Amazon CEO Andy Jassy: AWS AI Capacity Sold Out Through 2027

Long-term dedicated AI compute capacity is booked solid through 2027, but on-demand buffers remain available as power grids and cooling set the physical ceiling on cloud expansion.

Appearing on the venture capital firm Andreessen Horowitz’s video podcast “The a16z Show” on Oct. 23, 2026, Amazon CEO Andy Jassy detailed the state of AWS cloud computing and artificial intelligence infrastructure. Jassy stated that AWS compute capacity is fully sold out for 2026, largely sold out for 2027, and that a significant portion of 2028 capacity is already spoken for in advance reservations.

Pushing back against Wall Street concerns over AI overspending by hyperscalers (the large cloud companies), Jassy argued that actual customer demand continues to outstrip the pace of physical hardware deployment, leaving significant AI demand unserved.

Large-Scale Compute Capacity Booked Through 2027

The sellout Jassy described centers on massive dedicated compute clusters reserved through multi-year contracts by leading foundation model developers and major enterprises. In the interview, Jassy explained, “We always leave a certain amount for on demand, but ’26 is obviously sold out. ’27 is largely sold out. A good chunk of ’28 is sold out.”

Two primary forces drive this surge: frontier labs training next-generation foundation models on massive GPU clusters, and the rapid deployment of AI agents (programs that carry out multi-step tasks for you on their own). Crucially, resource consumption has shifted beyond training into inference (using a finished model to produce answers), creating round-the-clock compute loads that exceed earlier projections.

As a result, leading AI companies are competing to lock down accelerator capacity two to three years in advance. Across the cloud industry, spinning up large clusters of thousands of high-end GPUs on short notice has become virtually impossible.

The On-Demand Buffer and the Developer Reality

The sellout does not mean day-to-day cloud operations are frozen for developers. AWS continues to hold back dedicated buffer capacity for standard workloads, on-demand EC2 instances, and managed API calls through services like Amazon Bedrock.

However, locking up massive blocks of capacity for years creates real ripple effects across the ecosystem:

Limited access to next-gen accelerators: Securing the latest Nvidia GPUs or high-end accelerator instances on an on-demand or spot basis is increasingly difficult.
Higher spot instance interruption rates: Spare compute capacity typically offered at spot discounts is redirected toward long-term contracted demand, leading to frequent spot terminations during batch machine learning runs.
Regional capacity contention: Workloads are clustering into specific regions where power and space are secured, raising the likelihood of InsufficientInstanceCapacity errors during sudden scale-outs.

Teams building custom models or self-hosting inference clusters must plan multi-region fallback architectures or rely on managed serverless endpoints rather than assuming unreserved instances will always be on tap.

$220 Billion in CapEx Hitting Utility Bottlenecks

Amazon expects capital expenditures (CapEx) to reach roughly $220 billion for 2026 to support its data center buildout. Yet despite this historic capital deployment, physical infrastructure limits—not funding—are capping the speed of expansion.

The steepest bottlenecks are electrical grid capacity and cooling infrastructure. Powering a gigawatt-scale AI campus requires the energy output of a mid-sized city, and obtaining utility interconnect approvals and building substations often takes several years. Dissipating the extreme thermal loads of high-density server racks through advanced liquid cooling further complicates facility construction.

Hardware packaging also remains constrained. Beyond Nvidia silicon, global production of HBM (memory stacked next to an AI chip so data moves much faster) and high-speed network switches faces packaging backlogs. Jassy noted that Amazon is not withholding investment, but rather encountering the physical supply limits of the globe: “We are leaving a meaningful amount of AI demand on the table.”

Custom Silicon and the Path to a $1 Trillion Run-Rate

To bypass external supply bottlenecks and rein in serving costs, Amazon is accelerating deployment of its in-house custom chips. Jassy emphasized that AWS is scaling up its Trainium and Graviton processors to drive down the cost per query for enterprise workloads.

For generative AI to move from experimental pilot projects into core production systems, the cost of inference must drop dramatically. Custom application-specific chips allow cloud providers and customers to maintain viable operating margins without paying premium prices for generic GPU compute.

Jassy pointed out that only 10% to 15% of global IT spending has moved to the cloud, predicting that AI will drastically accelerate the remaining migration. If this infrastructure wave continues, Jassy projected that AWS could eventually scale into a $1 trillion annual run-rate (the most recent month or two of revenue scaled to a year) business.

What to Watch

In the near term, developers should monitor the price-performance metrics and fleet availability of AWS Trainium2 instances scheduled for broader rollout in late 2026. Over the longer horizon, the regulatory approval and deployment timelines for Amazon’s nuclear energy partnerships and small modular reactor (SMR) initiatives will determine whether the cloud giant can secure enough power to keep expanding its AI capacity.