According to Reuters, which reviewed confidential documents, Anthropic is preparing for a U.S. stock market listing and targeting a valuation of about $2 trillion. If it happens, it would be the largest IPO in history.
The company confidentially filed with the U.S. Securities and Exchange Commission (SEC) in June, but the documents have not yet been made public. The prospectus seen by Reuters revealed a previously hidden picture of the lab's economics: enormous revenue growth, enormous losses, and — most importantly — compute commitments of more than half a trillion dollars over a decade.
Key numbers: $4.6B revenue, $42B loss
Per the prospectus, Anthropic earned about $4.6B in revenue in 2025 — roughly 12 times more than the year before. At the same time the company posted a $42B net loss. While the figure looks frightening, nearly $34B of it is a non-cash accounting entry reflecting the increased value of financing instruments that may convert into shares in the future — i.e., not real money spent running the business.
Excluding these entries, the operating loss was $8.06B. Compute and infrastructure expenses were $7.33B — three times more than in 2024 and nearly 58% of total operating expenses of $12.65B. At year-end, the company's cash and short-term investments stood at $20.28B.
$518 billion: non-cancellable agreements
The prospectus's most sensational part: Anthropic plans to spend at least $518B on AI infrastructure with six partners over ten years. This is one of the largest AI build-out commitments in history. Nearly 80% of it is non-cancellable or payable regardless of service usage.
Under long-term infrastructure commitments, the company agreed to pay Google at least $111.1B over 7–10 years, Amazon $110B, and Microsoft $31.4B — "regardless of usage." The Google agreement runs from April 2026 to July 2033, Amazon's from May 2026 to April 2036. The document states directly: "If our actual spending is less, we must pay the difference to Google" — the same condition applies with Amazon.
A separate line covers equipment-leasing commitments with Broadcom of about $161.2B; these too are nearly non-cancellable (only in case of default). The $31.4B agreement with Microsoft (November 2026 to May 2033) can be terminated only upon a serious unremedied breach by Microsoft.
Additional partners: xAI and AMD
The prospectus also disclosed agreements with Elon Musk's xAI: up to $84.5B in spending on Nvidia-chip-based compute through 2029 — though these commitments are mostly cancellable with 90 days' notice. AMD also committed to buy up to $5B of Anthropic shares and deliver more than $20B in compute capacity.
Why agree to such terms?
Anthropic explains the necessity of these commitments to investors thus: access to compute is becoming the main constraint on AI development — future demand for advanced AI systems is expected to exceed supply and be "largely constrained by the availability of compute." In other words, the lab is now insuring itself against a future GPU shortage — whatever the price.
The company also said it is gradually moving away from a "cloud-only" model toward its own data centers and directly leased chips. This is a move to strengthen infrastructure control and optimize costs long-term.
Two-sided dependence: partners are also rivals
One of the document's most candid passages is the risk of dependence on Big Tech companies. Amazon, Google, and Microsoft simultaneously act as investor, customer, cloud provider, distributor, AND competitor — they're also developing their own AI models. The prospectus acknowledges these interests "may not be fully aligned with Anthropic's interests."
The numbers show the degree of dependence: in 2025, 47% of revenue came from sales through Amazon and Google (32% in 2024). Nearly a quarter of revenue comes from just two customers, and many large customers aren't bound by long-term contracts — they can cut spending at any time. "If third-party compute is limited, becomes more expensive, or is discontinued... our business and financial condition could be seriously harmed," the document says.
Warning in the IPO filing: AI's dark scenario
Anthropic wrote its "risk factors" section far more candidly than usual. The company acknowledged that broadly deployed AI could concentrate power and wealth to a degree harmful to society and destabilizing to the geopolitical order. It also noted that agentic AI technology creates "serious and unpredictable legal risks."
"Broadly deployed artificial intelligence could concentrate power and wealth to a degree harmful to society and destabilizing to the geopolitical order." — Anthropic IPO prospectus (Reuters)
These warnings sounded against the backdrop of CEO Dario Amodei's call this month for industry players to slow the pace of AI capability releases — the tension between the safety call and aggressive expansion will be the IPO's main drama.
Race with Stargate: the half-trillion club
The $518B plan is being compared in scale to OpenAI's Stargate project — a $500B AI infrastructure plan whose costs are expected to be shared by OpenAI, SoftBank, Oracle, and MGX. Both projects confirm one truth: the frontier AI race is no longer about model architecture but about gigawatts of electricity and millions of chips.
Note: the Broadcom deal we covered earlier came from exactly this prospectus — we have a separate article on Broadcom's agreement to lend Anthropic up to $42B.
