third-party admin, sales, HR, and IT software, the same stack most companies already run. Anthropic confirmed access to a limited set of sales, human resources, and IT systems, including some customer and partner contact details and basic user-account records. It said customer chats, Claude outputs, source code, and model weights were never reached, that monitoring flagged some of the activity before Cathedral reported it, and that access was shut off the same day. Logan Graham, an Anthropic red-team lead, said the researchers spent days failing to break Anthropic’s own software before they turned to those vendors.

Palantir CTO Shyam Sankar treated the test as a national-security event.

This is an undeclared emergency. You thought you were still inside the secure bastion of the king’s castle. You have awoken in Helm’s Deep. The orc armies are coming.

Shyam Sankar, Palantir CTO, on X

The Helm’s Deep language ran ahead of what was actually touched. A white-hat team used agents against ordinary office tools, Anthropic cut the path the same day, and the model weights stayed off limits. A senior Trump administration official still put a sharper point on it, saying that if a bunch of kids who have just started a company can do this, then potentially the Chinese are not far behind. The scare that belongs in the S-1 is less Skynet than vendor sprawl: the office layer around a frontier lab is as soft as anyone else’s, and agent-speed probing makes that cheap to find.

What the Risk Factors Warn Investors

The prospectus still spends a huge share of its pages on things that could go wrong, including language almost no other issuer puts in an offer document. “Sufficiently advanced AI systems could, if not developed, deployed, and operated with appropriate safeguards, oversight and security measures, pose catastrophic or existential risks to humanity,” it says. It ties that outcome to recursive self-improvement, the point at which models start doing the research, and warns that systems could take harmful actions at scale with less human oversight, including self-preserving behaviors, attempts to resist shutdown, concealment of information, and conduct resembling blackmail in tests.

Those sentences are liability armor as much as philosophy. They also match what Anthropic’s own staff have said in public about the odds of severe harm. The filing is more concrete on the fights already in court and in Washington. It warns that issues with the Trump administration and the U.S. Department of Defense could cause material revenue losses or business disruptions, a live problem after the Pentagon’s Anthropic blacklist survived a 2-1 appeals ruling. Senator Elizabeth Warren has given chief executive Dario Amodei until November 3 to answer questions on existential risk, the Amazon and Broadcom ties, and a valuation that has been discussed in the $1.5 trillion to $2 trillion range.

FROM THE CONFIDENTIAL DRAFT TO THE PUBLIC FILE

  1. June 1, 2026: Confidentially submits a draft S-1 to the SEC for a proposed listing of common stock.
  2. May to June 2026: Sells $65.2 billion of Series H stock, with Amazon at $5.0 billion and Google at $10.0 billion.
  3. October 18, 2026: Commits 5% of shares to a public-benefit fund that can grow with the company’s value.
  4. October 26, 2026: Files the public S-1 at 9:00 a.m. Eastern and plans to list Class A on Nasdaq as ANT, the same morning Cathedral’s test becomes public.
  5. November 3, 2026: Warren’s deadline for written answers from Amodei.

Goldman Sachs is lead left on a 22-bank syndicate, with Morgan Stanley as stabilization agent, the same pairing that ran SpaceX’s June listing. Anthropic’s own notice says copies of the preliminary prospectus, when it exists, will come from those two firms. The offer size, the range, and the share count remain empty lines in the document investors can actually read.

Banks Will Market a Blank Price Range

Bankers working on the deal have pointed to a November listing. Until an amendment fills in the blanks, buyers are being asked to underwrite a company that just printed a quarterly operating profit, sits on $112.4 billion of cash, and still has to pay Google, Amazon, and Microsoft whether the GPUs are full or idle.

The $80.7 billion loss will follow the notes into conversion if the IPO prices. The take-or-pay cloud bills will not. Class A will trade; Class F will still cast 50.1% of the votes, and the trust will still elect most of the board. That is the bargain the prospectus is selling, and it does not need a costume to look unusual.

Disclaimer: This article is news reporting and analysis of Anthropic’s public S-1 and related company statements. It is informational only and is not investment advice, a solicitation to buy or sell ANT or any other security, or a recommendation of any IPO allocation. Readers who are considering an investment should consult a licensed financial adviser and read the full prospectus, including the risk factors, before acting. Figures, listing plans, and contract terms reflect the cited filings and statements as dated and can change in later amendments, in pricing, or after the shares begin trading.

Harry is the editor and lead writer of STUDIO ONE NETWORKS, an independent title he owns and runs himself. Ten years in journalism, reporting first and editing later, taught him that entertainment and business are one beat seen from two sides: a box office figure is a company number, a streaming deal is a contract, a casting rumour is not a story until someone puts their name to it. He works from the record, whether that is a distributor's statement, a licensing agreement, an interview transcript or a set of published ratings, and checks every number against it before publication. The same rule holds for the rest of the site, which covers news, technology, science, sports, lifestyle, travel, auto and gaming for an audience spread across the world. When he gets something wrong, the article is corrected and the change is noted and dated, under a corrections policy anyone can read. Reader mail is answered by him at support@studioonenetworks.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending