OpenAI's Greg Brockman: Guaranteed Capacity discounts compute for 1-3 year commits

On May 19 OpenAI President Greg Brockman posted on X (1,509 likes, 60 retweets by May 20) the most concrete public explanation yet of OpenAI's Guaranteed Capacity offering: customers pre-commit to 1-3 year compute purchases and receive both discounted token pricing and capacity-availability guarantees. Brockman's blunt framing — 'we expect that the world will feel increasingly capacity constrained for the next while, as models continue to get much more useful' — is the new fact, replacing prior speculation about pricing structure.

Mechanically, Guaranteed Capacity functions as a forward-contract market for inference. OpenAI sells future tokens at a discount in exchange for demand certainty, customers buy future capacity at a discount in exchange for not getting rate-limited during peak demand. It mirrors AWS Reserved Instances and is the financial counterpart to the physical infrastructure deals (SpaceX-Anthropic-Google compute, Stargate, Dell on-prem) that hyperscale-tier customers have been signing.

The signal value matters more than the mechanics. OpenAI publicly admitting 'capacity constrained' contradicts the bear case that the AI buildout is overshooting demand. Combined with NVIDIA's 95% revenue guide and Dell's 320x token-consumption claim, the picture is one of demand running ahead of supply — not the inverse. The skeptical read is that capacity is constrained partly because OpenAI is selling discounted forward contracts, creating its own constraint.

Competitively, this also creates a moat: customers locked into 2-year OpenAI commitments don't comparison-shop Anthropic or Google daily, which structurally slows churn even if a rival ships a better model. Watch for Anthropic and Google to launch matching products — Anthropic's compute deals with SpaceX/Amazon/Google announced last week make a similar forward-commit pricing tier almost inevitable.