Building the Cloud for AI Agents | AWS CEO Matt Garman

Video summary

AWS CEO Matt Garman details $220 billion in 2026 capex, 2 million planned NVIDIA GPUs, and cloud tools built for AI agents.

AWS CEO Matt Garman says AWS generated about $169–170 billion in revenue, growing 37%, and argues startups remain vital: an estimated 30–40% of AWS revenue comes from companies that began as startups. To meet AI demand, AWS plans $220 billion in 2026 capital spending and says it will buy 2 million NVIDIA GPUs over the next few years, while deliberately reserving capacity for startups; about 60% of requests are eventually fulfilled, sometimes later or in another region. Garman says agent-ready cloud design means fast, predictable services, simpler account setup, temporary permissions and secure sandboxes. He describes Graviton as roughly 20% cheaper with 20% better performance, and says Trainium powers most Bedrock traffic. Enterprise agents are still mostly supervised; safe autonomy depends on trust, permissions, evaluations and guardrails. Garman also points to Continuum for AI-assisted security and Amazon Quick agents used across HR, finance and software development, where agent-led teams are accelerating product releases.

Chapters

1. 0:00169B, growing 37: AWS reaches $169–170 billion in revenue at 37% growth
2. 2:51Why startups are the lifeblood: AWS estimates 30–40% of revenue comes from startup-grown companies
3. 5:23Why startups are the lifeblood: AWS context, S3 tail latency, and sub-30-second account setup for agents
4. 12:26Rethinking cloud for agents: Firecracker sandboxes and short-lived agent permissions
5. 16:24The GPU allocation problem: $220 billion in 2026 capex and capacity for startups
6. 21:27Is the AI CapEx a bubble? AWS cites diversified customers and positive enterprise ROI
7. 24:54Is the AI CapEx a bubble? AWS plans power and supply chains years ahead
8. 30:22Clearing up data center myths: AWS cites $5,000 lower annual taxes in one county
9. 33:07The Graviton and Trainium bet: 20% savings and sold-out Trainium 3 capacity
10. 40:06Where enterprises are stuck on agents: safety, evaluations, and 45-day enablement
11. 45:43AI risk, Hugging Face, and Continuum: Bedrock privacy and SageMaker fine-tuning
12. 49:38AI Risk, Hugging Face, and Continuum: AWS’s Continuum Targets Machine-Speed Security
13. 52:41AI Risk, Hugging Face, and Continuum: Quick, Agent-First Development, and Smaller AWS Pods