AWS "Sold Out" Through 2027? What Jassy Said and What Is Verified

AWS is sold out through 2027 and a good chunk of 2028, according to headlines that spread on October 23, 2026. The underlying statements are real, but they are not the same statement. Andy Jassy's July earnings-call remarks describe a capacity shortfall for 2026 and 2027. The 2028 line comes from a later podcast conversation, relayed by The Information and others, with no percentage attached.

This post lays out the quote trail, the numbers around it, and what is still unverified, then what it means if you build on AWS. We could not open Amazon's own transcript or the paywalled reports directly, so where we rely on secondary coverage we say so.

TL;DR: what is claimed and what is verified

What Jassy said on the July earnings call

According to Fortune's report of the July 30 call, Jassy said Amazon would not "have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027, too." Fortune also reports AWS at a $169 billion annualized revenue run rate, Q2 AWS revenue of $42.2 billion up 37 percent, and a $496 billion backlog.

The same coverage says Amazon raised its 2026 capital expenditure forecast to $220 billion from $200 billion, mainly because of higher memory costs. Fortune says Jassy indicated AWS is on pace to double its power capacity by the end of 2027, compared with 2025.

Note what the wording is and is not. "Not enough capacity to meet all demand" is a shortage statement: some demand goes unmet. It is not the same as "no capacity available to anyone." AWS typically keeps some capacity for on-demand and managed services such as Bedrock and EC2, and the podcast coverage says Jassy made that point himself.

What changed on October 23: the 2028 line

On October 23, The Information ran a briefing headlined that AWS is sold out of capacity into 2028, Jassy says. Other write-ups attribute the remarks to a podcast hosted by Andreessen Horowitz and quote Jassy as saying "a good chunk of '28 is sold out." Social posts summarizing the conversation went further, saying AI compute is "sold out beyond 2028," which does not match the quoted line.

That is a useful example of how a claim inflates in transit. Step by step:

1. Earnings call, July: shortfall in 2026, same expected in 2027.
2. Podcast, October: AWS sold out for 2026, largely sold for 2027, a good chunk of 2028 (as reported).
3. Social summaries, October 23: "sold out beyond 2028."

Only the first step is something we could trace to a direct quote. The second depends on the reports' paraphrase, and the third is not supported by anything we found. We could not open the full recording or a transcript, so we are not able to confirm the exact words, the date, or whether "good chunk" referred to committed contracts, reserved capacity, or expected demand.

Update (October 30, 2026): Amazon's Q3 call puts numbers on the demand

Amazon reported third-quarter results on October 29, and the call gave the capacity story its first hard figures. According to AI Weekly's summary of CNBC's reporting, AWS revenue grew 39 percent to about 46 billion dollars, its fastest pace in nearly five years. Jassy said AWS has a contracted-revenue backlog of 578 billion dollars, "more than double" a year earlier, and guided AWS growth of more than 40 percent in both 2027 and 2028.

On spending, the same summary says Amazon still plans roughly 220 billion dollars of capital spending in 2026 and will spend "meaningfully more" in 2027. The same summary puts Q3 property-and-equipment spending at 77.3 billion dollars, nearly 20 billion above expectations. We have not opened the call transcript, so treat the quotes as secondary until you check it. The backlog is consistent with the "sold out" framing above, but it does not by itself show which years the contracts cover: a backlog counts signed commitments, which can stretch well beyond 2028. For how this connects to AI chip supply, see Anthropic's S-1 risk factor on TSMC.

The timeline of Jassy's capacity remarks

The "sold out" theme did not start in October. Seen in order, the statements have steadily stretched further into the future:

• April 2026, shareholder letter. Coverage of the letter reports Jassy saying capacity for services running Amazon's Trainium3 chips was nearly fully subscribed, that two large customers asked to buy all of Amazon's Graviton capacity for 2026 and were refused, and that power was the main bottleneck, with 3.9 gigawatts added in 2025 and a plan to double total power capacity by the end of 2027. These details come from secondary write-ups, so check the letter itself.
• July 30, 2026, Q2 call. A shortfall for 2026 and an expected one for 2027, with the capex forecast raised to about $220 billion.
• October 2026, podcast. Sold out for 2026, largely sold for 2027, and a good chunk of 2028, as reported.

Each step is consistent with the one before, which is a reason to think the October line is not a misquote. It is also why precision matters: a reader who sees only the last headline loses the difference between a forecast of shortage and a statement of contracts already signed.

What people are asking

Does sold out mean I cannot launch GPU instances today? No. It means total demand exceeds supply, so the marginal customer waits, pays more, or takes a different instance family. Many accounts still get capacity on demand in many regions; availability varies by region and instance type.

Is this only about AI? Not entirely. The reports describe growth in both AI and non-AI workloads, which is why memory has to be rationed between accelerators and ordinary servers.

Will other clouds be better? We have no data in the sources we could open comparing providers. Other hyperscalers have described similar constraints in their own results, but check their statements directly rather than relying on this post.

Should I buy hardware instead? The memory shortage makes that harder as well, as the April Register report suggests, so owning servers is not an easy escape.

Why memory keeps showing up

The reports of the interview tie the shortage to physical limits: only so much data-center space, power and memory, and Jassy reportedly saying he does not expect the memory shortage to end in 2027. A trade-off he is said to describe is allocating memory between high-bandwidth memory for AI accelerators and DRAM for ordinary servers. One source in our search says HBM takes about three times the wafer capacity of standard DRAM per gigabyte. We could not check that figure against a primary source, so treat it as indicative.

The memory story has its own trail on explainx.ai. Chip prices have been rising, as in our post on US chip producer prices up 27 percent. Earlier in the year The Register reported AWS saying the server memory shortage was pushing customers toward cloud rental instead of buying their own servers, which fits a picture of demand outrunning supply.

What this means for what you build or pay

This is not a stock story; the practical effects land on builders.

• Reserve rather than hope. If you rely on GPUs or large instance families, tight supply favors customers with reservations and commitments. AWS already raised Capacity Blocks rates by roughly 15 percent on October 7, which we covered in AWS Capacity Blocks GPU rates rise.
• Expect quota friction. New accounts and large increases may take longer to approve. Ask early and ask for specific regions and instance types.
• Design for portability. Containerized inference, standard APIs and a second region or provider reduce the risk of a capacity wall. Our guide to GPU schedulers and fair-share shows ways to squeeze more from fixed capacity.
• Use managed inference where it fits. Services with pooled capacity can absorb spikes better than a single reserved instance, though they carry their own limits.
• Watch prices, not only availability. Shortages tend to show up first as price rises on scarce options.

Does it mean a data-center bubble or a boom?

Opposite readings are possible, and the evidence here supports neither on its own. Sold-out capacity for years ahead suggests strong demand, which is how Ben Horowitz is reported to have framed it when contrasting the moment with the dot-com era. But backlog and bookings are contracts, not delivered revenue, and customers' commitments depend on their own funding. Financing concerns around cloud operators are a live debate; see our coverage of Larry Fink's data-center moratorium warning.

Supply is also a political question. Local pushback against new sites is growing, for example protests in Thane, India and New York's data-center moratorium debate. Power and siting limits are part of why Jassy talks about capacity in years, not quarters.

How to check the claim yourself

• Read Amazon's Q2 2026 earnings release and call transcript on its investor relations site, and search for "capacity."
• Find the full podcast recording and search for the 2028 remark; note the date and the exact phrase.
• Compare the backlog figure in Amazon's filing with the $496 billion in coverage.
• Check your own AWS console: Service Quotas and Capacity Blocks availability for the instance types you need.

Bottom line

Verified: Jassy said in July that AWS will lack capacity for all 2026 demand and expects the same in 2027, per Fortune. Reported, not fully verified: a good chunk of 2028 is already sold, from a podcast. Not supported: "sold out beyond 2028" and any fixed percentage of 2027 or 2028 booked. For builders the safe reading is a persistent supply squeeze, not a hard stop.

Facts are accurate as of October 23, 2026. Primary sources were not fully accessible when we wrote this; figures should be checked against Amazon's filings.

• AWS Capacity Blocks GPU rates rise about 15 percent
• US chip producer prices up 27 percent on AI data centers
• Larry Fink and the data-center moratorium warning
• Amazon Thane data-center protests
• New York data-center moratorium explained
• GPU cluster scheduling and fair share
• Tencent and Oracle 100,000-chip lease
• Sources: Fortune on the July call, The Information briefing