May 20, 2026 · 1h 29m · latent-space
The Agent-Native Cloud: 3M Users, 100K Signups/Wk, Data Centers, & Death PRs — Jake Cooper, Railway
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Railway founder and CEO Jake Cooper joins Latent Space to discuss building an independent, agent-native cloud platform by replacing legacy DevOps abstractions with custom bare-metal data centers and Linux kernel patches. He details the economic and architectural realities of scaling infrastructure to support millions of users and high-velocity AI coding agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 17.3% of the talking time here. How this is scored →
The hosts as informed peer 4.2 Guest teaching 3.9 Guest disagreement 2.0 The hosts pushing back 1.9
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:07:50 Stop writing code by hand
Jake forcefully dismisses traditional developer workflows, stating unequivocally that anyone writing code by hand in the modern LLM era is doing their job fundamentally wrong.
Hardest push from the hosts ▶ 25:09 Swyx pinning Jake on orbital thermodynamicsSwyx directly challenges Jake's dismissive stance on space data centers, pinning him on the exact thermodynamic impossibility of dissipating high server wattage in a hard vacuum.
Biggest teaching moment ▶ 17:02 The economics of bare metal vs cloud marginJake walks through the exact operational arithmetic of running bare metal hardware, revealing an astonishing 3-month capital payback period that renders public cloud markups economically unjustifiable.
The host holds their own ▶ 1:01:20 Swyx drilling into workflow engine determinismSwyx draws upon his deep domain expertise from Temporal to interrogate how state machine replays fail under real-world scale, demonstrating peer-level technical command.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Host Message: Support Latent Space by Subscribing | 1 | 1 | 0 | 0 | The opening includes Swyx's solo subscriber appeal followed by lighthearted studio banter about business cards and internal titles like 'Trainiacs'. Cooper introduces Railway's high-level premise as a collaborative platform without technical friction. | |
| Jake Cooper's Technical Philosophy and Custom Kernel Work | 4 | 4 | 1 | 2 | Swyx probes Jake's background across Bloomberg, Uber, and Cadence, noting it does not fit the typical PaaS founder profile. Jake explains how his obsession with user experience drove him from front-end work all the way to patching the Linux kernel for storage primitives. | |
| Reimagining Git, Open Source Distribution, and Rollouts | 3 | 4 | 2 | 1 | Jake critiques GitHub's core model as a series of broken pointers and proposes percentage-based non-deterministic rollouts. Swyx pulls up Railway's growth chart, prompting Jake to candidly explain the brutal financial reality of bleeding $500k/month on free tier abusers. | |
| Agentic Workloads and Deploying Bare Metal Data Centers | 3 | 5 | 1 | 1 | Jake outlines the technical reasons Railway avoids Kubernetes in favor of custom orchestration and bare metal data centers. He details the economic payoff of racking custom servers, noting a 3-month payback period compared to renting public cloud compute. | |
| Multi-Cloud Overlays, Hardware Financing, and Venture Debt | 4 | 3 | 2 | 2 | Alessio and Swyx explore multi-cloud network overlays and alternative financing models. Jake reframes venture debt as hardware asset-backed debt rather than dilutive venture equity, detailing how each VC round bought specific partner capabilities. | |
| Space Data Center Debates and Agent Infrastructure Bottlenecks | 5 | 4 | 4 | 4 | Swyx challenges Jake on his viral skepticism of orbital data centers, pointing out the thermodynamic laws of heat dissipation in a vacuum. Jake counters by separating theoretical engineering optimism from the immediate physical bottlenecks facing parallel agent compute. | |
| CLI Handholds for Agents and Canvas Context Anchoring | 4 | 4 | 2 | 2 | Swyx offers a counter-thesis that Railway's flagship visual canvas is becoming obsolete compared to deep CLI interfaces for coding agents. Jake partially agrees, explaining that while the canvas shifts from input to observational output, it remains essential as an organizational context anchor. | |
| Central Station, Public Transparency, and Incident Management | 4 | 3 | 1 | 1 | Jake presents Central Station, an internal system that aggregates telemetry and user issues into semantic incident clusters. He discusses balancing transparent public postmortems with responsible disclosure during upstream vendor failures. | |
| AI Site Reliability Engineering and Self-Replicating Infrastructure | 5 | 5 | 3 | 2 | Swyx brings up insider intel from Erica Brescia that Jake was originally an AI SRE skeptic. Jake confirms he still rejects unconstrained agents on production databases, arguing safe agentic SRE requires copy-on-write sandboxes and strict alignment between specs, code, and tests. | |
| Fluid Compute, Heroku's Legacy, and Temporal Workflows | 7 | 5 | 2 | 3 | Swyx leverages his insider background at Temporal to interrogate Jake on the friction points of long-running workflows and non-determinism bugs. Jake provides a deep practitioner breakdown of state machine complexity from his Uber Cadence days, prompting a detailed discussion on alternative workflow engines. | |
| Railpack Dependency Resolution and Enterprise AI Agent Spending | 4 | 4 | 3 | 2 | Jake explains why Railway moved away from Nixpacks to custom content-addressable dependency engines. He defends his mandate that developers must stop writing manual code, justifying spending $300k monthly on coding tokens as an accelerated roadmap play. | |
| The Death of Pull Requests and Snapshotted Environments | 6 | 4 | 3 | 2 | Swyx and Jake discuss the obsolescence of traditional PRs in favor of feature flags and shadow deployments. Jake rejects the classic 'cattle vs pets' cloud dichotomy, arguing that with instant snapshotting and copy-on-write file systems, pet environments can be cloned safely. | |
| Solo Founder Realities and the Independent Cloud Vision | 4 | 4 | 2 | 2 | The conversation covers Jake's unorthodox path as a solo founder avoiding traditional YC dogma, his intense weekly focus cycle, and his refusal to resell rented GPUs. Jake explains the long-term vision of an independent, vertically integrated cloud built without copying hyperscaler architectures. |
Statements from this episode (25)
Cooper: Current DevOps tooling stacks entropy on top of entropy
“Most of the tooling right now is kind of like stacked up, like you're stacking entropy on top of entropy on top of entropy, right? So you have like Docker and cube and then like Ansible scripts and all of these other things, right?”
Cooper: Railway is writing custom Linux kernel patches for agent infrastructure
“Like, you know, I was, Adding patches to the kernel this week. Right. Just to like get the experience there.”
Cooper: GitHub's original sin is that cloning creates broken pointers
“I think the, it's funny, because, like, I think GitHub's original sin is that it's like almost a series of broken pointers. It's like you have essentially this thing, and then you clone it, and then, okay, great, like, I've just lost That whole upstream, right…”
Cooper: Railway burned roughly $500k monthly during free tier era
“So like free tier era, losing, I think half a million dollars a month.”
Cooper: Natural language agents will dominate software development within ten years
“That will be kind of the dominant species over the next like 10 years is, is we've moved from assembly to C to C++ to JavaScript to now like words.”
Cooper: Three-month payback buying bare-metal versus renting cloud
“Our payback period when we go to metal if we rent it in the cloud, our payback period is about three months.”
Cooper: Railway's cash and server value exceeds total venture capital raised
“Even now the amount of money that we've raised is less than the amount of money that we have in the bank, plus how, what the value of the servers are, because the servers have actually appreciated in value, because RAM has gone up in, in general, right?”
Cooper: Rebuilt Railway's network overlay in one weekend for multi-cloud
“I spent a weekend rebuilding our entire, like, Network, like overlay essentially, so that we could straddle five different clouds.”
Cooper: Railway achieves roughly 70% margins on bare-metal infrastructure
“Because we've built out our metal data centers, like our margins on metal are like quite high for the like 70%.”
Cooper: A new version control system outside Git will ultimately emerge
“And I do think that something ultimately outside Git will emerge in terms of how we're going to go into version a lot of these things over time.”
Cooper: Agent infra requires 1000x scale rather than new primitives
“And so like the branching and forking stuff, like it's not different. Like we're just moving a thousand times quicker than we used to. And so some of these things like look like you really need like something massively, massively different, but it's just, you …”
Cooper: AI coding agents prefer complex CLIs with hundreds of flags
“Things that were prohibitively annoying to humans are not actually prohibitively annoying to agents. They're really, really nice, right? And so, if I wanted to hand you a CLI and I said, hey, guess what? The CLI has 40 arguments and. Right. 600 flags. You'd be…”
Cooper: The traditional push-pull-rebuild deployment loop will disappear
“Like push, pull, whatever kind of like loop, I just fundamentally believe it's going to go away, right? Like it's, we're going to get to a point where you make a small change in production that changes version across your entire kind of infrastructure. You're …”
Cooper: Infrastructure canvases will become output review tools instead of inputs
“We have previously used it a lot as an input and its goal moving forward is actually a lot more like an output. What I mean by that is you would go to the canvas and you'd make some changes and all these other things, whatever. Right. And you see them and, you…”
Cooper: Slack and Discord are awful tools for team communication
“I think Slack, I think, and incidentally, I think Discord's awful too.”
Cooper: Railway has scaled to 3 million users
“When you go and kind of roll out to those, like that, like large user base of like three million people, right. You know, like you have a lot of different disparate behaviors that, that can kind of come up.”
Cooper: AI SREs without safe primitives will destroy production databases
“If you just unleash an AI SRE on your production infrastructure and you don't have like safe primitives for like copying volumes, making sure that this is fine, it's gonna nuke your production database. Like it's not a matter of if, it's a matter of when it's …”
Cooper: Agentic workloads will destroy traditional static AWS instances
“I think the era of, like, I have an AWS instance, and I'm going to, you know, get four vCPU and 16 gigs of RAM, It's going to get, like, completely destroyed, right? Because it's like, if you do that for agents or anything else like that, you now need a thousa…”
Cooper: Temporal requires holding full workflow states mentally, risking replay bugs
“Temporal was always like really, really, really great in theory. And it was really, really great when you got it working the way that you wanted to in production. It's just, it required you to like model that entire journey in your head. And if you didn't have…”
Cooper: Developers writing code by hand instead of reviewing AI are wrong
“If you're writing code by hand, you are doing this wrong, right? Like, they're code, the tools are good enough at this point, like, that you move extremely, extremely quickly, and like, yes, there are issues and pain points and all these other things but, You …”
Cooper: Companies will reserve massive token budgets for elite AI developers
“I think what you're going to end up with is actually, you're going to almost, like, end up with the, like, F-one driver concept. Which is if you have a, if you have somebody who's like really, really adept at these things it makes sense to go and put them into…”
Swix: Pull requests and code reviews are dying, replaced by prompt requests
“The pull request is dying. Yeah. Right? It's going to be the prompt request. And then beyond that, code review is also kind of dying”
Cooper: Developers can treat servers as pets with instantaneous cloning
“I think you can move towards having pets so long as you have a, and this is gonna be a jump, so long as you have a cloning machine for your pets. If you can snapshot every single thing at every frame, then, like, It actually doesn't matter if, you know, that t…”
Cooper: Two is the worst number of co-founders due to deadlocks
“And this is why I was saying like two is the worst number of co-founders is because you have no tie break, right? You basically are like, well, I disagree on this thing and I disagree on this thing, right? I was like, well, how do you resolve that, right?”