AWS has sold the large blocks of 2027 capacity before much of the equipment is bought

Amazon chief executive Andy Jassy says AWS's 2026 cloud capacity is sold out, the large blocks for 2027 are largely sold and "a good chunk" of 2028 is already committed. Amazon keeps some room for on-demand EC2 and Bedrock workloads, so this does not mean an ordinary customer cannot start an instance. It means the hyperscale reservations sought by AI laboratories and large enterprises have absorbed most of the capacity Amazon can confidently promise years ahead.

Demand at that horizon is striking because Amazon does not buy every server and network component years in advance. Jassy told the Andreessen Horowitz podcast that much of the equipment is ordered only months before deployment. That creates an option: if reservations weaken, Amazon can stop purchasing some gear rather than complete every projected cluster. But land, substations, power contracts and buildings have longer lead times and are less reversible. Amazon plans roughly $220bn of capital spending this year and still says it cannot meet all demand.

The commitments therefore answer one question and leave another open. They provide evidence that today's build-out is backed by customers rather than solely by speculative supply. They do not reveal the price, cancellation terms, credit support or concentration of those reservations. A model company that books future capacity may itself depend on another fundraising round; an enterprise may reserve more than it ultimately consumes.

AWS's position also strengthens Amazon's custom-chip strategy. Jassy called Annapurna Labs perhaps Amazon's best acquisition and said the relationship with Anthropic accelerated Trainium development. If power and buildings are scarce, lower-cost, more efficient silicon is not merely a margin story: it determines how much sellable inference fits inside each constrained megawatt.