a16z Podcast: AWS CEO Matt Garman on the Cloud Built for AI Agents

SPEAKER_1 (0:01)

So Matt Garman just made a really bold claim in his conversation with a16z. He says AWS is being rebuilt for a completely new era, and here's the kicker: "Agentic workflows tend to perform better on AWS than anywhere else." That's a huge statement about where cloud infrastructure is heading. It really is. And the scale behind this is massive. AWS launched back in 2006, and now they're pulling in roughly $170 billion in revenue while still growing at 37 percent. But what struck me is Garman saying they're only in the early stages because so much compute still lives on premises. Right. And a major part of their strategy comes down to startups. Garman recalls that early analysis showed startups were the most interesting segment, and that focus hasn't changed. They believe today's startups are tomorrow's enterprises, estimating that 30 to 40 percent of their revenue comes from companies that began on their platform. That startup focus actually shapes their hardware strategy in interesting ways. Garman says AWS could sell every accelerator to frontier labs, but they intentionally hold capacity for emerging companies. They're buying 2 million NVIDIA GPUs over the next couple years and spending about $220 billion in 2026

Which brings us to this agent-first cloud they're building. AWS is creating compute sandboxes, gateways, permissions, and faster setup paths so agents can move through systems safely. They're rolling out tools like Bedrock, Agent Core, and AWS Context to make the cloud traversable for AI agents, not just humans. And agents change the economics entirely. Garman points out that some workloads need five nines of durability, but transient agent tasks may not. So they're designing lighter, disposable environments like Firecracker microVM sandboxes that can still scale into full production systems later.

Security is absolutely central to this vision. Enterprises worry about mistakes, data access, and autonomous actions. That's why AWS is investing in permissions, human-in-the-loop controls, and evaluation systems. Garman emphasizes that enterprise data is a company's most valuable asset, which is why Bedrock keeps customer data inside the VPC and doesn't use it to train outside models. What's fascinating is the infrastructure challenge behind all this.

Garman says they're not just solving for GPUs, but also power, memory, networking, TSMC capacity, and components deep in the supply chain. He adds there's always a latest constraint, one bottleneck disappears and another appears.

That's driving their push into custom silicon. Nitro, Graviton, Trainium, and Inferentia are all designed for performance, cost, and security.

Garman says Trainium is already a strong inference chip and powers much of Bedrock.

And here's the meta point: AI is changing AWS itself. Internally, they're using Amazon Q across employees and experimenting with smaller, faster product teams. The message is clear: AWS wants to be where agents, startups, and enterprises build the next generation of software, with everything ready to support it.