Anthropic Files for IPO: Its Must-Pay Computing Deals Need About Twice Today's Revenue, and Open-Weight Rivals Are Catching Up
Published: October 27, 2026
Claude Code is my bread and butter today, and I love it. But I see a very hard climb ahead for Anthropic, and rivals are catching up fast. Covering the deals it must pay for looks within reach. Covering all of them, and holding the lead for years, is the hard part.
By my rough estimate, the deals it must pay for need about twice today's revenue. Revenue grew 45% in July-September on the quarter before, after 144% the quarter before that.
If growth slows to, say, 30%, 25% and 20% a quarter, it gets there by April-June 2027, while the bills rise as new capacity arrives.
At 30 June it held $132 billion of the $141 billion it has raised, which gives it time.
The other risk is how close open-weight models are getting. NIST puts GLM-5.3, from China's Z.ai, about four months behind the best US models on its cyber tests. Separately, Anthropic's own filing says others have tried to copy and distill its models, and that open-weight models could take parts of the market and push prices down.
The filing also says that if strong models become easy to build, competition could depend more on access to computing on good terms. Most of the computing deals run to 2033 or later, so the lead has to hold for many years. If it slips, prices fall while the bills stay fixed.
The deck has the rough working, its caveats, and my read.
A few nuances the deck could not fit
The filing gives no yearly split, so I spread each deal evenly over its term, and took SpaceXAI at its full $3.1 billion a month. The $115 billion a year is roughly the picture while all seven overlap, around 2027-29. SpaceXAI can be cancelled on 90 days' notice from early 2027, others end after 2029, and more TPU capacity is already committed for 2028-31.
April-June had an operating profit of about $1 billion. That is before $26.6 billion of staff share-award costs, which start once it lists.
Anthropic had about 5,300 staff on 30 Sep, up from about 940 at the end of 2024. It puts revenue per employee over the last 12 months at more than $6.6 million, and says over 90% of the code added in September was written by Claude.
Staff costs were $3.2 billion in January-June, 19% of revenue against 46% a year earlier. About 70% of it was stock-based pay. Even on the September headcount, that is about $1.2 million a head a year.
The filing's results stop at 30 June. July-September results are expected in updated filing before the shares are priced. Investors expect annualized revenue of about $110 billion by year-end, per The Wall Street Journal as reported by CNBC.
Related analysis
A correction in AI valuations is looking more likely, 31 Aug. The view now comes from a central bank, an asset manager and a bond investor: tigzig.com/post/ai-valuation-correction-aug2026
Could this time be different? 24 Oct. Past drawdowns, AI valuations and cracks across markets and the economy: tigzig.com/post/drawdowns-this-time-different-oct2026
Anthropic IPO could reach $2T. The headwinds are already here, 30 Sep: tigzig.com/post/anthropic-ipo-2-trillion-headwinds-sep2026
GLM-5.3 from Z.ai can build cyber exploits on its own, 30 Sep. Anthropic's findings and NIST's assessment: tigzig.com/post/anthropic-glm53-open-weight-cyber-sep2026
A crescendo of warnings on AI debt and returns, 23 Oct. S&P Global, Gartner, Morgan Stanley, the IMF and the World Bank: tigzig.com/post/ai-warnings-crescendo-oct2026
The biggest tech build meets the bond rout of 2026, 28 Sep. Hyperscaler spending, free cash flow and debt: tigzig.com/post/ai-hyperscalers-debt-bond-rout-sep2026
Analysis of public sources, for information only. Not investment advice.
Full Deck Content (Text Format)
Text below was extracted from the source deck. Chart visuals stay in the PDF and as slide images above the post.
Slide 1
Anthropic files for IPO. By my estimate, its must-pay deals need about $125 billion of revenue a year
That is about twice its current pace. Revenue grew 45% in July-September, on April-June. Of the $141 billion raised, it held $132 billion at 30 June.
Revenue vs break-even, $bn a year 46 Apr-Jun filing 67 +45% Jul-Sep press ~125 Must pay deals 200+ All 7 deals Apr-Jun and Jul-Sep are quarterly revenue x 4. Jul-Sep was 45% above Apr-Jun.
| Revenue a year, latest $16.8bn in July to Sept, x 4 | $67bn |
|---|---|
| Break-even, must-pay deals, a year, page 4 | ~$125bn |
| Break-even, all 7 deals, a year, page 4 | $200bn+ |
| Computing deals | $800bn+ |
| Raised so far | $141bn |
| Still in hand cash and investments | $132bn |
Break-even means zero operating profit. Rough working and caveats: page 4.
Where the revenue comes from, January to June 2026
How customers pay Per use 80% Subs 20% Where customers are US 72% Rest 28% Sold via Amazon, Google, Microsoft 48% Rest 52% Government customers Agencies, about 1% Rest 99%
Subscriptions include the Claude Pro and Max plans.
SourcesAnthropic S-1, 26 Oct. Jul-Sep revenue: CNBC, 15 Oct. Break-even: my arithmetic.
Amar Harolikar, ACA · Decision Sciences & Applied AI1 of 6tigzig.com
Slide 2
The latest quarter in the filing
| April to June 2026 | $ billion |
|---|---|
| Revenue | 11.6 |
| Compute cost | 5.6 |
| Operating profit, 8% margin | 1.0 |
| Net loss about $68bn of it is non-cash revaluation of notes | -66.3 |
| Cash from operations January to June | 3.9 |
Who put in the $141 billion
It still holds $132 billion of it in cash and investments, at 30 June.
$ billion, shares and notes, since founding
- $23bn Google $18bn Amazon $100bn Everyone else
Google: $20bn of shares in 2026, $3.3bn of notes before that
Amazon: $10bn of shares in 2026, $8bn of notes. Both will hold over 5% going into the IPO
Everyone else includes: Microsoft (announced $5bn in November 2025) Spark Capital
Bessemer
Salesforce Ventures
Menlo
Lightspeed
ICONIQ
GIC
The May round was led by Altimeter, Dragoneer, Greenoaks and Sequoia. Several hundred other investors are not named.
SourcesAnthropic S-1, SEC, 26 Oct 2026: quarterly results, notes 9, 10 and 16, Exhibit 4.2. Google and Amazon amounts are my sums of the pieces the filing names, so they are floors. Microsoft, Bessemer and the May round leaders: CNBC, 26 Oct.
Amar Harolikar · Decision Sciences & Applied AI2 of 6 · 26 October 2026tigzig.com
Slide 3
$800bn+ of computing to pay for
The biggest deals $ billion, total over each contract
Must pay whether used or not Can cancel, or pay as delivered Estimate
Broadcom, Google TPU leases $161bn SpaceXAI, Nvidia GPUs up to $121bn Google $115bn More TPU capacity, 2028-31 $125-150bn est. Amazon $110bn Microsoft $45bn Nscale, Nvidia GPUs up to $45bn Lambda, Nvidia GPUs $41bn AMD, from 2027 over $20bn est.
At least $431 billion must be paid whether or not the computing is used. That is the Google, Amazon and Microsoft deals and the Broadcom leases, my sum.
Some power and data centre deals run 20 years. They are take-or-pay, and fuel costs are passed on without a cap.
Data centres face pushback. New York has paused permits for big new sites, and Texas can cut power to large users in a grid emergency. Anthropic has promised to cover any rise in consumer electricity prices its data centres cause.
SourcesAnthropic S-1: MD&A commitments, risk factors. My total of the filing's minimums: about $853bn.
Amar Harolikar · Decision Sciences & Applied AI3 of 6 · 26 October 2026tigzig.com
Slide 4
Back of the envelope: what the bill means for revenue
Revenue needed to break even
| $ billion a year | Apr-Jun x4 | Must-pay deals | All seven deals |
|---|---|---|---|
| Revenue | 46 | 126 | 219 |
| Computing | 22 | 66 | 115 |
| Other operating costs | 20 | 60 | 104 |
| Operating profit | 4 | 0 | 0 |
Revenue ran at about $67 billion a year in July to September. Break-even is about twice that for the must-pay deals and about three times for all seven.
Working: what Anthropic would pay a year for computing
No yearly split in the filing: each deal spread over its term, SpaceXAI at its monthly rate.
| Deal | Total | Term | A year |
|---|---|---|---|
| SpaceXAI | up to $121bn | $3.1bn a month | $37bn |
| Broadcom TPU leases | $161bn | 5 years | $32bn |
| $115bn | 2026 to 2033 | $16bn | |
| Amazon | $110bn | 2026 to 2036 | $11bn |
| Microsoft | $45bn | 2026 to 2033 | $7bn |
| Nscale | up to $45bn | 2026 to 2033 | $7bn |
| Lambda | $41bn | 2026 to 2034 | $5bn |
| Seven biggest deals | $115bn |
SpaceX's filing shows a customer, which CNBC says is Anthropic, paid it just over $5bn in July to September.
Left out, and all would raise break-even: about $215bn of other commitments, $125-150bn more TPU capacity in 2028-31, over $20bn with AMD, and $26.6bn of share awards.
A rough calculation: how other costs will move with computing is not known. At the April-June ratio, must-pay needs $126bn and all seven $219bn. If other costs grow with revenue instead, about $117bn and $203bn. The 85% gross margin in press reports counts only the computing that serves customers.
SourcesAnthropic S-1: MD&A commitments, quarterly results. CNBC 15 Oct, CNBC 26 Oct. The filing's 86x figure puts Jul-Sep at about $17bn. Yearly figures are my arithmetic.
Amar Harolikar · Decision Sciences & Applied AI4 of 6 · 26 October 2026tigzig.com
Slide 5
Other things in the filing
5% to 7% of revenue in January to September came from customers later found breaking its usage rules. Anthropic expects the share to fall, but some to continue.
Open-weight rivals may take part of the market and push prices down. If strong models become easy to build, competition could depend more on access to computing.
Staff share awards carry $26.6 billion of unbooked cost, plus $5.8 billion granted after June. They vest when Anthropic lists: $2.9 billion at once had it listed on 30 June, the rest over about two years.
Broadcom has a key-person clause. If certain named people leave, Anthropic must buy all the leased TPU equipment or be in default. The names are redacted.
Claude wrote over 90% of the code added to Anthropic's codebase in September.
New models carry the revenue. Models released in the six months to September made up over 80% of July to September revenue.
Google can take up to $10 billion of shares at the IPO price, in place of the same amount of notes.
A September class action claims AI labs including Anthropic agreed to slow model development. Anthropic disputes it.
Revenue per employee was over $6.6 million in the past year, on about 5,300 staff at 30 September.
The founders control the vote. A single share held by the seven co-founders carries 50.1% of votes and no share of profits. A trust elects most of the board.
SourcesAnthropic S-1: risk factors, MD&A, stock compensation, Exhibits 4.4 and 10.11.
Amar Harolikar · Decision Sciences & Applied AI5 of 6 · 26 October 2026tigzig.com
Slide 6
My read
A steep climb ahead, and rivals are catching up fast
On the technology Claude Code is my bread and butter. All my work runs on it, tigzig.com is built with it, and it is the tool I recommend to my clients.
On the numbers On my rough sums (page 4), revenue has to about double to cover the deals it must pay for, and roughly triple to cover all seven. Revenue grew 45% in July-September, after 144% the quarter before. If growth keeps slowing, to 30%, 25% and 20% a quarter, it reaches the must-pay level in about three quarters, while the bills rise as new capacity arrives. At 30 June it held $132 billion of the $141 billion raised, which gives it time.
Rivals Open-weight models are close behind, and the filing says copying and distillation could wear down its edge. Most deals run to 2033 or later, so the lead has to hold for years. If it slips, prices fall while the bills stay fixed.
I believe AI will ultimately pay off. Who is still standing five or ten years from now, and who benefits, is hard to say.
Earlier notes (tigzig.com ➜ Analysis)
30 Sep Anthropic IPO could reach $2T. The headwinds are already here
30 Sep GLM-5.3 can build cyber exploits on its own, Anthropic finds
31 Aug A correction in AI valuations is looking more likely
Not investment advice. This is my analysis of public sources, for information only. Care has been taken, but errors are possible. Use of AI. Parts of this work were prepared with AI under my direction. I reviewed, edited and approved every page. The analysis and conclusions are mine.
Amar Harolikar · Decision Sciences & Applied AI6 of 6 · 26 October 2026tigzig.com
Working on something similar? How I work covers the rates, the availability and what I take on.