AWS Is Sold Out Through 2027 as AI Demand Outruns the Buildout

Sold out, and still not enough

Amazon CEO Andy Jassy said AWS has sold out its cloud capacity for 2026 and most of 2027, with a good chunk of 2028 also committed, as AI demand outpaces the company's buildout. Amazon has guided to about $220 billion in 2026 capital expenditure, and Jassy signaled it could spend more. That is the sound of the world's largest cloud provider admitting it cannot build fast enough for the customers lining up.

The sold-out picture matters beyond Amazon. When AWS runs out of capacity, it signals that spending on AI infrastructure will stay elevated across every hyperscaler, supporting demand for accelerators, memory and data center equipment. It also gives Amazon unusual pricing leverage in contract renewals: customers who need capacity cannot credibly threaten to leave if there is nowhere to go.

Jassy added a striking detail: AI inference remains roughly ten times too costly for widespread business use. Training gets the headlines, but inference, the cost of actually answering questions at scale, is where most business AI spending will eventually land. If inference stays ten times too expensive, the AI boom has a cost problem that no amount of model cleverness alone will fix.

Trainium is the cost answer, if it works

Amazon's answer to that cost problem is its own silicon. Jassy said Amazon's Trainium chips offer 30 to 40 percent better price-performance than comparable processors, and the company is betting that in-house chips can erode Nvidia's share of inference workloads, where cost per query matters more than raw training performance.

That bet is also visible in Amazon's footprint. In Korea, AWS is building a 100-megawatt AI data center in Ulsan with SK Group, with operations expected to begin in stages in the second half of 2027. The pattern is consistent: build where power and partners are available, fill it with your own chips where possible, and sell the capacity before the concrete dries.

The capacity crunch is not unique to Amazon. Anthropic's $750 billion cloud bill exposed the same capacity war, and AWS is already paying billions for Nvidia GB300 racks it cannot switch on yet because power, not hardware, is the bottleneck. The industry's constraint has moved from chips to electricity, and from electricity to the time it takes to build anything at all.

What this means for cloud customers

If you buy cloud capacity, the sold-out headline has direct consequences:

• Book early, and book longer. Capacity committed through 2028 means short-notice, short-term deals will cost more or simply not exist. Planning horizons that used to be quarters are now years.
• Expect renewal pressure. When a provider is sold out, your negotiating leverage at renewal shrinks. Multi-cloud strategies and portable architectures are worth more in a seller's market.
• Watch inference costs, not just model quality. A model that is ten times cheaper to run can be a better business choice than a model that is slightly smarter. Trainium's pitch is aimed exactly at that calculation.
• Follow the power story. Data center announcements now hinge on grid access as much as on chips, as DayOne's gigawatt-scale Johor build shows. Where a provider can get power is where your capacity will actually come from.

Amazon reports third-quarter results on October 29, and traders expect the AI chip strategy to feature prominently. Prediction market Kalshi assigns a 96 percent probability to executives mentioning Trainium and 95 percent to Bedrock, while pricing only a 15 percent chance that layoffs come up. The market is telling you what it cares about: chips, capacity and cost, in that order.

The bigger picture

A sold-out cloud is a strange kind of success. It means demand is real, not speculative, and that customers are willing to commit years ahead for capacity that does not exist yet. It also means the AI buildout is now limited by construction crews, power grids and supply chains rather than by belief. Belief was the easy part. The next phase of the AI boom will be decided by who can actually pour concrete, secure megawatts and ship racks on time, and right now even the biggest player in the business is telling you it cannot keep up.