Anthropic's pre-IPO math: $518 billion in commitments against $4.6 billion in revenue
In a confidential prospectus for its planned stock market listing, Anthropic discloses decade-long infrastructure agreements with six partners, most of which, according to the company, require payment regardless of how much of the capacity is used. The figures — which cannot be independently verified — come weeks before a listing that prospective investors believe could value the company at $1.8–2 trillion.
What has been disclosed
Anthropic expects to spend at least $518 billion over ten years building AI infrastructure together with six partners. That is according to a confidential IPO prospectus seen by Reuters, reported on 29 September 2026. The company itself describes the plan as one of the largest AI buildout commitments ever recorded.
What is striking is not just the size, but the rigidity: According to the prospectus, around 80 percent of the total is either non-cancelable or requires payment "regardless of usage," as the wording is reported by Reuters.
Partner by partner: Where the money goes
The prospectus breaks the commitments down as follows, according to Reuters' review of the document:
- Google (Alphabet): at least $111.1 billion
- Amazon: at least $110 billion
- Microsoft: at least $31.4 billion
The three cloud agreements are long-term contracts running seven to ten years, payable regardless of usage. On top of that come roughly $161.2 billion in equipment lease obligations tied to the chipmaker Broadcom, which are also largely non-cancelable.
Adding up the disclosed items, these commitments total over $410 billion — a sum AIMag has derived from the partner figures in the prospectus, not a number the prospectus itself states. What flexibility remains sits at the edge of the portfolio:
- xAI (Elon Musk): agreements worth up to $84.5 billion in Nvidia-based capacity through 2029, but these are largely cancelable on 90 days' notice — by far the most flexible item.
- AMD: an arrangement in which AMD can buy up to $5 billion of Anthropic stock, against over $20 billion in capacity.
Why rigid commitments matter
The distinction between fixed and cancelable commitments is the core of the company's risk profile. Fixed contracts function somewhat like fixed-rate loans on computing power: They provide guaranteed access and price predictability in a market where capacity is scarce, but they bind the company to payment streams for the next ten years regardless of how revenue develops. Cancelable agreements like the xAI deal, by contrast, provide a safety valve — Anthropic can reduce costs quickly if demand falls.
According to Reuters, the prospectus itself offers a rationale for the strategy: Anthropic says access to computing power will become the key bottleneck for AI development, and that the company is therefore moving from a pure cloud model toward its own dedicated data centers and directly leased chips. The capacity agreements the company has already announced show the scale: reservations of Google/Broadcom TPUs of around five gigawatts from 2027, Amazon's Project Rainier cluster with over 500,000 Trainium2 chips, and an agreement with SpaceX from May 2026 for more than 300 megawatts and over 220,000 Nvidia GPUs in the Colossus 1 data center.
The numbers behind it: Growth against losses
For 2025, Anthropic reported, according to figures cited from the prospectus, revenue of around $4.6 billion against an operating loss of $8.06 billion, up from $2.98 billion in 2024.
One important nuance: the company's net loss in 2025 was nearly $42 billion, but almost $34 billion of that was a non-cash accounting charge related to financing instruments. The operating loss of $8.06 billion is thus the most comparable measure of the real operating burn, while the $42 billion figure must be read with that accounting deduction in mind.
At the same time, the company points to steep revenue growth: In April 2026, Anthropic said its annualized revenue run-rate had reached $30 billion. A separate, secondary account claims the run-rate had passed $65 billion as recently as July 2026. The two figures come from different sources and cannot be reconciled based on what is available — the discrepancy is unexplained, and both numbers should be verified before conclusions are drawn from them. Whichever is correct, revenue is growing far faster than the $4.6 billion for all of 2025 — but from a base where the operating loss still exceeds revenue.
The listing in sight
The prospectus has not been made public, but it is part of the preparations for a listing that prospective investors believe could value Anthropic at between $1.8 and $2 trillion, according to Bloomberg as reported by Quartz. A roadshow could, according to the same sources, begin the week of 9 November, with the aim of listing before Thanksgiving. Both the valuation and the timeline are thus expectations among investors, not confirmed facts.
The spreadsheet behind the valuation is therefore uniquely visible to the market before the listing: a fast-growing revenue base, a substantial operating loss, and decade-long commitments that exceed annual revenue by a factor of more than one hundred.
Open questions
Several issues remain unresolved. The prospectus is confidential, so none of the figures — neither the $518 billion total, the 80 percent, the partner amounts, nor the financials — can be independently verified from the document itself. Anthropic had not responded to Reuters' inquiries when the report was published, so the company has not itself confirmed the cited terms.
Beyond that, it remains to be seen how the balance between fixed commitments and revenue growth develops. The cost side is largely locked in for seven to ten years; the revenue side is far more uncertain, as the gap between the two run-rate figures illustrates. The analysis firm TrendForce estimates, according to a secondary account, that high-bandwidth memory prices will rise 121 percent year over year in 2027 — a signal that the inputs to the buildout itself could become more expensive even as the commitments stand firm.
Anything in this could change the math. The only certainty is that future shareholders now get to see the terms before they subscribe.