Daytona

There is a slightly absurd fact at the center of the AI boom, which is that when a large language model writes code, the code has to run somewhere. Not on the model - the model is a probability distribution, it cannot execute a for-loop - but on an actual computer, with a filesystem and a network connection and all the ordinary machinery that computers have. And the code it writes is, definitionally, code that no human reviewed before it ran. This is a lovely arrangement if you enjoy watching untrusted programs touch your production servers. Most people do not.

Daytona sells the escape hatch. The company builds what it calls programmatic, composable computers for AI agents: sandboxes that spin up in under 90 milliseconds, run whatever the agent wrote in complete isolation, and then get thrown away without ever having touched anything that matters. You can start one, pause it, snapshot it, and resume it later, which sounds mundane until you remember the customer here is not a person but a piece of software that may want to do this several thousand times before lunch.

The framing the company likes is blunt: give every agent a computer. It is the title of the press release announcing its $24 million Series A, and it is also, more or less, the entire thesis. If you believe that autonomous agents are going to do real work - write software, analyze data, click around the internet on your behalf - then each of those agents needs a place to actually do the work. That place is the product.

What makes this genuinely interesting - and not just another pin on the AI-infrastructure map - is that Daytona did not set out to build it. The company began life as a developer-environment manager, an open-source tool that first appeared on GitHub in February 2024. Think Gitpod or Coder: standardized, reproducible workspaces so that human engineers stop losing afternoons to "it works on my machine." A perfectly good business. Just, as it turned out, not the business.

Sometime in early 2025 the founders noticed that a growing share of the things asking for a clean, disposable, fully-configured computer were not humans at all. They were agents. And agents have very different requirements than people: they don't need a nice IDE or a pretty terminal, they need to spin up by the million, run for seconds, and vanish. So Daytona did the thing that is easy to say and hard to do - it rebuilt the company around the customer it actually had rather than the one it had planned for. By the second quarter of 2025 the pivot was complete.

Daytona was founded by Ivan Burazin (CEO), Vedran Jukic (CTO), and Goran Draganic (Chief Architect). Burazin is the kind of founder who has been circling this problem for most of his career: he co-founded Codeanywhere, a browser-based cloud IDE, back when "coding in a browser" sounded slightly deranged, and he built Shift, a developer conference in Croatia that grew into one of the largest in Southeastern Europe before Infobip acquired it. The through-line is roughly two decades of asking where developers do their work - and now asking the same question about the software that is starting to do the developing.

The team is small, which is worth sitting with for a moment. Around the time of the Series A, Daytona was about 20 people serving on the order of 850,000 sandbox runs a day, growing 74% month over month. That is a lot of computers for a company that could fit around a couple of dinner tables. The plan for the funding is unglamorous in the best way: add capacity, expand into more regions, and hire - because the constraint isn't demand, it's hardware.

The economics follow from the shape of the thing. Daytona charges for compute consumed - CPU, memory, GPU, storage - metered by usage, with an open-source core and a managed cloud on top. The customers span the full range of the AI market, from Y Combinator startups on their first product to Fortune 100 enterprises. Named ones include LangChain, Writer, Turing, SambaNova, Mintlify, Clay, Snorkel AI, and CoreWeave. The company reached a $1 million forward revenue run rate within three months of the pivot, then doubled it six weeks later, which is the sort of curve that stops looking like a business plan and starts looking like weather.