Building the Cloud for AI Agents | AWS CEO Matt Garman

Summary · 2 min read

Matt Garman is AWS CEO and was its first GM for EC2, a role the host dates to 2006. Garman says he first worked with AWS as a business-school intern in 2005, when it was an internal project.

Garman argues that AWS must combine an easier cloud on-ramp with infrastructure designed for agents, including short-lived resources, fine-grained permissions and sandboxes. He says AWS is reserving scarce accelerator capacity for startups as well as serving frontier labs and enterprises, and does not expect investment to slow given demand and customer returns. He describes infrastructure constraints as shifting across power, chips, memory and construction, and explains how Graviton and Trainium fit into AWS’s chip strategy. For enterprises, he says wider agent adoption depends on rethinking workflows, improving evaluation and security practices, and protecting customer data; AWS is also using agents internally to speed product work.

Minute by minute

1. 12:49

   Agents need new infrastructure primitives. Garman says agent workloads may create a database briefly, so AWS is working to make resources quick to create and discard without losing the option of production-grade durability. He cites agent-specific permissions, gateways and compute sandboxes as new building blocks. He says Firecracker microVMs can provide a rapid, secure boundary for those sandboxes.

2. 16:22

   GPU allocation spans the ecosystem. Garman says AWS plans $220 billion in CapEx for 2026 and does not expect to slow, citing massive demand. AWS recently announced plans to buy 2 million NVIDIA GPUs over the next couple of years. He says AWS serves frontier labs and large enterprises while intentionally reserving capacity for startups; it eventually says yes to about 60% of requests, sometimes later or in a different region or configuration.

3. 32:54

   Custom chips target cost and scale. Garman traces AWS chips from network and storage offload work with the Annapurna team to Graviton and Trainium. He says Graviton is 20% cheaper with 20% better performance, and more than 90% of AWS’s top 100 customers use it in some way. He says Trainium 3 capacity is sold out toward the end of next year, and that most Bedrock traffic runs on Trainium.

4. 47:59

   Agents accelerate internal product work. Garman says Amazon rolled out Quick to employees, and teams in HR and finance are using agents to handle work that once took groups of people weeks. He says software development and product development have seen the biggest gains, with agents writing code and engineers managing agent teams. AWS is testing smaller, more flexible teams, including three or four people instead of ten.

Published Oct 8, 2026.