The Agent-Native Cloud: 3M Users, 100K Signups/Wk, Data Centers, & Death PRs — Jake Cooper, Railway
Summary
Railway’s central bet is that agents will become the dominant software-building “species” over the next 10 years, making agent-native infrastructure a secular platform shift even if today’s boom hits an inference wall. Jake Cooper traces the abstraction ladder from assembly to C to C++ to JavaScript to “words”; the destination is thousands of agents working concurrently, where coordination, safe intervention, and compute efficiency become as important as code generation.
The company’s growth curve was earned through six years of expansion, painful compaction, and a deliberate repair of bad unit economics. Railway once burned roughly $500,000 a month against perhaps $50,000 of monthly revenue and a $20 million bank balance while free users attracted bots and crypto miners; it temporarily restricted free usage, rebuilt the business, and now operates with 35 people while adding about 100,000 users a week. The discussion cites both 2 million and 3 million user figures. Cooper rejects the idealized always-up-and-right chart: “You actually don’t want charts that look like that.”
Bare metal gives Railway both its margin engine and the economic headroom to support massively parallel agents. Cooper says hardware pays back in about three months versus renting equivalent cloud capacity, despite being depreciated over four years, while metal workloads carry margins around 70%; Railway reports that the vast majority of workloads now run in its own data centers and uses hyperscalers for bursts. The operating rule is categorical: “You never ever ever want to be waiting on compute anymore. You always want to be waiting on intelligence.”
Railway’s architectural moat is control over network, compute, storage, and orchestration at a workload scale where familiar abstractions may break. Agents need many of the same primitives as humans—version control, feature flags, logs, traces, files, snapshots—but “a thousand times quicker”; Cooper expects conventional CI/CD to “melt,” thinks something beyond Git may emerge, and wants systems whose components can be ripped out whenever a super-exponential workload exposes the next bottleneck.
The agent interface reverses conventional product design: complexity becomes useful input, while the graphical canvas becomes an approval and context surface. A human may hate a CLI with 40 arguments and 600 flags, but an agent sees “so many handles”; Railway measures where agents leave the happy path and adds handles to close the loop. Its canvas consequently shifts from input to output—a “port in the storm” where humans understand changes, preserve shared context, and approve or reject what agents did.
Autonomous remediation is unsafe without cheap, production-like forks, progressive delivery, and reversible state. Cooper remains skeptical of an AI SRE unleashed directly on production: without copy-on-write volumes, read-only production data, PII transformations, observability, and limited blast radii, “it’s not a matter of if, it’s a matter of when it’s going to nuke that database.” Railway’s answer is to let agents clone services and state, validate a hypothesis close to production, then merge or discard the result.
Railway’s roughly $300,000 monthly coding-agent spend is a bet that engineering output—not token thrift—is the relevant denominator. Cooper personally uses about $25,000 and told the 35-person company, “If you are writing code by hand, you are doing this wrong”; engineers should review and reconcile generated code while architectural judgment matters more than ever. His proposed ROI measure is the percentage of tokens that reach production, with the best operators treated like F1 drivers worth placing in a “$3 million car.”
Competitive focus is as important as technical ambition: Railway does not want to become merely “the new Heroku,” and it is declining GPUs for now. Cooper attributes Heroku’s stagnation to being peripheral to Salesforce’s core business, while Railway aims to own the full build-and-deploy loop without copying hyperscalers’ architecture. He is explicit that Railway will not offer GPUs now but “100% will” eventually because vertically integrated infrastructure ultimately needs FLOPs—an unusually clear statement of sequencing rather than scope denial.
Deep dive
Not yet available upstream; scheduled sync will retry.