Anthropic, the AI company behind the Claude models, has committed about $518B to computing infrastructure over the next decade, and roughly 80% of that total is non-cancellable or must be paid regardless of whether the capacity is used. The figures come from the company's confidential IPO prospectus, filed with the US Securities and Exchange Commission in June and reported on 29 September.
The commitments are spread across most of the industry. Broadcom equipment leases account for $161.2B and are largely non-cancellable. Google is committed at $111.1B from April 2026 to July 2033, Amazon at $110B to April 2036, and Microsoft at $31.4B to May 2033, cancellable only for material breach. A deal with Elon Musk's xAI worth up to $84.5B through 2029 is largely cancellable at 90 days' notice, and AMD has up to $20B of capacity plus a $5B stock purchase commitment.
"If our actual spend falls short, we must pay Google the difference," the prospectus says of one arrangement.
Why Lock In Capacity This Way
Non-cancellable commitments look alarming until you consider what they buy. Advanced AI chips, the power to run them and the data centre space to house them are all scarce, and suppliers allocate them to customers willing to guarantee payment.
A company planning to train successively larger models cannot risk discovering in 2029 that its capacity went to someone who signed a firmer contract. The trade is certainty of supply for loss of flexibility, and every large AI developer is making some version of it.
The structure also reveals the state of the market. When buyers accept take-or-pay terms across a decade, suppliers hold the power, which matches what Oracle, CoreWeave and the chipmakers have been reporting about demand.
The Comparison Across The Industry
Anthropic is not an outlier. OpenAI's Stargate programme carries a headline figure around $500B, with costs shared among several parties, and Oracle's backlog of $664B is roughly half attributable to OpenAI alone.
What differs is where the obligation sits. Oracle borrows to build and carries the construction risk; Anthropic commits to pay and carries the demand risk. If AI revenue growth slows, those contracts become expensive in a way that cancelling a cloud subscription would not be.
The Revenue Behind The Spending
The spending only makes sense against the growth Anthropic reports. The prospectus puts 2025 revenue at $4.6B, a run rate of $65B based on July sales, and a projection of $200B by 2028, according to analysis in Forbes of the filing.
One disclosed risk deserves attention: two customers accounted for 25% of 2025 revenue, without long-term contracts. A business with fixed obligations of $518B and concentrated, uncommitted revenue is carrying an asymmetry that investors will examine closely.
The Companies On Both Sides Of The Deal
The prospectus flags another structural issue. Amazon, Google and Microsoft are simultaneously investors in Anthropic, customers for its models, suppliers of its computing capacity, distributors of its products and competitors building rival models.
Money therefore flows in several directions between the same parties, and a cloud provider that invests in a customer which then spends the investment on that provider's capacity is reporting revenue that originated with itself.
This pattern now runs through the sector, from Nvidia's investments in companies that buy its chips to the arrangements between cloud providers and model developers. It is not improper, but it makes the industry's reported growth harder to interpret, since the same dollar can appear in several companies' results.
What Business Buyers Should Take From It
For companies buying AI services, the filing is a useful window into supplier economics. A provider with large fixed obligations has strong incentives to fill that capacity, which argues for competitive pricing in the near term and for pressure on margins later if growth disappoints.
It also indicates that supply constraints are expected to persist for years, since nobody signs decade-long take-or-pay contracts for something they expect to become cheap and plentiful. Buyers negotiating multi-year AI commitments should weigh that against the risk of locking themselves into terms as prices and models change.
The Whole Sector Is Betting The Same Way
Anthropic's $518B is one line in a much larger pattern. Oracle is borrowing heavily to build capacity it has pre-sold, smaller providers such as Australia's Sharon AI are mortgaging GPUs to fund deployments, and Toshiba is doubling hard drive production on the expectation that AI storage demand holds.
Each of those decisions rests on the same assumption: that demand for AI compute keeps rising through the decade. The commitments are now large enough that if the assumption fails, the losses land across chipmakers, cloud providers, storage suppliers and the model developers at once.
Certainty Of Supply, At The Price Of Flexibility
What Anthropic has bought with these contracts is the ability to plan. It knows what compute it will have through 2036, which is a genuine advantage in a market where capacity is the binding constraint on building better models.
What it has given up is the option to change its mind. An 80% non-cancellable share means the company must grow into those obligations rather than adjust to conditions, and that is the bet investors will be asked to take a view on when the listing comes. The prospectus is unusually clear about the risk, which is more than most of this cycle's infrastructure commitments have offered.
