AWS Is Largely Sold Out Through 2027 and Part of 2028, Jassy Says, Even With 220 Billion Dollars of Capex

Amazon Web Services has already sold the bulk of its computing capacity through the end of 2027 and has commitments on a substantial slice of 2028, according to chief executive Andy Jassy, a sign that the constraint on the AI buildout has shifted from demand to the physical supply of data centers, power and memory.

Jassy made the comments in a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, recorded on Friday and reported by Reuters and Chinese business outlets. His account is blunt: 2026 capacity is gone, 2027 is effectively gone, and 2028 is going. AWS will hold back some room for on-demand instances and for Bedrock, its model-serving business, but the large blocks that anchor multi-year contracts have been claimed.

The spending number behind that is roughly 220 billion dollars of capital expenditure this year, a figure Amazon disclosed earlier in 2026 and which Jassy said is still not enough. He argued that spending more would not be an unwise use of capital, because the alternative is turning away customers who are willing to commit years ahead.

The way AWS sizes that bet explains why the money takes so long to show up in results. Data centers have to be committed roughly two years before they carry a workload and are depreciated over about 30 years, while servers and networking gear are bought months ahead of use, break even in about three years and last five to six. Most of the customer contracts behind the buildout run about five years, which lines up with the hardware rather than the buildings.

Jassy described demand as dumbbell-shaped. On one end sit foundation model labs and a handful of fast-growing AI applications that want enormous amounts of compute immediately. On the other sit enterprises using AI to cut costs in customer service, fraud detection and back-office process automation. The middle is thin: he said most of the applications companies actually run in production still do not lean on AI inference at all, and he expects that to change over the next five to seven years rather than the next five to seven quarters.

The growth curve he offered as context is unusually steep. AWS took three years to reach about 60 million dollars in annualized revenue, he said, while the company's generative AI business passed 25 billion dollars annualized within roughly three years of the current wave starting. AWS itself reported 42.2 billion dollars of revenue in the second quarter, up 36.7 percent, or about 169 billion dollars annualized.

Jassy also warned that tightness in some infrastructure components will last several years, a forecast that lands in the middle of a memory shortage already reshaping consumer hardware prices. Amazon reports third-quarter results on October 29, when investors will be looking for how much of the backlog converts into revenue and how much margin the pre-sold capacity protects.

The strategic read is double-edged. Pre-selling years of capacity gives Amazon rare visibility and bargaining power, and it makes the capital spending look less like a bet on demand than a queue management problem. It also concentrates the business in a small number of very large buyers, and it means the ceiling on AWS growth for the next two years is set by transformers, turbines and DRAM supply rather than by whether companies want AI. That is a better problem to have than the reverse, but it is still a ceiling.