Six hours. That's how long some coding agents now run, straight through, rewriting entire repositories. Financial research bots pull earnings data across dozens of companies for days at a time. Customer service agents never clock out. And somewhere in the gap between what developers expected and what actually happens in production, a new infrastructure market is being born.
The original assumption—spin up an agent, complete a task, shut it down—turned out to be wrong. Or at least incomplete. "Agent workloads are shifting from ephemeral to always-on," Barnaby Malet wrote earlier this year. Malet, a solo founder whose startup machine0 went through Y Combinator in 2026, is building infrastructure for exactly this scenario: agents that need to stay alive.
He's not alone. And the money following the problem is hard to ignore.
Together AI closed an $800 million Series C at an $8.3 billion valuation. Lambda secured a $1 billion credit facility to meet what it called "gigawatt-scale AI infrastructure demand." CoreWeave inked a deal with Meta in April worth $21 billion running through 2032. Gartner's forecast put AI spending on track to hit $2.59 trillion in 2026, up 47% year-over-year, with infrastructure gobbling up more than 45% of that. A chunk of that capital is chasing a question that didn't really exist two years ago: where do you run an agent that never stops?
When the Thesis Crystallized
Something shifted in the developer conversation this year, and it happened fast. In July, a16z published an investment memo with the headline "Agents just want a computer," arguing for fully stateful, OS-level environments. Days later, LangChain echoed the sentiment with "Agents need their own computer," emphasizing fast-boot isolated compute. Google Cloud released production guides that treated long-running orchestration patterns as table stakes, not edge cases.
These weren't abstract think pieces. Anthropic's Computer Use feature—launched as a research preview—lets Claude control desktop environments directly. The Model Context Protocol, or MCP, is standardizing how agents discover and invoke tools across machines, with a roadmap published mid-year. LangChain detailed checkpointing and recovery for agents that can run 200-turn coding sessions without losing state.
What changed isn't mysterious. The agents got more capable. The tasks got more complex. A coding agent rewriting a repository isn't done in three minutes. A research bot pulling quarterly earnings across fifty companies doesn't wrap up before lunch. These workloads demand persistence, isolation, state management—in other words, they need what we used to just call "computers."
The Race to Build the Layer
Malet's pitch for machine0 is almost aggressively straightforward: one CLI command gives your agent its own cloud computer. Persistent NixOS or Ubuntu VMs. Minute-based billing starting at $0.013 per hour for a single vCPU and 1GB of RAM. GPU options scaling up to eight H100s at $26.312 per hour. Five regions spanning US East and West, the UK, EU, and Asia. Agents interact via CLI or MCP endpoints, with every command supporting JSON output for programmatic control.
Malet, who previously worked on Upflow (another YC company), initially started with an MCP gateway. Then he pivoted. The real bottleneck, he realized, wasn't protocol standardization. It was provisioning. The product changelog shows the pace: profiles and variables went generally available, OAuth rolled out for MCP clients, NVMe storage tiers were introduced, GPU snapshots improved. It's moving fast because the demand is there—or at least, enough people think it will be.
machine0 has company. Orgo markets itself as offering "instant cloud computers your AI agents can see, control, and operate," with REST APIs, Python SDKs, WebSockets, and an OpenAI-compatible chat endpoint. Computer Agents promises each agent its own persistent cloud computer for browsing, code execution, package installation. Celesto released an open-source framework for lightweight sandboxes aimed at agents. Matrix OS offers always-on cloud computers with repo access and live previews. Heyo Computer focuses on micro-VMs using KVM and Firecracker, with snapshot and fork capabilities baked in.
Then there are the established players adjusting course. E2B, which raised a Series A last year, positions its sandboxes for usage-based billing but emphasizes that you need to actively kill sandboxes to stop charges—a model built for ephemeral tasks, not persistent agents. Vercel published guidance comparing its sandbox offering to E2B's, highlighting concurrency limits and policy differences. AWS documented patterns for running agent sandboxes on EKS, using stateful IT helpdesk bots as a use case. Modal wrote about pricing serverless GPUs for bursty, agentic workloads, noting that per-second billing works well when agents have irregular duty cycles.
The startup land grab is underway.
Why Persistence Actually Matters
The technical arguments for persistent agent environments cluster around three themes, though not everyone agrees on the ranking.
State management is the obvious one. LangChain described reducing checkpoint bloat for 200-turn coding agents—a problem that only exists when agents run long enough to generate that much state in the first place. AWS's guide on building stateful IT service desk agents with LangGraph on EKS walks through checkpointing, background runs, recovery patterns. Google Cloud's production agent guides emphasized long-running orchestration as a first-class design pattern. These aren't fringe use cases anymore.
Isolation matters because agents do risky things. Anthropic's Computer Use documentation includes privacy scopes and guidance to avoid dangerous operations—installing packages, browsing the web, running arbitrary code. Each of those actions is safer in a dedicated VM than in a shared environment. The NSA published MCP security considerations this year. The Cloud Security Alliance noted vulnerabilities in agentic command-and-control chains and framework CVEs across several months. When an agent can execute arbitrary commands, you want blast radius containment, not shared tenancy.
Operational simplicity is less discussed, but perhaps it's the real sell. Cursor published lessons from building "cloud agents," noting that parallelism and long-running tasks—hours, not minutes—require infrastructure abstraction. If you're manually provisioning and tearing down VMs every time an agent needs to run, you're not building a product. You're babysitting infrastructure. The value proposition of machine0, Orgo, and the rest is that you don't have to think about it. One command, one VM, minute-based billing, agent-controllable endpoints.
Whether that's worth the premium is another question.
The Economics of Always-On
The pricing models reflect the shift, at least in theory. machine0 bills per minute. Modal and Runpod emphasize per-second GPU billing. Baseten introduced fractional H100 GPUs via MIG slicing, with current pricing pages confirming continued availability. The pattern is consistent: charge for what you use, down to the second, because agent workloads are bursty and unpredictable.
Compare this to traditional cloud pricing. Reserved instances and committed use discounts assume you know your capacity needs in advance. That works for web servers. Not for agents that might run for eight hours today and two hours tomorrow. The economics favor short-term, fine-grained billing—in theory.
GPU availability remains the constraint that matters most. Voltage Park advertises H100s starting around $1.99 per hour, but community threads discuss availability bottlenecks. Together AI's Series C announcement mentioned projecting a roughly 50× infrastructure footprint increase over five years, with customer cost reductions as a selling point. CoreWeave's deals with Meta and Anthropic signal where the capacity is flowing. Lambda's $1 billion facility is explicitly for gigawatt-scale demand. The supply chain is scrambling, perhaps more than the founders expected.
Power and real estate are the real gating factors. Goldman Sachs analysis projected data center power demand could increase 50% by 2027 and as much as 165-175% by 2030, with a potential 45 GW shortfall in the U.S. by 2028. NVIDIA's Blackwell platform promises better inference throughput per watt, but if you can't power the racks, the efficiency gains don't matter. TechRadar Pro argued for moving creative production AI to the edge due to cost predictability and latency—a case that extends to persistent agent workloads when cloud capacity is constrained.
The math only works if the power shows up.
What Comes Next
The EU AI Act obligations took effect in August 2026, with staggered compliance dates extending into next year and beyond. High-risk AI systems face quality management, monitoring, and database registration requirements. That pressure favors auditable, isolated agent environments—exactly what these startups are building. The NSA's guidance on MCP security and the Cloud Security Alliance's work on agentic vulnerabilities suggest that compliance and security will shape the infrastructure layer as much as performance and cost.
The Blackwell rollout will unevenly impact availability and pricing across providers. Together AI's 50× scaling projection, Lambda's gigawatt-scale facility, and CoreWeave's $21 billion Meta deal all point to capacity concentration among a few players. That creates opportunity for startups like machine0 to focus on developer experience—CLI simplicity, minute billing, MCP endpoints—rather than competing on raw GPU count. Whether that's a sustainable moat is unclear.
The technical patterns are stabilizing, at least. Persistent VMs with suspend and snapshot workflows. Per-minute or per-second billing. CLI or SDK control with JSON outputs. MCP endpoints for agent-to-machine communication. These aren't speculative architectures anymore. They're production requirements, or close enough.
It's early enough that a solo founder with a CLI and a good thesis can launch through YC and compete. It's late enough that a16z, LangChain, and Google Cloud are all saying roughly the same thing: agents need their own computers. The infrastructure layer for always-on agents is being built right now, and the startups racing to define it are betting that the shift from ephemeral to persistent isn't a trend.
It's the new default. Or it will be, if the power grid holds.
