Matt Garman says data center operators who boast about ignoring regulations have given the industry a bad name.
“All data center producers are not created equal,” the AWS CEO tells me.
I flew up to Seattle to speak with Garman on Friday, just days after the midterm elections. I was particularly interested in hearing from him now: candidates had campaigned against data center expansion, and voters had overwhelmingly backed restrictions in local referendums. I wanted to understand what he thought the results meant for the AI compute buildout.
Garman argues that Amazon’s approach to power, water, and community engagement should distinguish it from operators who disregard local concerns. But when I press him on what the industry needs to do differently, he says AWS is working with governors to turn some of those practices into “more structured programs” everyone has to follow. He’s convinced there’s a way to keep building while giving communities a better deal.
That political challenge sits alongside an enormous business ambition. Amazon thinks AWS could become a trillion-dollar revenue business within the next five to ten years. Garman says AI is accelerating the move into the cloud, as companies that put off migrating their core systems now want agents to work across their data.
He also thinks inference, the computing required to run AI models, is less than 1 percent of what it will eventually become. AWS is adding one GPU every 14 seconds, he says, yet he expects capacity shortages for another couple of years. We discuss how he decides who gets that compute, why he reserves capacity for startups that haven’t been founded yet, and why he sees AWS’s risk differently from providers dependent on debt and a handful of large customers.
I ask about the reset inside Amazon’s Nova team, its ambitions to build frontier models, and the decision to sell Trainium chips outside AWS. We also get into the growing market for inference providers, and a research bet Garman thinks could make companies more comfortable letting agents work on their own.
And after my recent conversation with Sundar Pichai about putting data centers in space, I wanted Garman’s take. He thinks the economics and practical challenges will keep that prospect out of reach for another decade or two. Of the alternatives we discuss, he’s much more interested in deserts.
Sources is available wherever you get podcasts, including YouTube, Spotify, and Apple Podcasts.
Thanks to the show’s premiere sponsors: Mercury, Granola, and Atlassian.
A MESSAGE FROM MY SPONSOR
Still the OG. Now With AI Answers.
The original guide to the internet just got smarter. Yahoo Scout brings AI answers into Yahoo Search, with links to the sources behind every answer so you can check the details or read the original. Search the way you always have, or just ask.
Try Yahoo Scout at scout.com.
Episode highlights
AI is pushing companies into the cloud, and that’s central to Amazon’s trillion-dollar ambition. Garman says customers who put off moving their databases and other core systems are accelerating those projects because they want agents to work across their businesses. “It has to have all of the data sources accessible,” he says of an agent. A company may be comfortable leaving a database in its own data center until that becomes an obstacle to what it wants its agents to do. I hadn’t thought about AI accelerating the older cloud business quite that way. That’s a big part of his case for AWS reaching $1 trillion in annual revenue: AI creates new demand for computing while pulling existing workloads into the cloud.
Amazon’s AI model reset hasn’t changed its ambition to compete at the frontier. I ask whether the changes inside the Nova team mean Amazon is backing away from building its own frontier models. “We’re just as invested in that strategy as we’ve ever been,” Garman tells me. He describes a reset around talent, data, and infrastructure, informed by lessons from the team’s earlier approach. He still thinks customers want a frontier model built by AWS itself, alongside the outside models the company offers. On what comes next, his message is “stay tuned.”
He thinks inference is less than 1 percent of what it will become. Running models already accounts for the majority of AWS’s AI workloads, Garman says. But when I ask how far along that market is, he says, “I don’t think it’s even 1%.” He says each 50 percent drop in inference prices has brought an order of magnitude more usage, with revenue continuing to rise. He also sees providers like Fireworks, Baseten, and Modal as partners that help customers use open-weight models, even when they compete with AWS’s own services. His aim is to make AWS useful beneath whichever models and services customers choose.
AWS is betting that mathematical proofs can make AI agents reliable enough to do more work on their own. Garman points to automated reasoning, which he describes as probably Amazon’s biggest internal research investment. He wants to combine it with AI models to create “provably correct agents” that can work flexibly while obeying specific rules. His example is an agent processing an insurance claim: it could investigate the relevant information while being prevented from violating defined policies. He thinks that could let companies automate work they still require a person to oversee, and expects some deployments next year. The commercial connection is direct: “If you do unlock that, the usage suddenly goes through the roof even more.”
ICYMI
Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.
