Take-or-Pay: Why AI Compute Contracts Don't Cancel

The name is the mechanism. A take-or-pay clause obliges the buyer to pay for a contracted quantity whether or not it takes delivery: take the capacity, or pay for it anyway. In a market built on five-year assets financed with debt, this is the clause that makes the financing possible.

Why suppliers insist on it

A data centre takes three to five years and billions of dollars to build. Nobody lends against a forecast of demand, so the lender asks for contracted revenue instead. The supplier cannot sign that contract without shifting the volume risk somewhere, and the only party with the risk appetite is the buyer — at a discount, in the form of a lower rate per unit of capacity.

What it looks like in a filing

The clearest recent example is Anthropic's prospectus, as reviewed by Reuters in September 2026: $518 billion of commitments over a decade, with about 80% described as non-cancellable or payable regardless of usage, against 2025 revenue of $4.6 billion. That ratio is the whole argument in one line — the obligations are two orders of magnitude larger than the revenue currently arriving to service them.

What it does to the buyer

It converts a variable cost into a fixed one. If utilization disappoints, the buyer does not pay less; the buyer pays the same and earns less. That is why the clause matters more than the headline price: a 30% cheaper contract with a hard take-or-pay floor can be more expensive than a spot rate, in any scenario where demand does not arrive on schedule.

Five things to check before signing

How “take” is defined and where it is measured; whether shortfalls carry forward; whether the commitment is by capacity or by spend; what happens on termination for convenience; and which force-majeure carve-outs survive. None of these are exotic. All of them are where the risk actually sits, and none of them appear in a headline price per GPU-hour.