Anthropic’s $125B TPU Lease Meets Falling Token Prices

In September, the average cost of a million AI tokens had fallen to 97 cents, from $2.07 on May 28. Yet Anthropic had committed $125.2 billion to five years of TPU leases, while OpenAI was offering discounts for one- to three-year guarantees of compute access.

Frontier labs are adding capacity contracts to usage-priced AI inference. Fixed, sometimes non-cancelable infrastructure bills give them reason to exchange discounted rates and guaranteed throughput for committed spend, shifting some utilization risk to enterprise customers.

An hourly rental became a multiyear obligation

In the 2018 coverage record, Google Cloud offered beta Cloud TPUs starting at $6.50 an hour. That price described access to a machine for an hour, not a customer’s obligation to keep buying its output. Eight years later, Anthropic has committed $125.2 billion over five years for TPU leases, while Broadcom has agreed to lend it up to $42 billion through a convertible note that could help finance them. The hourly rate and the lease total are not comparable prices: one quotes access by the hour, the other commits a model provider for years.

Anthropic’s prospectus describes more than $518 billion in expected spending over ten years with six AI-infrastructure partners, roughly 80% of it non-cancelable or payable regardless of usage. A model provider can reduce the price of its next token; it cannot reduce a fixed lease payment by leaving the chip idle. Filled capacity matters to its commercial decisions, not just the advertised token rate.

Those obligations do not all work alike. Anthropic’s agreement to pay SpaceX up to $84.5 billion for Nvidia-based compute through 2029 was reported as mostly cancellable on 90 days’ notice. The cancellation clause, not the headline value, determines how much demand risk Anthropic must carry.

A discount can buy a demand forecast

OpenAI offers Guaranteed Capacity at a discount through one- to three-year commitments based on spend levels. Earlier competition for startup customers featured token credits, promotions and one-time bonuses. A credit encourages a customer to try an API; a spend commitment gives the provider a longer view of the demand it must serve. OpenAI’s offer establishes that bargain, though it does not establish that its customer commitments have the same cancellation terms as Anthropic’s infrastructure leases.

Sources also say Anthropic has ended discounts, typically around 15%, after customers exhausted purchased-token allowances and required renegotiations. Anthropic has not confirmed that account as company policy, and it should not be treated as an industry rule. It does identify the boundary a buyer needs to inspect: the price after the committed volume runs out.

The buyer has a reason to resist signing too early. The LLM Token Expenditure Index put the average cost per million tokens at 97 cents in September, down from $2.07 on May 28. If spot prices keep falling, a customer that reserves capacity may pay for certainty it did not need. A customer whose workflow stalls without throughput may value the same contract differently. Neither buyer can settle that choice from a discount percentage alone.

Agents make a seat a poor unit of demand

SemiAnalysis estimates that agentic AI accounts for about 90% of Anthropic’s business; sources cited in the same report say two clients generated nearly a quarter of its 2025 revenue. Those are estimates and sourced claims, not a count of the tokens each customer consumed. They nevertheless point to the operational problem: a coding or autonomous workflow can call a model repeatedly while one employee oversees it, so a software seat does not reliably describe its compute demand.

Anthropic has placed Claude Code CLI and Claude for Microsoft 365 in early access. Buyers testing either need to measure model calls per workflow before committing to capacity: a seat count cannot tell them how much compute an agent will consume.

The chip supplier now helps finance the customer

Broadcom’s convertible note ties a chip supplier’s financing decision to Anthropic’s ability to fund leased compute. Elsewhere in the stack, General Compute obtained a $400 million loan described as the first to use inference-specific chips as collateral. Lenders are no longer financing only buildings and general-purpose equipment. They are assigning value to hardware whose use depends on sustained inference demand.

CME Group and Silicon Data have also announced futures contracts based on daily GPU benchmarks for on-demand rental rates. Such a contract could address exposure to a compute price; it would not guarantee that a particular model remains available on the same API terms. A buyer needs to separate those risks before treating “compute” as one purchasable input.

The buyer can replace a model more easily than a signature

An enterprise can put model access behind a backend interface and preserve the option to change providers when prices, policies or available models change. It can also measure a workflow by completed work rather than by cheap tokens. Neither design choice cancels a capacity reservation already signed. Procurement must therefore specify what happens when consumption exceeds an allowance, falls short of a minimum, or moves to another model.

September’s 97-cent average measures the price of tokens used. A reserved slot puts a price on the risk of leaving capacity unused.