Anthropic S-1: $33.2B Revenue in 9 Months, $104B Net Loss, $895B Compute Bill

Nov 10, 2026 · 7 min · Ajay Kumar

Anthropic, the company behind Claude, made its registration statement public on 26 October 2026 and filed an amended S-1 on 2 November that adds third-quarter results. The headline number is revenue of $33.24 billion in the nine months to 30 September 2026, up from $2.55 billion a year earlier, with $16.96 billion of that in the third quarter alone. The other headline number is a $104 billion net loss over the same nine months. Both are true, and the gap between them is mostly an accounting artefact, which is worth understanding before you read anyone's hot take. The company has applied to list Class A shares on the Nasdaq Global Select Market under the ticker "ANT". The price range in the amended filing is still blank.

What happened

Anthropic filed confidentially in June and went through five amended draft registrations between July and October, according to the EDGAR filing history. The public S-1 landed on 26 October. CNBC's live coverage that day worked from first-half numbers; the 2 November amendment replaced them with nine-month figures. Goldman Sachs and Morgan Stanley head a syndicate of more than 20 banks on the cover page.

The last private price was set in May, when Anthropic raised $65 billion at a $965 billion valuation. IPO valuation talk since then has ranged widely. SiliconANGLE, summarising Wall Street Journal and Financial Times reporting, says Anthropic hopes for about $1.5 trillion, one shareholder suggested about $1 trillion, and the FT reported a plan to list before Thanksgiving. None of that is in the filing. Until a further S-1/A sets a range, every per-share number you read is a leak or a guess.

The numbers

All figures below are from the 2 November S-1/A, in US dollars.

Metric Value Period Source
Revenue $33.24B (vs $2.55B) 9 months to 30 Sep 2026 S-1/A
Revenue, Q3 2026 $16.96B (vs $1.38B) Quarter S-1/A quarterly table
Compute and infrastructure cost $18.81B 9 months S-1/A
Operating loss $1.26B (vs $5.22B) 9 months S-1/A
Change in fair value of notes, derivatives, SAFEs $94.32B charge 9 months S-1/A
Net loss $104.04B 9 months S-1/A
Net loss, full year 2025 $41.97B on $4.59B revenue FY2025 S-1/A
Cash plus short-term investments $95.6B + $36.3B 30 Sep 2026 S-1/A
Compute and lease commitments ~$895.4B, of which ~$447.5B before end-2030 As of and after 30 Sep S-1/A
Net revenue retention ~900% 30 Sep 2026 S-1/A
Customers above $1M trailing revenue 1,900+ 30 Sep 2026 S-1/A
Employees ~5,300 30 Sep 2026 S-1/A
Anthropic quarterly revenue vs compute cost ($M) Revenue Compute and infrastructure Q1 2025 376 944 Q2 2025 787 1,390 Q3 2025 1,382 2,047 Q4 2025 2,040 2,945 Q1 2026 4,731 3,387 Q2 2026 11,555 5,558 Q3 2026 16,956 9,863
Quarterly revenue and compute and infrastructure cost from the unaudited quarterly table in Anthropic's 2 November 2026 S-1/A. Revenue first exceeded compute cost in Q1 2026.

Why the loss is $104 billion when the operating loss is $1.3 billion

The filing's footnote explains it: the $94.3 billion charge is the change in fair value of convertible notes, derivative liabilities and SAFEs, "primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation." In plain terms, when your valuation goes from hundreds of billions to nearly a trillion, the notes that will convert into your stock become more expensive liabilities, and that increase runs through the income statement. Another $10.4 billion of other expense, which the filing ties mainly to fair-value accounting on partner financing arrangements, also sits below the operating line. The note charge is non-cash, and the filing says the notes convert into stock in connection with the offering. It is also a real transfer of ownership to the note holders, so calling it "fake" is wrong too.

The operating line is the one engineers should look at. Compute and infrastructure cost was 251% of revenue in Q1 2025 and 58% in Q3 2026, by the filing's percentage table. That is the curve every AI business is trying to bend, and Anthropic's version of it is now public and audited for the annual periods.

The compute bill

The figure that surprised me most is the commitment total: roughly $895.4 billion of purchase commitments and leases "substantially all of which is related to compute and infrastructure", including about $447.5 billion expected before the end of 2030. One line item shows how much of the industry is now selling to everyone: between May and October 2026 Anthropic agreed to spend up to $120.5 billion with SpaceXAI, the former xAI, for Nvidia-powered compute, at about $3.1 billion a month once capacity is fully delivered, which the filing expects in December 2026. SpaceX's own third-quarter release credits new cloud services agreements for most of its AI segment growth.

Against that, the balance sheet held $95.6 billion of cash and $36.3 billion of short-term investments at 30 September, after $141.1 billion of capital raised to date. The filing also warns that take-or-pay terms "obligate us to pay for resources regardless of actual utilization."

What sceptics say

The bear case is not hard to assemble from public material. Independent research firm New Constructs valued Anthropic at $150 billion before the filing went public, a fraction of the private mark. A net revenue retention of about 900% is extraordinary, but it describes a customer base that is still in its land-grab phase, and the filing itself says model versions released in the prior six months produced over 80% of Q3 revenue: the product is only as good as the next model. Concentration is real as well, with the top five customers at about 25% of nine-month revenue. And $895 billion of commitments against $33 billion of nine-month revenue is a ratio that only works if growth continues for years.

The bull case is in the same document: an operating loss that shrank from $5.2 billion to $1.3 billion while revenue grew thirteenfold (though Q3 alone swung back to a $567 million operating loss after a $974 million operating profit in Q2), more than 1,900 customers paying over $1 million a year, and no single customer above 10% of revenue.

What it means for engineering and DevOps teams

My take, and it is opinion: the S-1 is the most detailed public look anyone has had at the unit economics of selling frontier model access, and three things in it matter to people who buy that access.

First, consumption pricing is the business. The filing says that in Q3, Claude Enterprise users who did not use Claude Code spent under $75 a month on average, while the top 5% of users who did spent about 20 times that. If your bill looks like that, you need the same token budgeting and attribution discipline you apply to cloud spend; I covered the tooling in FinOps for AI in 2026.

Second, vendor risk changes shape after an IPO. A public Anthropic will report quarterly, which helps you plan, and will face pressure on gross margin, which may show up in pricing and rate limits. Keep a gateway in front of your model calls so switching or splitting providers is a config change, as in AI Gateways in 2026.

Third, the compute commitments tell you capacity is being bought years ahead across AWS Trainium, Google TPUs, Nvidia GPUs and now SpaceX. Model providers are multi-cloud and multi-chip underneath. Your applications should not be single-model on top.

FAQ

When is the Anthropic IPO date?

No date is in the filing. The 2 November S-1/A leaves the price range, share count and delivery date blank. Press reports point to a listing in late November 2026, but a priced S-1/A on EDGAR is the only reliable signal.

Why did Anthropic lose $104 billion?

About $94.3 billion of it is a non-cash charge from revaluing convertible notes and similar instruments as the company's valuation rose. The operating loss for the nine months was $1.26 billion.

What is Anthropic's ticker symbol?

Anthropic has applied to list Class A common stock on the Nasdaq Global Select Market under the symbol ANT.

Related: GPU Clouds in 2026: H100s at $2.59 an Hour, $104B of CoreWeave Backlog on $35B of Debt, and a Six-Year Bet on Silicon, Coding Agents in 2026: Cursor at $2B Then Sold to SpaceX, Cognition at $1B, 58% on Terminal-Bench, and the Only RCT Still Says Slower and FinOps for AI in 2026: 98% Manage Token Spend, 3 in 4 Can't Prove the Value, GPUs Run at 5%, and FOCUS Gets a Token Column in 1.5.

Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.

I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.

Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.


Related