Anthropic's $42 Billion Loss and Its Compute Commitments

Anthropic recorded a net loss approaching $42 billion in 2025, according to figures from a confidential draft IPO prospectus reviewed by Reuters. About $34 billion of that loss was a non-cash charge tied to financing that could convert into shares. The company also recorded an operating loss of $8.06 billion. Operating loss measures the shortfall after operating revenue is set against operating expenses; it is not the amount spent to run the company.

The reported financing charge grew as the estimated value of the convertible instruments rose alongside Anthropic’s valuation. Recording that increase as a loss did not require a $34 billion payment in 2025. Yet setting the charge aside does not produce a cash-burn figure. Operating loss and operating cash flow answer different questions. The confidential draft has not been publicly released, leaving its detailed accounting notes and cash-flow reconciliation unavailable for that comparison.

A Claude developer cannot pay an API bill with an accounting adjustment. Anthropic generated nearly $4.6 billion in 2025 revenue while its operating loss reached $8.06 billion, up from $2.98 billion in 2024. Revenue was roughly 12 times the previous year’s figure. Compute and infrastructure spending more than tripled to $7.33 billion in 2025 and accounted for over half of the reported $12.65 billion in operating expenses. Those figures, drawn from Reuters’ review of the draft prospectus, describe an expanding business with a growing operating shortfall. Compute spending sits within operating expenses. Adding it to the operating loss would count an expense and the resulting shortfall as separate bills.

Across 2024 and 2025, sales grew fast enough to change the scale of Anthropic’s business, but not fast enough to close its operating shortfall. The reported compute bill grew alongside that expansion. One year of growth cannot tell a maintainer whether the price of Claude requests will stay workable for a project with a small budget. Nor does it establish whether more customers, cheaper model serving or a different product mix will eventually close Anthropic’s gap. The financing charge explains much of the net-loss headline; the operating figures leave the recurring-cost question open.

Capacity Comes with Commitments

Anthropic has at least $518 billion in future computing and infrastructure commitments, extending over approximately a decade, according to figures attributed to Reuters’ examination of the prospectus. Around 80% of the commitments reportedly cannot be cancelled or require payment regardless of actual usage. The reported arrangements include at least $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft, plus approximately $161.2 billion in Broadcom-related equipment lease commitments. The aggregate spans years, but the payment schedule is unavailable. It cannot tell readers how much comes due in any given year.

The prospectus reportedly identifies compute availability as a potential constraint on Anthropic’s growth; the company is also moving beyond cloud services toward leased processors and dedicated data-center infrastructure. Securing capacity ahead of demand could help it serve customers who need dependable access to Claude. If demand falls short while usage-independent payments continue, Anthropic has less room to reduce that portion of its costs. The same arrangement can buy room to grow and create exposure to unused capacity. Its value depends partly on future usage that a commitment total cannot predict.

An open-source maintainer deciding whether to build a Claude integration has a more immediate concentration problem. Code that assumes one provider’s request format, model behavior and pricing takes work to move. A second provider might offer an alternative, but someone has to maintain the adapter and test whether the application still behaves as expected. Anthropic’s supplier agreements make none of those choices for a small project. Nor does the $518 billion total establish what API access or pricing that project will receive.

Nearly a quarter of Anthropic’s 2025 revenue came from two clients, according to people familiar with the filing cited by Ars Technica. Several major clients lack long-term contracts, while some reported supplier obligations require payment regardless of use. A client’s departure is not inevitable, but revenue from customers free to leave offers less contractual assurance against those payments. Anthropic held $20.28 billion in cash, cash equivalents and short-term investments at the end of 2025. Dividing that snapshot by the multi-year commitment total would produce a bogus runway: future sales, financing, contract conditions and payment dates all belong in that assessment.

For an API-dependent community, customer concentration adds a question to the usual concern about being locked into one model. A provider with a few large buyers may have an incentive to prioritize their requirements. That is a possibility, not a claim about Anthropic’s product decisions. A maintainer can ask who will keep an alternative integration working if the project’s requirements and its provider’s priorities diverge.

A Buyer’s Promise and a Builder’s Budget

Large technology companies face infrastructure bets of their own, on different terms. In February, Futurum estimated 2026 capital expenditure of $660 billion to $690 billion across Microsoft, Alphabet, Amazon, Meta and Oracle. That annual estimate includes spending beyond AI: Amazon’s plan, for example, also covers logistics. Anthropic’s reported $518 billion concerns obligations extending over approximately a decade, some of them payable regardless of use. Placing the figures side by side illustrates the scale of the buildout, not equivalent annual bills or equivalent contractual liabilities.

Alphabet offers a closer look at the other side of a cloud deal. By September, its expected 2026 capital expenditure stood at $195 billion to $205 billion, Barchart reported, above the $175 billion to $185 billion plan Futurum cited in February. The figures are dated views of an annual spending plan, rather than a measure of which contracts Alphabet could cancel. Alphabet builds infrastructure for its own services and cloud customers. Google Cloud chief Thomas Kurian said the aggregate payback period for Google’s AI servers is less than two years, as Barchart relayed. That is Google’s claim about its servers, not an established return for Anthropic’s supplier deals.

Google can seek customers for infrastructure it operates; Anthropic buys or leases resources to deliver models to its customers. Where Anthropic owes payments regardless of usage, slower model demand could leave it paying for resources that bring in less revenue than expected. Stronger demand could instead make advance access valuable. Neither outcome follows from the commitment total alone.

The $42 billion net loss cannot supply a cash runway, and the $8.06 billion operating loss cannot be substituted for cash burn. A public prospectus with cash flows and contract payment dates would make Anthropic’s financial flexibility easier to assess. Until then, a small team considering Claude still has a concrete choice: decide how much of its software to tie to one provider before it needs an exit path.