Every gold rush eventually produces someone who notices that the miners need somewhere to sleep. Daytona has just been valued at a billion dollars for noticing it early, on roughly ninety-one million raised to date, which is the part of the story worth sitting with.
The company is building what it calls the agentic cloud: infrastructure for software agents to actually run in. Not the model, not the orchestration layer that decides what the agent should attempt, but the room it occupies while attempting it. A sandbox that boots in a moment, holds state, executes code it has never seen before, and can be destroyed without taking anything else down with it. The pitch sounds unglamorous next to the frontier labs, and that is precisely its commercial shape.
The demand profile here is genuinely new. A human developer opens an environment in the morning and closes it in the evening, which is an industry built on long sessions and predictable load. An agent opens ten thousand environments in an hour, uses each for forty seconds, and discards it. The unit economics, the security model and the cold-start problem all change character under that pattern, and existing cloud primitives were not shaped for it. Whoever makes ephemeral, isolated and instant into a single product owns a toll booth on every agent that does anything other than talk.
It is the espresso machine problem. The cafes that made money through the specialty boom were rarely the ones with the most romantic story about single-origin lots. They were the ones that bought the right grinder, understood that the machine’s recovery time between shots was the actual constraint on a morning rush, and plumbed the water line properly before anyone queued. The flavour is what the customer notices. The throughput is what the business is. Daytona is selling recovery time between shots to an industry that has spent three years talking about flavour.
The capital efficiency is the detail that separates this from the general noise of the cycle. Ninety-one million raised in total against a billion-dollar mark is a narrow base for that number, and narrow bases tend to mean the market is pricing a position rather than a run rate. Infrastructure companies earn those multiples when the thing they built turns out to be a dependency rather than a choice, because dependencies compound: every agent framework that standardises on your sandbox brings its own users, and the switching cost grows with the volume that flows through it. That is a very different asset from a product people like.
The risk is equally structural and worth naming without flinching. The hyperscalers can see this demand curve as clearly as anyone, and sandboxing is a feature they could plausibly bundle. The counter-argument is the one infrastructure founders have been making since the first managed database, which is that bundled features are built for the average case and the agentic workload is nothing like the average case. Being specialised at the exact moment a workload becomes weird is how independent infrastructure survives a platform’s attention. It is also how it gets acquired, which from a cap table’s point of view is not a complaint.
What the billion really marks is a shift in where the industry thinks the scarcity sits. For three years the scarce thing was intelligence, and the money followed the models. The valuation says the scarce thing is now somewhere safe for that intelligence to put its hands, which is a quieter proposition and a stickier one.
Agents, it turns out, are tenants. Someone was always going to be the landlord.
