The short version
On October 23, 2026, Amazon CEO Andy Jassy said AWS capacity is sold out for 2026, “largely sold out” for 2027, and a good chunk of 2028 is also booked, even with about $220 billion of capital spending planned for 2026. AWS still holds some capacity back for on-demand EC2 and Bedrock, but the big blocks are already committed. If your product needs GPUs or a lot of compute in the next 12–24 months, treat capacity as a planning item in your cloud and DevOps roadmap. Reserve it early, use it efficiently, and keep the option to move.
What did Andy Jassy say about AWS capacity?
Jassy was speaking on The a16z Show, the Andreessen Horowitz podcast, with co-founder Ben Horowitz and general partner Erik Torenberg, in an episode released on Friday, October 23. He said AWS has already sold out its capacity for this year and that 2027 is “largely sold out” too. He added that AWS always leaves some capacity for on-demand use and for its inference service Bedrock or EC2, but that it “largely sells out all the big chunks.” He also said “a good chunk of ’28 is sold out.” Reuters and The Information both reported the remarks the same day.
This is happening while Amazon is spending more than ever. The company raised its 2026 capital spending plan to about $220 billion in July, partly because memory has become more expensive. Jassy said on the podcast that Amazon “could invest a lot more and it wouldn’t be a bad use of capital.” Demand is coming from AI labs and from enterprises building AI into their products. Neither Amazon nor the reports named specific customers, regions or chip types behind the reservations.
Why is cloud capacity running out?
Money is not the bottleneck here. Physical supply is. Jassy pointed to three limits: data center space, power, and memory. None of these can be added in a quarter. New sites take years to permit and build, grid connections are scarce in many regions, and high-bandwidth memory for AI accelerators is supply-constrained across the whole industry. When the biggest cloud provider says it is booked into 2028 even at this level of spending, its rivals are almost certainly working under similar pressure.
For buyers, the key detail is how AWS is allocating what it has. The biggest share goes to multi-year commitments from large customers, mostly AI labs and enterprises that reserved capacity early. What’s left is kept for on-demand use and for managed inference. In practice, a mid-sized team can still launch instances and call models today, but large or specialized capacity, such as big GPU clusters for training or fine-tuning, is the hardest part to get on short notice, and it may cost more when you do get it.
What it means for US & EU software teams
First, capacity planning is now part of product planning. If your roadmap includes training or fine-tuning models, high-volume inference, or a big migration in 2027, the time to secure compute is before the work starts, not when the backlog reaches it. AWS already offers ways to do this, such as capacity reservations, committed-use pricing and scheduled GPU blocks for ML. The trade-off is that you commit to spend before demand is proven, so size reservations to forecasts you can defend. Over-reserving just moves the risk from availability to cost.
Second, managed services become a safer default. Jassy said AWS keeps capacity aside for inference through Bedrock and EC2. Teams that call foundation models through a managed service such as Amazon Bedrock are drawing on that pool and don’t need to secure their own GPU fleet. For most product teams, that is a reason to prefer managed inference over self-hosted model clusters, unless data rules, cost at scale or model control clearly justify running your own.
Third, efficiency now affects availability as well as cost. When compute is scarce, the team that needs less of it ships sooner. Smaller or distilled models, response caching, batching, and right-sized instances lower your bill and also lower how much scarce capacity you depend on. Portability works the same way. Workloads defined in infrastructure-as-code and containerized on Kubernetes can move to another region or provider when one runs short. For EU teams, check that any fallback region still meets your GDPR and data-residency obligations before you rely on it.
How should teams plan capacity now?
- Forecast 12–24 months of compute. Break it down by workload: training, inference, batch, steady-state. GPU-heavy items go first in the queue.
- Reserve what you can defend. Use capacity reservations or commitments for predictable baseline load, and keep spiky or experimental load on demand.
- Prefer managed inference. Call models through managed services where possible instead of standing up your own GPU clusters.
- Make efficiency a release criterion. Track compute per request or per user, and use caching, batching and smaller models before scaling hardware.
- Keep workloads portable. Use infrastructure-as-code and containers so a capacity shortfall in one region or cloud doesn’t block a launch.
- Check fallback regions for compliance. Confirm that any alternate region meets your data-residency and regulatory requirements before an incident forces the move.
- Revisit pricing assumptions. Scarce capacity can mean less discounting. Budget for that in 2027 cloud plans.
Frequently asked questions
Is AWS really sold out of capacity?
Not completely. On October 23, 2026, Amazon CEO Andy Jassy said AWS capacity for 2026 is sold out, 2027 is “largely sold out,” and a good chunk of 2028 is also sold. He said AWS always keeps some capacity back for on-demand use and for inference through Bedrock and EC2. What is sold out are the large committed blocks reserved by big customers. Teams can still launch instances and call models, but large or specialized capacity, especially GPUs, is harder to get on short notice.
Why can’t Amazon just build more data centers?
Amazon is already planning about $220 billion of capital spending in 2026, and Jassy said it could invest more. The limits are physical. Data center sites take years to build, power and grid connections are scarce, and memory for AI hardware is in short supply industry-wide. Jassy named data centers, power and memory as the bottlenecks.
How can a software team secure cloud capacity?
Forecast compute needs 12 to 24 months ahead by workload, and reserve capacity for the predictable baseline using the provider’s reservation or commitment options. Keep experimental and spiky load on demand. Prefer managed services for model inference, since AWS says it keeps capacity aside for those. Reduce how much compute you need with caching, batching and smaller models, and keep workloads portable with infrastructure-as-code and containers so you can move regions or providers if one is short.
Does the AWS capacity shortage affect EU teams differently?
The reports did not break capacity down by region, so there is no published figure for EU regions specifically. The practical difference for EU teams is compliance. If you plan to fail over or expand into another region because your main one is short, that region must still meet your GDPR and data-residency obligations. Check that in advance, and include compliant fallback regions in your capacity plan instead of picking one during an incident.
Sources
The a16z Show — Amazon CEO Andy Jassy on the $220 Billion AI Buildout (primary source, October 23, 2026)
Reuters (via U.S. News) — Amazon cloud capacity sold out through 2027, CEO Jassy says (October 23, 2026)
The Information — AWS largely sold out of capacity into 2028, Jassy says (October 23, 2026)
CNBC — Amazon hikes 2026 capex to $220 billion due to higher memory costs (July 30, 2026)