Depot is making builds fast enough for the AI era

Kyle Galbraith, CEO & founder of Depot, has been coming to re:Invent for six years. He's watched it grow from 25,000 people in one hotel to 80,000 spread across the entire Las Vegas Strip. This year, he's watching something else grow: the strain AI is putting on the software delivery pipeline.

Depot started by accelerating Docker image builds in CI environments like GitHub Actions and CircleCI. Kyle and his co-founders were platform engineers frustrated with slow builds, so they built the solution they wanted. After Y Combinator in 2023, they expanded into general build acceleration—GitHub Action runners, remote caching, agent sandboxes, and their own container registry.

The bottleneck has moved

It's almost trite to say AI is producing exponentially more code. But the downstream effects are real. More code means more builds, more tests, more deployments. The delivery pipeline—historically thought of as that weird middle ground between development and production—is becoming the bottleneck.

Kyle's observation: we need to stop treating CI/CD as something code just passes through. It's a production system that needs to be understood and optimized like any other production system. When the October AWS outage hit, companies discovered their build systems had hidden dependencies on third parties that failed silently. Without visibility, they couldn't fix it or ship updates.

Correctness first, then speed

Depot's philosophy is that correctness comes before performance. If you make builds faster but corrupt the output, you've accomplished nothing. Every optimization they make is transparent and documented, because customers need to trust the build system completely.

The new frontier is egress controls on GitHub Action runners. You can now monitor all outbound network traffic from CI jobs and put ACLs in place—if a job tries to phone out to a domain not on the allow list, the build fails. This prevents data exfiltration during the build phase, a vector most companies don't think about until it's too late.

The API-first future

Everyone's asking for APIs for everything now. Kyle notes that MCP had a moment, but people are realizing they don't want to stuff their entire context window with tools. They'd rather have a clean API and write a quick script—or have an agent write the script—that works reliably every time.

This maps to a broader point about developer tools: the best ones are flexible. They expose building blocks so customers don't outgrow them. Developer, human, machine, agent—whatever's calling the API, the building blocks should be composable.

Kyle's favorite part of re:Invent is still the infrastructure talks, the deep glimpses into how AWS thinks about redundancy and reliability at scale. His second favorite? The happy hours sponsored by three or four different companies, where you can learn about different problems and domains without sitting through a sales pitch.

This interview was conducted at AWS re:Invent 2025.