The developer of the Claude AI model has filed for an IPO in the US, and the disclosed documents reveal an alarming signal: Anthropic's total obligations to its partners are approaching the $1 trillion mark. This is the kind of debt load that can either sink a company or make it invulnerable — depending on whether it can keep revenue growth in line with its payment schedule.

Who the company owes and how much

Anthropic has little in the way of classic debt. As of the end of June 2026, it amounts to $6.2 billion in convertible notes held by Amazon, which will convert into shares at IPO. A credit line from a banking syndicate led by Morgan Stanley grew to $15 billion in September, but by mid-year the company had not drawn on it. It can borrow another $77 billion under agreements with Google, Amazon, and Broadcom.

The main problem lies elsewhere. Anthropic has signed multi-year contracts for computing capacity and is obligated to pay for them regardless of whether it uses that capacity or not. The largest of these are:

  • Broadcom — about $161.2 billion for a five-year lease of AI accelerators;
  • SpaceXAI — up to $120.5 billion for access to Nvidia GPUs through 2029, about $3.1 billion per month, although from 2027 the contract can be terminated with 90 days' notice;
  • Google — a minimum of $114.5 billion through 2033 and another $125–150 billion for additional TPU capacity in 2028–2031;
  • Amazon — a minimum of $110 billion through 2036;
  • Microsoft — a minimum of $44.9 billion through 2033;
  • Nscale — up to $44.6 billion, Lambda — about $41.3 billion, AMD — more than $20 billion.

The company did not disclose the total amount in its preliminary prospectus, but it can be pieced together from the notes to the financial statements. As of mid-2026, contractual obligations stood at $265.6 billion, with another $321.4 billion in lease agreements that had not yet taken effect. After June, the company added contracts worth roughly $245 billion and another $21 billion in short-term ones. In total, that is more than $850 billion, and including the additional TPU capacity — about $1 trillion.

To cover these obligations, Anthropic has $131.7 billion in cash and short-term investments — less than one-sixth of the contract total. Revenue for the first half of 2026 was $16.3 billion, and in the second quarter the company reached operating profit of $974 million.

The main risk. The company's revenue depends almost entirely on consumption. Clients can cut spending at any moment, while capacity payments are scheduled years in advance.

If revenue falls behind schedule, cross-defaults on the credit line and notes will be triggered, and creditors will be able to demand early repayment of all debt. Some of the data center and electricity contracts are for 20 years with annual indexation and no cap on fuel costs. The Broadcom notes are structured so that if revenue targets are missed, the rate rises from 9% to 13%, and conversion occurs at a larger discount. Debt becomes more expensive precisely when the business gets worse.

The structure also has a flip side. Amazon, Google, and Microsoft are simultaneously investors, suppliers, and clients of Anthropic, while Amazon and Broadcom are also creditors. A crash of the company would hit their balance sheets and deprive them of a major buyer of capacity. Too many parties have an interest in not letting Anthropic fall.

Historical examples

The closest example is WeWork. In its 2019 IPO filing, the company showed about $47 billion in office lease obligations for 15 or more years, while income came from tenants on short, often month-to-month contracts. The IPO was canceled, SoftBank poured billions into the company, but in November 2023 WeWork still filed for bankruptcy. The pattern of "long fixed costs against short flexible revenues" is the same as at Anthropic.

Global Crossing in the late 1990s laid undersea fiber-optic cables in anticipation of explosive growth in internet traffic. Traffic grew more slowly than expected, capacity prices collapsed, and in January 2002 the company went bankrupt with about $12 billion in debt. Capacity was built before demand appeared. Anthropic listed this scenario among its risks.

AIG in 2008 sold default insurance to banks around the world and ended up owing everyone at once. It was not allowed to fail: the US government allocated about $182 billion for the rescue, because the insurer's collapse would have dragged down all its counterparties. Lehman Brothers, with $613 billion in debt, received no such support.

Evergrande owed banks, bondholders, contractors, and homebuyers more than $300 billion. Size did not save it: in 2021 the developer defaulted, and in January 2024 a Hong Kong court ordered the company to be liquidated.

For investors, the key indicator is the pace of revenue relative to the payment schedule. As long as the former outpaces the latter, the structure works in the company's favor. But Anthropic warned in its filing that growth will slow, and that 5–7% of revenue comes from clients the company disconnects for rule violations. A slowdown in such a model would hit not one creditor but the entire chain at once.

Cryptalist analysis: Anthropic's liability structure reflects a new norm for flagships of the AI sector: the capital intensity of computing infrastructure is growing faster than the monetization of models, and the gap is being closed with long-term contracts with fixed payments. What becomes key for the market is not the absolute size of the debt but the dynamics of revenue relative to the payment schedule — it is precisely this spread that will determine the company's valuation at IPO and the resilience of the entire chain of chip and data center suppliers. The intertwining of interests of Amazon, Google, Microsoft, and Broadcom as simultaneously investors, creditors, and clients reduces the likelihood of a hard default but does not eliminate the risk of a sector repricing at the first slowdown in growth rates.