The bottom line: Railway founder Jake Cooper explains how cloud infrastructure for autonomous agents is being redesigned: bare-metal data centers with 3-month payback periods, 35 employees serving 3 million users, and why pull requests may become obsolete.
Railway founder Jake Cooper discusses the future of cloud infrastructure for a world of autonomous agents. From bare-metal data centers to rebuilding a cloud system for AI-native workloads – Railway is redefining how software goes to production.
Railway was not planned as an AI infrastructure company. When Jake Cooper – formerly at Bloomberg and Uber – founded the platform in 2020, his focus was radically simple: activation energy for deployment should be nearly zero. Push code, get a URL, iterate. No Dockerfiles, no Kubernetes manifests, no Ansible scripts on top of Ansible scripts.
The early years were a grinding struggle. Railroad spent its first 18 months manually acquiring the first 100 users – Jake personally greeted every Discord signup on a second monitor. Today, Railway has raised 124 million dollars and is growing rapidly. A 35-person team serves three million users and registers approximately 100,000 new signups weekly.
The secret lies in Railway’s own bare-metal data centers. With an amortization period of only three months compared to cloud rental – at 70% margins – these investments generate aggressive cloud-bursting capacity. Remarkably: the hardware has increased in value as RAM prices have risen. The value of their infrastructure now exceeds the capital deployed.
In the podcast, Cooper and the hosts discuss the technical architecture: bare-metal data centers, cloud bursting, Railpack and Nixpacks, Temporal workflows, feature flags and progressive rollouts specifically for agent-safe production environments. Central to this is Railway’s Central Station – an internal system for clustering customer feedback and incidents.
Cooper’s core thesis: agents as the next software species require version control, observability, compute, storage and orchestration at 1000x scale. This makes the classical pull-request workflow obsolete. CLIs could become more important than graphical interfaces. And perhaps “cattle, not pets” was yesterday – when you can clone your pets.