Amazon Cloud Capacity Is Largely Sold Out Through 2027: What It Signals for AI Infrastructure

Key Takeaways

• Amazon CEO Andy Jassy said on Friday 23 October 2026 that Amazon Web Services cloud computing capacity is largely sold out through 2027, with significant commitments extending into 2028, according to Reuters.
• Jassy said 2026 capacity is already sold out and that Amazon could invest a lot more in capacity without it being a bad use of capital, in comments made on a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg.
• The statement is about booked capacity, not about Amazon running out of cloud services altogether: Jassy said Amazon always leaves a certain amount for on-demand use and for its inference service Bedrock and its EC2 server service.
• Amazon had said earlier in 2026 that it would spend about 220 billion dollars in capital expenditure for 2026, a figure Reuters repeated in its report on the remarks.
• For businesses, the practical message is to plan AI capacity early, confirm commitments in writing, and avoid assuming that large blocks of accelerated computing will be available on short notice.
• Amazon is scheduled to report third-quarter results on 29 October 2026, which may add detail to a picture that currently rests on one podcast interview.

Amazon Web Services is, by its chief executive’s own account, struggling to keep pace with demand for artificial intelligence infrastructure. On Friday 23 October 2026, Amazon CEO Andy Jassy said that Amazon Web Services cloud computing capacity is largely sold out through 2027, with significant commitments extending into 2028. Reuters reported the remarks, which Jassy made on a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg.

That distinction shapes everything else in this article. A sold-out year for a cloud provider does not look like a sold-out concert. It looks like a calendar in which the biggest rooms are booked, the remaining space is fragmented, and anyone arriving late must either take what is left, wait, or pay for priority. For companies planning AI projects for 2027, that is a planning problem first and a pricing problem second.

What Jassy Actually Said

Jassy’s comments came in a conversation, not in an earnings call or a formal filing, and that context matters for how much weight to put on them. His central line, as reported by Reuters, was: “2027 is largely sold out. We always leave a certain amount for on demand and our inference service bedrock, or EC2, but largely sold out all the big chunks. A good chunk of 2028 is sold out.” He said Amazon is struggling to keep pace with what Reuters described as insatiable demand from AI labs and enterprises. He also said that 2026 cloud-computing capacity is already sold out, and added that the company could invest a lot more and it would not be a bad use of capital.

Earlier in 2026, Amazon had said it would spend about 220 billion dollars in capital expenditure for the year, according to Reuters. That figure gives the podcast remarks their scale. This is not a company trimming investment while demand softens. It is a company already spending at an enormous rate and still finding that demand runs ahead of supply.

It is equally important to note what Jassy did not say. The coverage reviewed for this article attributes no specific capacity figure to him, in megawatts, numbers of chips, or numbers of data centres. It does not name the customers behind the commitments, and it does not break down how much of the booked capacity is for training large models versus running them for users, which the industry calls inference. Any article attaching precise customer names or capacity totals to this announcement should be treated with caution unless it cites a separate, named source. This article does not invent those details.

Why AI Demand Is Eating Cloud Capacity

To understand why a cloud provider can be sold out two years ahead, it helps to separate ordinary cloud computing from AI computing. Traditional cloud workloads, such as hosting websites and running business software, grow steadily and can often share hardware efficiently. AI workloads, especially training large models, behave differently. They need large numbers of specialised chips working together at once, they consume far more power per rack than typical servers, and they run for weeks or months at a time. That combination creates a bottleneck that money alone cannot remove quickly: chips must be manufactured and delivered, data centres must be built and connected to power grids, and power itself must often be secured years in advance.

Inference deserves special attention here, because Jassy explicitly carved it out. When he said Amazon leaves a certain amount for on-demand use and for Bedrock, he was signalling that everyday AI use, where a user sends a prompt and gets an answer, is not the part that is fully booked. The scarce resource is the large, committed block: the reserved capacity that a laboratory or enterprise books to train a model or to guarantee heavy, sustained use.

What Sold-Out Capacity Means for Customers

If you are a business leader reading this news, the first question is practical: does this affect my plans? The honest answer is that it depends on what you are buying.

If your company uses cloud services for ordinary hosting, software, and AI features delivered through managed services, a sold-out headline about large capacity blocks is unlikely to change your day-to-day experience. On-demand capacity is, by Jassy’s own account, deliberately held back for that purpose. Most smaller companies work through exactly this kind of flexible, managed access, as our guide to how AI can help with your business shows. If, on the other hand, your company is planning a large AI training run, a dedicated cluster, or a multi-year reserved commitment, the message is clear: start the conversation early, get terms in writing, and build contingency into your timeline.

There are three concrete steps worth taking now. First, ask your cloud provider or reseller for a written statement of what capacity is actually reserved for you, for which period, and under what conditions it can change. Second, separate your must-have capacity from your nice-to-have capacity: training can often be scheduled flexibly, while a customer-facing service that must respond in real time cannot. Third, consider a multi-cloud or hybrid approach for parts of the workload that do not depend on a single provider’s specialised hardware. Running models locally is one option for some teams, and our guide to running AI models locally on your Mac in 2026 explains what that route realistically requires. None of this means panic buying: committing to capacity you do not need, at prices you have not tested, is a worse outcome than waiting.

What It Signals About the AI Infrastructure Race

Jassy’s remarks also tell a broader story about competition among cloud providers. When the chief executive of the largest cloud provider says demand exceeds supply through 2027, he is describing an industry-wide condition, not a company-specific failure. Building data centres, securing power, and obtaining specialised chips are constraints that affect every major provider, even if their exact positions differ. Finimize, covering the same remarks, notes that pre-booked multi-year capacity can act like a backlog, giving AWS unusually clear visibility into future demand, while also shifting attention to execution and cash flow as new data centres are built.

There are two ways to read the claim that Amazon could invest a lot more without it being a bad use of capital. The generous reading is that demand is genuinely strong enough to justify further spending. The sceptical reading is that a CEO talking on a podcast has an interest in presenting demand as overwhelming, because that framing supports continued heavy investment and reassures customers that they should commit early. Both readings can be true at once. The test will come on 29 October, when Amazon reports results and analysts can compare the podcast language with revenue, backlog, and guidance.

The Risks Behind the Headline

A sold-out cloud market carries risks as well as opportunities. The first is cost: when supply is tight, prices for scarce capacity tend to rise, and smaller companies can find themselves priced out of the largest configurations. Managed services and shared infrastructure blunt that effect, but they do not remove it entirely. The second is concentration: if the largest blocks are booked years ahead by the largest customers, newer entrants may struggle to secure the resources they need to compete. The third is environmental: data centres consume large amounts of electricity and water for cooling, and rapid build-out puts pressure on local grids. Jassy’s remarks do not address these issues, but any honest account of a data-centre building boom must acknowledge them. Finally, there is the risk of reading too much into a single interview. Podcast remarks are not audited financial statements, and they should be weighed alongside earnings data and independent reporting before being treated as settled fact.

What to Watch Next

Three disclosures will sharpen this picture: Amazon’s third-quarter results on 29 October 2026, where analysts will test the podcast language against revenue and guidance; any formal AWS announcement about new regions or data-centre projects; and comparable commentary from rival cloud providers, which will indicate whether sold-out conditions are industry-wide or specific to AWS.

For most readers, the actionable summary is simple. AI infrastructure demand is strong enough that the largest cloud provider says its big capacity blocks are largely booked through 2027. That does not mean AI is closed to smaller businesses, and it does not mean every workload needs a multi-year reservation. It does mean that large, committed AI capacity should be treated as a strategic purchase with a long lead time, negotiated early and documented carefully.

Sources

• Reuters, “Amazon cloud capacity sold out through 2027, CEO Jassy says”, 23 October 2026: https://www.reuters.com/business/retail-consumer/amazon-cloud-capacity-sold-out-through-2027-ceo-jassy-says-2026-10-23/
• Finimize, “AWS Says Its AI Cloud Capacity Is Sold Out”: https://finimize.com/content/aws-says-its-ai-cloud-capacity-is-sold-out
• Andreessen Horowitz video podcast with Ben Horowitz and Erik Torenberg, as reported by Reuters, 23 October 2026.