AWS CEO Matt Garman: the cloud is being rebuilt for AI agents
Key points
Key takeaways from the Source podcast interview with AWS CEO Amazon Web Services and Matt Garman (October 2026):
Still early for cloud. Garman says AWS revenue is around $169–170bn, growing 37% a year, yet most workloads still run on-premises, so the tailwind from both AI and cloud migration has room to run.
Startups are the engine. He estimates 30–40% of AWS revenue today comes from companies that were once AWS startups, and argues AWS deliberately keeps capacity available for two-people-in-a-garage founders because they push the platform harder than banks or governments.
A $220bn AI capex bet. AWS recently announced plans to purchase roughly two million NVIDIA GPUs over the next couple of years and a capital spend of about $220bn for 2026, with no plan to slow the buildout as demand keeps rising.
Cloud built for agents. AWS is optimising services so agents work as well as people — such as starting a database in three seconds — and is previewing AWS Context, a context layer that lets agents find data across S3, Aurora and other stores.
Frictionless agent on-ramp. New AWS accounts can now sign up with a Gmail address and no credit card and be operating in under 30 seconds, with VPC and IAM defaults handled behind the scenes and no migration needed when the workload later scales.
New primitives for agent workloads. Garman cites compute sandboxes, gateways and fine-grained, time-boxed agent permissions versus people permissions as genuinely new building blocks, and notes Firecracker microVMs — invented at AWS a decade ago — have proven ideal for fast-spinning, securely isolated sandboxes.
Agentic workflows perform better on AWS. He argues agent workflows run better on AWS than elsewhere because of its latency, throughput and tail-latency behaviour, which agents are far more sensitive to than human users.
Read more: a16z on YouTube
