AWS compute is sold out through 2027, says CEO Andy Jassy

The cloud a team planned to run its AI workload on in 2027 may already belong to someone else.

The Information reported on Friday that Amazon CEO Andy Jassy said AWS is largely out of capacity. He was talking about 2026 and 2027. The outlet quotes him directly. "'26 is obviously at this point sold out. [2027] is largely sold out. We always leave a certain amount for on demand and our inference." The quote came from an Andreessen Horowitz podcast. The podcast was released on Friday. Catherine Perloff wrote the briefing.

What "sold out" means for a cloud this size

AWS is the largest cloud provider by revenue. Its capacity is measured in regions. Each region is a cluster of data centers with its own power contract and its own build timeline. A region fills up through long contracts for reserved compute. Reserved compute is how large customers plan AI training and large inference tiers.

Jassy's words name two different things. The 2026 figure refers to committed capacity for the current year. The 2027 figure is about capacity a buyer would be reserving today. His caveat about on-demand and inference matters. AWS keeps a share of each region unbooked. A developer running a small inference job will still get the instances. A team wanting to reserve a GPU cluster for training is in the smaller remaining pool.

Two separate scarcity numbers

The outlet's reporting does not state how far through 2027 the sold-out line falls. It does not say how much of 2028 is already spoken for. It does not break the number down by region, by instance family, or by GPU type. All of those matter more than the whole-fleet number for a team planning a workload.

Buyers can read two things from the briefing. The first is that reserved AWS capacity for next year is scarce. The second is that AWS is willing to say so in public, which is itself new. A cloud provider has reasons to keep a scarcity number quiet, because the number is also an invitation to competitors. A CEO stating it in a podcast is a signal that AWS now wants customers to hear it.

The planning problem it creates

A team planning an AI model release in the first half of 2027 now has three options. One is a reserved pool booked now. One is a smaller on-demand pool at a higher per-hour price. One is a different cloud. Each option has a date attached. The reserved pool depends on when the capacity physically exists. The on-demand pool depends on the region having a free slot the day the workload runs. A different cloud starts a different procurement timeline.

The reporting is behind a subscription paywall and names one quote on one podcast. The scarcity it describes comes from a single CEO's statement about one company. A market-wide figure for cloud capacity, measured independently, is a separate thing nobody has published this week.

Source

AWS Largely Sold Out Of Capacity Into 2028, Jassy Says, The Information, by Catherine Perloff.