Kernel Powers 1,000+ AI Agents on $444/Month of Railway Infrastructure

For a Y Combinator startup providing AI infrastructure to over 1,000 AI companies, every millisecond of latency and every deployment matters. Kernel spins up cloud browsers for customers building AI agents that need to access the internet, serving as critical infrastructure for the exploding AI agent ecosystem.

When co-founder and CTO Rafael Garcia started Kernel in January, he brought hard-won lessons from his previous company Clever, which sold for $500 million in 2021 after building a six-person infrastructure team to manage AWS.

"Clever started on Heroku in 2012, which was like the Railway of the time. We quickly outgrew it and switched to home rolling our own tooling around AWS. My big learning from that experience was that we should have tried to hold on to Heroku longer than we did."

Garcia watched his previous company dedicate six full-time engineers to building an internal platform-as-a-service, reinventing products that eventually existed in the market. This time, he was determined not to repeat that mistake.

"It ended up being a full-time team of six people running core infrastructure and doing a lot of stuff that towards the end felt like reinventing a lot of stuff that just existed as products. Railway is exactly the tool I wish I had in 2012."

The challenge for Kernel was particularly complex: maintain rock-solid websocket connections for Chrome DevTools protocol, handle explosive growth from zero to 1,000+ customers, and do it all without building another infrastructure team.

Garcia chose Railway from day one, using it for Kernel's API, dashboard, website, and the critical websocket infrastructure connecting users to cloud browsers.

"Railway is kind of the front door API. All requests flow through our API. The tricky part is the nature of automating a web browser requires this websocket protocol called Chrome Dev Tools protocol—it's extremely stateful."

The platform delivered immediate value on deployments, a problem Garcia knew was deceptively complex from experience.

"Zero downtime deployments—doing that correctly is not trivial. I think people underrate the difficulty of that problem. It's especially difficult if you're a company like us who has long-running websocket connections that can't get broken."

Railway's out-of-the-box observability eliminated setup overhead. When Kernel needed to scale, the process was simple: click to add replicas. The team runs a fleet of Temporal workers on Railway for workflow orchestration, scaling them with just a few clicks when queue times increase.

"We had an issue the other day where that basically boiled down to we needed more workers running. That was like a few clicks in Railway."

The platform's template system accelerated infrastructure provisioning. When Kernel needed Redis, they deployed it from a template in minutes. Staging environments allowed testing before production deployments, critical for maintaining websocket stability.

While Kernel runs their actual browser infrastructure on bare metal servers (requiring KVM access for VM standby modes), Railway handles everything customer-facing—the entire API layer through which 1,000+ companies access Kernel's browser fleet.

Railway enabled Kernel to achieve explosive growth while maintaining a lean six-engineer team, avoiding the infrastructure burden that consumed resources at Garcia's previous company.

• Zero to 1,000+ customers in seven months. Launching in May with no customers, Kernel quickly became the go-to browser infrastructure for AI companies, capturing significant market share among Y Combinator's 5,000 portfolio companies building AI agents.
• Infrastructure costs of just $444/month for all customer-facing systems—API, dashboard, websocket connections, Temporal workers, and website. This tiny infrastructure bill supports customers ranging from YC startups to Series B+ enterprises.

"I don't even know how much we spend. It's currently not registering on my radar of like, oh, that was a big bill. For what it's doing for us, it's a great deal."

• Zero dedicated infrastructure engineers required. Unlike Clever's six-person infrastructure team, Kernel's six engineers focus entirely on product development. The contrast validated Garcia's decision to choose Railway from the start.

The platform scaled seamlessly as Kernel expanded from YC companies to larger enterprises, handling the increased load without requiring architectural changes or additional operational overhead.

"I'm extremely happy with the value we're getting. I love that you guys are optimizing infrastructure at the API layer, doing the same for us that we do for Chromium. There's a lot of respect for what you're doing."

Looking forward, Kernel plans to explore Railway's upcoming VM runtime feature for additional use cases, while continuing to rely on the platform for all customer-facing infrastructure.

For a founder who learned the hidden costs of building internal platforms the hard way, Railway delivered exactly what was promised: enterprise-grade infrastructure without the enterprise-grade complexity or team.