AWS Capacity Sold Out for 2026, Says Andy Jassy

Amazon Web Services (AWS) has sold out its Cloud Computing capacity for 2026, with most of its capacity for 2027 already booked and a significant portion of 2028 capacity also committed, according to Amazon CEO Andy Jassy. The comments highlight the growing pressure on cloud infrastructure providers as demand for Artificial Intelligence (AI) computing continues to outpace available resources.

Speaking on an a16z video podcast released on October 23, Jassy said the rapid growth in demand for AI infrastructure could push Amazon to spend more than its planned $220 billion in capital expenditure for 2026. The company is investing in the infrastructure required to support both AI workloads and its established cloud computing services.

The capacity crunch reflects a broader challenge facing the technology industry: building enough data centres, securing sufficient electricity and obtaining specialised computing components to meet the needs of businesses adopting AI. Even major cloud providers are facing limits on how quickly they can expand their services.

AWS Cloud Capacity for 2026 Is Fully Booked

Jassy said AWS has already allocated its available capacity for 2026, while most of the major capacity blocks for 2027 have also been booked. A considerable portion of capacity for 2028 is committed as well, indicating that customers are planning their computing requirements years in advance.

“’26 is obviously at this point sold out. ’27 is largely sold out. We always leave a certain amount for on-demand in our inference service, Bedrock or EC2, but largely sold out, all the big chunks. A good chunk of ’28 is sold out,” Jassy said.

His comments do not mean that every AWS service is unavailable to new customers. Jassy noted that the company retains some capacity for on-demand usage through services such as Amazon Bedrock and Amazon Elastic Compute Cloud (EC2). However, the major blocks of capacity for the coming years are largely committed.

EC2 provides computing capacity through virtual servers, allowing businesses to run applications without purchasing and maintaining all the underlying hardware themselves. Amazon Bedrock enables customers to build generative AI applications using supported AI models and related tools.

Both services depend on the infrastructure AWS operates, making available computing capacity an important factor in how the company serves customers. As demand rises, Amazon must decide how to allocate its resources between AI-related workloads and other cloud services.

AI Demand Is Increasing Pressure on Data Centres and Power Supply

The growing demand for AI computing is creating challenges beyond the construction of additional data centres. Cloud infrastructure requires reliable electricity, physical space, networking equipment and large quantities of computing hardware. Expanding one part of the system may not be enough if other resources remain constrained.

Jassy specifically pointed to limitations in data-centre capacity, electricity and memory. He also suggested that the memory shortage could continue beyond 2027, underscoring the possibility that supply constraints will persist even as companies continue investing in new infrastructure.

“If you think about it, there’s only so much data center and power today. There’s only so much memory, and I don’t think that that shortage ends in 2027,” he said.

Memory is particularly important for AI systems because modern workloads often require large amounts of data to be processed quickly. High-bandwidth memory (HBM), used alongside AI accelerators, helps supply data at the speeds required by demanding AI applications. Traditional cloud workloads also rely on dynamic random-access memory (DRAM).

These requirements place additional pressure on the supply chain. Building more data-centre space does not automatically resolve a shortage of memory chips or guarantee access to sufficient electricity. Cloud providers must coordinate infrastructure expansion with the availability of hardware and other essential resources.

For AWS, the challenge is to increase capacity while maintaining the services used by businesses for websites, databases, enterprise software and other applications. AI workloads are adding to existing demand rather than replacing the need for traditional cloud infrastructure.

Amazon Could Spend More Than Its Planned $220 Billion in 2026

Jassy’s remarks suggest Amazon could increase infrastructure investment beyond its planned $220 billion capital expenditure for 2026 as it works to address customer demand. The figure represents a planned spending level, while any additional investment would depend on the company’s decisions and requirements.

Capital expenditure on cloud infrastructure can cover data-centre construction, servers, networking equipment, specialised chips and supporting facilities. These investments take time to translate into operational capacity, which helps explain why customers may reserve computing resources well ahead of their intended use.

The scale of Amazon’s planned spending also illustrates the financial demands of the AI boom. Companies operating large cloud platforms must commit substantial resources to infrastructure before they can deliver additional capacity to customers.

Even with higher investment, the pace of expansion may be affected by constraints in electricity, data-centre development and semiconductor supply. Jassy’s comments indicate that Amazon is dealing with several interconnected limitations rather than a shortage that can be solved simply by spending more money.

The situation also creates a planning challenge for businesses that depend on cloud computing. Organisations developing AI products or expanding existing applications need to account for the availability and cost of computing resources when deciding how quickly to scale their operations.

Amazon Wants to Make AI Inference Cheaper

Alongside expanding infrastructure, Amazon is looking to reduce the cost of running AI models. The process known as inference occurs when a trained AI model generates a response, processes a request or performs a task for a user or application.

Inference becomes particularly important when AI applications move beyond testing and begin serving large numbers of users. Every request consumes computing resources, and the cost can rise significantly when a business runs AI systems frequently or across a large customer base.

Jassy argued that inference remains too expensive for AI to be adopted as widely as it potentially could be. “It’s still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible,” he said.

An order of magnitude generally refers to a tenfold difference. In this context, Jassy’s assessment suggests that substantially lower inference costs could make AI more practical for a wider range of everyday business applications. It is his view of the scale of cost reduction needed, rather than a guarantee that prices will fall by that amount.

Reducing inference costs could help businesses use AI more frequently in customer support, software development, data analysis and other workflows. However, the financial benefits depend on factors such as model choice, computing requirements, usage levels and the infrastructure used to run each application.

For cloud providers, lower costs per AI task could also encourage greater adoption. If businesses can run AI applications more economically, they may have more opportunities to integrate the technology into products and internal operations.

Amazon’s Trainium Chips Aim to Improve Price-Performance

One part of Amazon’s strategy is the development of its own AI chips, known as Trainium. These chips are intended to support AI computing workloads while helping customers improve performance relative to cost.

Jassy said Trainium delivers 30% to 40% better price-performance than comparable alternatives. The claim reflects Amazon’s stated assessment of its hardware, rather than an independently established advantage across every workload or configuration.

Price-performance is an important consideration for companies running AI systems at scale. A chip that delivers more useful computing work for a given cost can help reduce the expense of operating AI models, provided it meets the application’s technical requirements.

Developing its own chips also gives Amazon another way to shape the hardware and infrastructure used across its cloud platform. The strategy complements investment in data centres and other resources by targeting the efficiency of the computing systems themselves.

However, specialised chips do not eliminate every infrastructure constraint. AI services still depend on adequate memory, electricity, data-centre capacity and supporting equipment. Improvements in chip efficiency can help, but they must be considered alongside the availability of the wider infrastructure.

What AWS’s Capacity Constraints Mean for AI Businesses

The fact that AWS capacity is largely booked years ahead offers a signal of the scale of demand cloud providers are preparing to serve. For businesses building AI applications, access to computing resources is becoming an important part of long-term technology planning.

Companies may need to forecast their infrastructure requirements earlier, evaluate the costs of different AI workloads and consider how efficiently their applications use computing resources. These decisions can affect the pace at which new products are launched and existing services are expanded.

AI developers must also distinguish between experimenting with a model and operating it at scale. A system that works well in a limited test may require substantially more computing capacity when deployed to thousands or millions of users. Inference costs and hardware availability can therefore become major considerations during commercial expansion.

For AWS, the challenge is to expand capacity without neglecting its traditional cloud customers. Jassy’s comments make clear that AI demand is competing for resources that are also needed by the company’s broader cloud business.

Amazon’s planned infrastructure spending and Trainium development represent two approaches to addressing the challenge: increasing the amount of available computing infrastructure and improving the economics of the hardware used to run AI workloads. Neither approach alone guarantees that shortages will disappear, particularly if demand continues to grow rapidly.

Outlook: AI Infrastructure Demand Could Keep Supply Under Pressure

Jassy’s assessment points to a continuing mismatch between the demand for AI computing and the infrastructure currently available to supply it. With AWS’s major capacity blocks for 2026 sold out, much of 2027 booked and a significant portion of 2028 already committed, the pressure extends well beyond immediate demand.

The constraints involve several parts of the technology supply chain, including data-centre space, electricity and memory. These factors can limit how quickly cloud providers turn financial investment into usable computing capacity.

Amazon’s potential to spend more than its planned $220 billion in capital expenditure for 2026 highlights the scale of the investment required. At the same time, its focus on Trainium chips and lower inference costs reflects the importance of making each unit of computing capacity more economical.

The central question for the AI industry is not only how quickly infrastructure can be built, but also whether the supply of power, memory and specialised hardware can keep pace with growing demand. For now, Jassy’s remarks suggest that capacity planning and computing efficiency will remain central challenges for Amazon and the wider cloud industry.