Compute is the binding constraint on AI, and the price is fixed

Frontier AI is no longer a software business. It is a utility business, and the binding input is compute.

Anthropic's confidential S-1, reported by Reuters, commits at least US$518 billion to infrastructure across six partners over a decade. The figure is the news. The structure is the story: roughly four-fifths of that is non-cancelable or payable even if the company never uses the capacity. The Google and Amazon contracts require minimum payments if actual spend falls short. The Broadcom lease obligations, around $161.2B, neither side can cancel except in default. The Microsoft commitment runs November 2026 through May 2033.

Most readers will read the $518B as a balance sheet. It is closer to a power purchase agreement. The only optional block in the stack, up to $84.5B of Nvidia-based capacity through xAI, is cancelable on 90 days' notice. That contrast, locked-in versus optional, is the picture: a serious frontier entrant now has to pre-commit decade-long infrastructure spend across multiple providers before it ships the product.

The mechanism is portable. Any lab that wants to sit at the frontier has to sign the same multi-vendor, long-horizon, capital-intensive contracts, or fall behind. Compute is the binding constraint on AI development, and it is now priced like long-haul energy, not like a cloud bill. That re-prices who can compete, and what it costs to try.