AWS capacity is largely booked for 2027, Jassy says, as Amazon builds for AI demand

Amazon Web Services has largely committed its 2027 computing capacity and already booked significant portions of 2028, chief executive Andy Jassy said in a podcast released Friday. The statement gives investors a fresh indication of demand behind Amazon's artificial-intelligence expansion. It also raises a more practical issue for customers: obtaining the right infrastructure when a project needs it.

Jassy preserved an important qualification, according to Reuters' account of the interview with Andreessen Horowitz's Ben Horowitz and Erik Torenberg. AWS keeps some capacity for on-demand services, including EC2 and Bedrock. The comments describe large commitments and tight supply, rather than an announcement that AWS has stopped accepting ordinary cloud customers.

Committed capacity and available services are different measurements

EC2 supplies virtual computing servers for applications and other workloads. Its on-demand model charges for usage without an upfront commitment, while other purchasing arrangements trade commitments for different prices. Bedrock provides access to AI models and tools for building applications. These are services with different ways of consuming computing resources, rather than a single interchangeable pool that every customer buys in the same form.

That helps explain how a provider can have substantial future capacity committed while continuing to serve on-demand demand. For a particular customer, the relevant questions include the location, hardware, service and time required. A broad description of next year's bookings cannot establish that every combination is either available or unavailable, nor does it supply a public percentage for each region.

This is where the physical bottlenecks behind AI computing become a commercial issue. A software team can change a deployment plan much faster than a provider can bring an entire new facility into operation. Capacity planning therefore links a customer's launch schedule to the provider's investment decisions, even when the customer never owns a server.

The spending is already visible in cash flow

The a16z episode describes Amazon's roughly $220 billion capital-expenditure plan for 2026 and Jassy's view that demand could support more investment. That is an account of management's spending expectations and assessment of opportunity. It is not a new disclosure that all of the money has already been spent, nor a separately verified forecast of the return the assets will earn.

Amazon's previously published second-quarter figures give the discussion a financial baseline. AWS sales reached $42.2 billion, up 37% from a year earlier, with operating income of $16.6 billion. Across Amazon as a whole, trailing twelve-month operating cash flow reached $161.4 billion, yet free cash flow was an outflow of $7.6 billion. The company attributed the decline chiefly to higher property and equipment purchases reflecting AI investment.

Those figures cover different scopes. AWS revenue and operating income describe a segment; the cash-flow measures describe the broader group. They should not be combined into a supposed stand-alone AWS free-cash-flow calculation. Our guide to free cash flow explains why spending on long-lived equipment can create a gap between reported operating profit and cash remaining after investment.

Bookings support the growth case, but do not settle the return question

A commitment for future computing gives a provider more information than an expression of interest. It can help shape the timing and size of an investment. It still differs from revenue already recognized for services delivered. The customer must use or pay for the contracted service under its terms, and the provider must supply the infrastructure required to fulfill that obligation.

Investors therefore need both sides of the equation. Faster consumption can support sales, while purchasing, powering and maintaining infrastructure affects its economics. Scarcity can demonstrate strong demand while also delaying the activity from which a provider expects to earn money. Neither a crowded order book nor a large construction budget alone establishes the eventual margin on the new capacity.

For enterprise buyers, a reservation also needs to match the intended workload. An application that runs economically on one configuration may need adjustments to use another. The cost of running AI after training includes utilization and workload design, not simply the advertised price of a unit of computing power. Availability and efficiency can consequently pull a deployment decision in different directions.

Friday's comments strengthen the evidence that Amazon sees continuing demand for its expansion. The next financial report will offer a more comparable test through actual AWS growth, operating results and investment cash flows. Until then, the distinction to preserve is straightforward: much of the large-scale future capacity is spoken for, while Amazon is still reserving resources for customers buying services on demand.