AWS Capacity Largely Sold Out Through 2027, With On Demand Headroom Reserved

AWS being “sold out” through 2027 refers mainly to large, committed capacity blocks. These are planned allocations that customers secure for future use, often at a scale ordinary users cannot obtain instantly. It does not mean every AWS server is unavailable today.

Jassy said 2026 capacity is already gone, 2027 is largely sold out, and a good chunk of 2028 is allocated. AWS still reserves headroom for On-Demand customers, EC2, and Bedrock, its inference service. That creates a difference between long-term reservations and daily availability.

David Linthicum argued that a company could probably still open the AWS console and obtain some GPU capacity. The concern is finding a specific, massive block for several years. Current reports also describe shortages affecting ordinary EC2 instances, so availability may be tightening even outside major AI deals.

The article gives no single quantity for the computing capacity already committed. Instead, Jassy describes its timing: AWS has no remaining capacity for 2026, has largely sold out its big 2027 blocks, and has already allocated a good chunk of 2028. AWS still holds some headroom for On-Demand, EC2, and Bedrock.

The investment figure is roughly $220 billion in planned capital expenditure for the year. That spending covers Amazon’s total capital expenditure, according to the article. It should not be read as a dedicated AI infrastructure budget.

Jassy said Amazon could invest well beyond that amount without making a poor capital allocation decision. Amazon’s third-quarter results, due October 29, are expected to include remaining performance obligations. That disclosure may show how much contracted revenue is connected to these commitments.

On-Demand Capacity Reservations let customers reserve instances in a named Availability Zone. Capacity Blocks for ML do something similar for machine-learning resources, but for a fixed future window. Reserved instances are another commitment tool that teams use to secure capacity against expected growth. Ordinary EC2 availability is what a customer can obtain day to day.

AWS Capacity Blocks cover listed systems such as P5, P5e, P5en, P4d, Trn1, and Trn2 instances. A customer selects a future start date and a fixed usage window. By contrast, a team checking the console may find some immediate EC2 or GPU capacity without owning a large future block.

These categories matter because “capacity” can mean different things. A provider may have daily capacity available while its largest future blocks are committed. The article also notes that AWS raised Capacity Blocks for ML rates by about 15% across every region in January.

When large customers reserve capacity years ahead, newer customers may find fewer suitable resources when they need them. The impact can include delayed deployments, difficulty expanding production systems, and pressure to accept less suitable or more expensive options. The article does not claim that every workload will be blocked.

One developer reported trouble obtaining ordinary EC2 instances, including On-Demand capacity, while Spot instances were especially difficult. Production Kubernetes clusters became concerningly constrained on some days. Another operations manager said his team adopted reserved instances to protect projected growth.

That response can change cloud behavior. Ian Rae warned that customers may reserve resources simply because they fear needing them later. This can reduce reservation efficiency and create an “anti-cloud pattern,” even though cloud services are normally valued for flexible, on-demand allocation. AWS provides EC2 Capacity Manager to track used and unused reserved capacity.

Large AI model providers need computing at a scale that differs from ordinary cloud usage. They may require many GPUs or specialized processors together, with predictable access over a long period. Securing capacity early reduces the risk that future infrastructure will not be ready when a model program begins.

AWS offers Capacity Blocks for ML, where customers reserve instances for a fixed window beginning on a future date. The listed hardware includes P5-family and P4d GPU instances, plus Trn1 and Trn2 processors. Jassy said the largest 2027 blocks are largely sold, with a good chunk of 2028 allocated.

The article does not identify the customers taking those blocks or state exactly why each reserved them. It does show the broader trade-off: AWS is considering major investment, while large commitments provide planning certainty and may leave less flexible capacity for smaller or newer customers.

If AWS cannot supply enough capacity, a company can broaden its search. It might use another Availability Zone, select a different EC2 instance type, or shift some workloads to another cloud provider. Spot capacity is another possible route when interruptions are acceptable, though the article reports that one team struggled to obtain it.

The practical mechanism is substitution. A workload may fit a different processor or region, while a multi-cloud design can draw resources from more than one provider. However, moving workloads may require engineering changes, and another Availability Zone may not offer identical capacity or deployment characteristics.

The article gives no guarantee that alternatives will solve the problem. One commenter reported insufficient compute from all three major providers, and another described Microsoft Azure East US constraints. Capacity problems therefore may extend beyond AWS and beyond GPUs, making early planning and flexible architectures more important.

Cloud capacity exists only where providers have enough physical infrastructure. Data centers need servers, GPUs or other processors, electricity, cooling, storage, and networking. Providers also need capital to buy equipment, construct facilities, and connect them. If any part is constrained, the amount of usable capacity can fall.

A large AI workload may require many matching GPUs in one place, rather than scattered machines. That makes capacity harder to supply than a simple count of unused servers suggests. AWS Capacity Blocks and On-Demand Capacity Reservations let customers secure particular resources or locations ahead of time.

The article focuses on the outcome, not a detailed inventory of these physical limits. It reports AWS’s roughly $220 billion total capital-expenditure plan and Jassy’s view that Amazon could invest more. It also shows demand reaching into 2028, while ordinary EC2 users report constraints today.