Amazon Web Services has already committed much of the computing capacity it expects to have available over the next several years, according to CEO Andy Jassy. In an interview released by venture firm Andreessen Horowitz (a16z) on October 23, Jassy said AWS capacity for 2026 was sold out, 2027 was largely sold out, and a good portion of 2028 capacity had been committed. The remarks describe bookings and customer commitments as Jassy characterized them, not an audited measure of every AWS server, region or service. [a16z interview] [Reuters]
What Jassy said about AWS commitments
Reuters reported on October 23 that Jassy described AWS cloud-computing capacity for 2026 as sold out, said 2027 was “largely sold out,” and said a good chunk of 2028 was sold out. The interview episode, titled “Amazon CEO Andy Jassy on the $220 Billion AI Buildout,” was published by a16z on Apple Podcasts at 10:00 UTC on October 23. Reuters said Jassy made the comments in a video podcast conversation with a16z co-founder Ben Horowitz and general partner Erik Torenberg. [a16z episode listing] [Reuters report]
The wording is important. Jassy did not say all AWS capacity through 2027 was unavailable to every customer, nor did he publish a utilization rate, a regional breakdown or a count of unfilled orders. “Largely sold out” leaves room for capacity not yet committed, while the statement about 2028 referred to a good portion rather than the entire year. Reuters also reported that Jassy said Amazon could invest considerably more and that doing so would not be a bad use of capital, given demand and sold-out 2026 capacity. This is an executive’s description of commercial demand and commitments, not a formal capacity disclosure or guarantee of future delivery.
On-demand and inference are explicit qualifications
Jassy qualified his description of 2027. As quoted by Reuters, AWS reserves some capacity for on-demand use and for inference services, including Bedrock and EC2, while saying the “big chunks” were largely sold out. The remarks therefore distinguish between major capacity blocks already committed and capacity that AWS keeps available for other customer access or workloads. They do not explain how those pools are sized, whether they cover every geography and instance type, or how the company accounts for reservations, expansions and customer usage over time. [Reuters]
That distinction matters for customers trying to translate a headline into practical availability. AWS sells many kinds of compute and managed services, and a broad statement about capacity commitments does not establish that a particular instance, model, region or service is inaccessible. Jassy’s comments do not specify whether “sold out” means signed contractual commitments, reserved capacity, expected demand against planned supply, or some combination. The safest reading is limited: according to the CEO, most of the large capacity blocks for 2027 have been committed, while AWS maintains some capacity for on-demand and inference needs.
Bookings are not the same as physical capacity
A commitment is a commercial signal; it is not itself a measure of how much equipment is installed and running. Customers can commit to future computing capacity before the facilities, power systems and servers that will support it are fully deployed. Conversely, AWS may have physical infrastructure that is not captured by a simplified “sold out” label because it is reserved for a different customer, service, location or workload. Neither the interview listing nor Reuters supplied a technical definition of the term or a reconciliation between commitments and operational capacity.
That is why Jassy’s statement should not be rewritten as proof that Amazon has run out of data-center space, electricity, chips or usable servers. It could reflect demand outpacing capacity AWS is willing or able to sell under particular planning assumptions, but the sources do not quantify any such gap. Reuters framed the remarks around strong demand for AI infrastructure and noted that AI labs and enterprises were seeking capacity. The comments do not isolate how much demand comes from AI training, inference, conventional cloud workloads or specific customers. [Reuters]
Why reserved capacity and service mix matter
On-demand compute, reserved commitments and inference services can compete for infrastructure but are not interchangeable products. On-demand access generally refers to capacity a customer can request without the same long-term reservation structure. Inference means running a trained model to answer prompts or produce outputs; AWS offers model access through Bedrock and compute options through services such as EC2. Jassy’s acknowledgment that AWS leaves capacity for on-demand and inference makes a categorical “nothing is available” interpretation especially misleading. It also suggests the company manages the allocation of supply across customer segments and services, although he did not spell out the rules or trade-offs.
For buyers, the relevant question is therefore not simply whether AWS is “sold out,” but what capacity is available for a specific workload, location and timeframe, and on what commercial terms. The interview offers no customer-level availability data. It also does not establish whether a customer facing a constraint could use another AWS region, a different instance family, an alternative service or another cloud provider. Those questions require current, workload-specific information rather than extrapolation from a company-wide executive comment.
Demand meets a larger investment plan
The remarks arrive as Amazon is investing heavily in its infrastructure and AI businesses. In its July 30 second-quarter results release, Amazon reported AWS net sales of $42.2 billion, up 37% year over year, and AWS operating income of $16.6 billion, compared with $10.2 billion a year earlier. The release also said trailing-12-month free cash flow was an outflow of $7.6 billion, driven primarily by a year-over-year increase in purchases of property and equipment, which Amazon said primarily reflected AI investment. [Amazon Q2 2026 results]
Those reported figures offer financial context, but they do not independently verify Jassy’s capacity estimate. Revenue growth measures sales during a period, while operating income reflects segment profitability. Free cash flow and property-and-equipment purchases illuminate investment intensity. None reveals how many megawatts, accelerators, servers or customer commitments AWS has available for each future year. Amazon’s Q2 release also said the company expected about $220 billion in capital expenditure for 2026; the a16z episode title and description likewise refer to roughly $220 billion in 2026 capex. The figure is a spending plan discussed by the company, not a direct measure of capacity already delivered. [Amazon Q2 release] [a16z episode listing]
The strategic tension is straightforward: cloud providers must build infrastructure ahead of expected demand, but large investments carry execution and financial risks if demand, customer commitments or technology requirements change. Jassy’s comments indicate Amazon sees enough demand to commit substantial future supply and continue investing. They do not prove that the planned investment will translate into a particular amount of usable capacity on schedule, or that every committed customer will consume the capacity in the way anticipated.
What customers and investors should watch
Amazon’s next scheduled quarterly report, according to Reuters, is October 29. Investors and customers can look for additional information on capital spending, AWS growth and management’s description of supply and demand. A useful update would distinguish future commitments from capacity already online, clarify whether the estimate applies globally or to selected regions and services, and explain how reserved pools for on-demand and inference are treated. No such detailed breakdown is contained in the interview or Reuters account available for this story. [Reuters]
The new comments are distinct from Amazon’s October 20 TorchNeuron announcement, which concerned software support for running PyTorch workloads on Trainium chips, and from Google and Voltus’s October 23 virtual power plant announcement. Those stories address developer tooling and electricity flexibility, respectively; neither is a report of AWS’s future customer capacity commitments. A screen of the site’s latest 20 published articles found no AWS capacity-bookings story, so this candidate is not a duplicate of those adjacent infrastructure topics.
For now, the headline finding is both significant and bounded. Jassy says the 2026 capacity pool is sold out and the larger blocks for 2027 are largely committed, with a good portion of 2028 also committed. The qualification about on-demand and inference capacity, along with the absence of a disclosed methodology, means that statement should be read as an executive estimate of bookings and demand—not as evidence that AWS has exhausted every physical resource or that customers cannot obtain cloud or inference services.
Frequently asked questions
Did AWS say every kind of capacity is unavailable in 2027?
No. Jassy said 2027 was largely sold out, not completely sold out. Reuters reported his qualification that AWS leaves some capacity for on-demand use and inference services such as Bedrock and EC2. Neither source provides a service-by-service inventory.
Does “sold out” mean AWS has no physical servers left?
The sources do not say that. The statement concerns capacity commitments as described by Jassy. It is not an audited inventory of servers, data centers, power or regional availability, and it does not establish a physical infrastructure ceiling.
When did Jassy make the statement?
The a16z episode was published October 23, 2026, at 10:00 UTC according to its Apple Podcasts listing. Reuters also published its report on October 23 and described the comments as made in the video podcast conversation.
What does this mean for AWS customers seeking AI inference?
It does not show that Bedrock or EC2 inference is broadly unavailable. Jassy specifically said AWS retains some capacity for on-demand and inference services. Actual availability depends on the service, workload, location and timing; the interview did not provide customer-level availability figures.
Does Amazon’s $220 billion capex figure verify the capacity claim?
No. Amazon’s Q2 release and the episode description refer to roughly $220 billion in 2026 capital expenditure, but spending plans are not a direct capacity metric. The figure supplies investment context; it does not independently confirm how much capacity has been booked, built or made available.