The AI agent economy has a new infrastructure problem: where do you park a fleet of autonomous workers that need to run code, train models, and spin up other agents for days or weeks at a time?
Barnaby Malet thinks he has an answer. His Y Combinator-backed startup machine0 sells persistent cloud virtual machines purpose-built for AI agents, starting at just over a penny per hour. Unlike the ephemeral sandboxes or serverless functions that dominate the market, these are full-fledged VMs with root access, static IPs, and the kind of persistence that matters when your software workforce runs around the clock.
"A persistent cloud VM, billed by the minute," Malet wrote when he introduced machine0 on Hacker News in August. "Agents self drive via CLI or MCP." The pitch landed with developers nursing similar frustrations: 83 upvotes and a thread full of founders describing agent deployments that outgrew traditional compute options.
The service, which evolved significantly through ongoing releases from March into the summer, provisions VMs that scale from a single vCPU up to 60-core machines with 240 GB of RAM and NVMe storage. Each VM gets a static IP and an HTTPS endpoint at <vm>.mac0.io. Customers access them through a command-line tool or via Model Context Protocol integration, a choice that reflects machine0's design philosophy: every CLI command outputs JSON for agent parsing, and the platform injects credentials and environment variables through what the company calls Profiles.
GPU configurations tell the real story about machine0's target market. Single RTX 4000 Ada cards rent for $0.836 per hour. Eight-GPU H100 clusters cost $35.28 per hour; H200 equivalents run $35.76. AMD's MI300X eight-GPU nodes go for $22.792 hourly. Those aren't impulse purchases for hobbyists tinkering with ChatGPT wrappers.
One customer, Malet said, runs hundreds of machines simultaneously for agent fleets where a pilot agent parcels out work to subordinate agents. Another keeps a suspended H100 on standby and fires up overnight optimization loops. A third issues fresh 60-vCPU instances per user session from versioned images—thousands of machines, most alive for roughly two minutes before vanishing.
The suspend-and-resume feature matters here in ways that separate machine0 from competitors. The platform snapshots full disk state while tearing down compute, charging $0.078 per GB monthly for suspended storage. RAM and running processes don't survive. "Anything that survives a reboot survives a suspend," Malet explained, a constraint that forces developers to architect with state persistence in mind but cuts costs when agents idle.
machine0 runs on DigitalOcean infrastructure, a partnership Malet disclosed in the Hacker News thread. The arrangement shows in the pricing: machine0 raised GPU rates effective August 1 to track DigitalOcean's list increases, according to changelog entries. "Their machines are good & they have a partnership program that's fast and compatible with the model," Malet said of DigitalOcean, though the company has bring-your-own-cloud on its roadmap without a public timeline.
The technical choices reveal someone who understands the agent developer workflow. NixOS gets first-class support with reproducible flakes and GPU-ready images like nixos-25-11-nvidia. Ubuntu images arrive preloaded with Docker, Node, Python, Claude Code, and Codex. The MCP gateway at app.machine0.io/mcp aggregates tools across per-profile endpoints. These aren't features you'd prioritize for traditional web hosting.
Malet previously co-founded Upflow, a Y Combinator Winter 2020 company where he served as CTO and chief product officer after studying computer science at Imperial College London. machine0 operates as Fundamental Software with Paris headquarters, though the YC directory shows a single employee—a lean setup even by startup standards. Y Combinator's current deal provides $500,000 ($125,000 for 7 percent equity plus a $375,000 uncapped SAFE), though machine0 hasn't disclosed exact terms.
The competitive landscape includes E2B's code-execution sandboxes, Modal Labs' serverless GPU jobs, RunPod's per-second billing, and Vast.ai's GPU marketplace. Malet positioned machine0 against Modal directly: "Modal is an ephemeral sandbox, whereas machine0 is a persistent VM you own … root, your own driver/CUDA/kernel." That distinction matters to developers who need control over the full stack, not just a function runtime.
The company operates in five regions—New York, San Francisco, London, Amsterdam, and Singapore—and claims 99.99 percent VM-level uptime. Installation requires either curl or npm; account usage requires a $5 minimum top-up with unused credits refundable. Minute-based billing starts at $0.013 per hour for entry-level VMs, granular enough that the two-minute sessions one customer described don't rack up waste.
July and August changelog entries trace rapid iteration: MCP enhancements, capacity-shortage handling, SSH improvements. The pace suggests a founder chasing feedback from early customers rather than executing a preset roadmap. Whether that agility translates to staying power in an infrastructure market crowded with well-funded competitors remains an open question, but machine0 has carved out a niche that didn't quite exist six months ago—and the agent developers arriving in that Hacker News thread seem grateful someone finally did.
