Daytona raises $60 million at a $1 billion valuation to build computers for AI agents

A unicorn built on agent sandboxes

Daytona, a start-up that rents out short-lived computers for AI agents, has raised a $60 million Series B at a $1 billion valuation. Fortune first reported the round on 20 October, and the company published its own announcement saying the round was led by FirstMark, which also led its $24 million Series A in February. Daytona has now raised $91 million in total, and Fortune reported that about $50 million of the new round came from existing investors.

The round has an unusual cast of angel investors, many of them chief executives whose own products run agents. Daytona and TechCrunch list Stripe's Patrick Collison, Shopify's Tobi Lütke, Perplexity's Aravind Srinivas, Zoom's Eric Yuan, Datadog's Olivier Pomel, Quora's Adam D'Angelo and Mercor's Brendan Foody. Qualcomm Ventures, Samsung Next and Okta Ventures joined as strategic investors, alongside existing backers including E2VC, Upfront and Pace.

Coverage differs on where Daytona is based. Fortune describes it as a San Francisco start-up and the company says it is hiring on site there, while TechCrunch calls it New York-based. The founders, Ivan Burazin, Goran Draganić and Vedran Jukić, started the company in 2023 and Daytona says Burazin previously built Codeanywhere, a cloud coding tool, and the Shift developer conference.

What an agent cloud actually sells

Daytona's argument is that the cloud was built for people, not agents. A normal virtual machine takes from seconds to minutes to boot, then runs until someone turns it off. An AI agent that writes code, tests it, browses the web or practises a task in reinforcement learning wants something different: a computer that starts almost instantly, can pause with its state intact, can be copied into many parallel versions and disappears when the job is done.

Burazin told Fortune that Daytona can start a new sandbox in about 60 milliseconds, against 30 seconds or more for a configured virtual machine at a big cloud provider. The company says it does this by spreading prepared snapshots across its machines in advance. Its sandboxes can be paused, forked into copies and rewound to an earlier state, which is useful when an agent wants to try several approaches at once or back out of a mistake.

Daytona also claims to be the only provider that offers Linux, Windows, macOS and GPU sandboxes on one platform, so an agent can tune a CUDA kernel, test a Mac app or drive a Windows desktop in the same place. That claim comes from the company and is hard to check independently, but it matches the direction of the market: agents are moving from chat windows to full computers.

The growth numbers, with caveats

Daytona says its revenue has grown about 25 times since January, and that its average month-over-month net revenue retention, a measure of how much more existing customers spend over time, is about 135 percent. It says customers from six months ago now spend nearly three times as much in aggregate. Burazin told Fortune the company has grown 50.9 percent a month on average.

None of those figures come with absolute revenue, and they are the company's own numbers. Growth from a small base can look spectacular in percentage terms, and monthly retention figures are not comparable with the annual figures most software companies report, as Daytona itself notes when it compares its number with Snowflake's annual figure at IPO. The valuation suggests investors have seen the underlying revenue, but outsiders have not.

The customer list is more concrete. Daytona names Meta, Cohere, Handshake, Snorkel, Surge AI, Turing, AfterQuery, Trajectory, Writer, Clay, n8n, Postman, Mintlify and bolt.new, and Fortune adds Anthropic, OpenAI, NVIDIA, Cursor and Cognition. Snorkel has run more than a million sandboxes on it. Writer ran its own sandbox system until it passed 20,000 sandboxes a day, then provisioned 300,000 in a single month on Daytona. AfterQuery, an applied AI research lab, says it replaced its in-house Kubernetes cluster on Google Cloud with Daytona in a day.

Burazin's CPU crunch warning

The most interesting part of the pitch is about chips. AI models run on GPUs, but the computers agents use to run code and tools run on ordinary CPUs. Burazin told Fortune that Daytona internally expects spending on CPUs to end up higher than spending on GPUs, and that he strongly believes there will be a CPU crunch because the machines needed for this growth do not yet exist.

He backed that with rough maths about Meta's Muse agents, which Meta said this week had passed 10 million users. By Burazin's back-of-the-envelope estimate, Muse would need $2.8 billion of sandboxes to serve 100 million users. That is his estimate, not Meta's, but it shows why agent infrastructure companies think the market could be very large.

Other parts of the chip industry are saying similar things. Fortune notes that AMD raised its server CPU growth forecast to more than 35 percent a year, that Arm has estimated agentic AI data centres could need 120 million CPU cores per gigawatt, and, citing Quartz, that Intel told some Chinese customers lead times for certain server CPUs had stretched to six months. Modal, a Daytona rival, recently said it would move its agent sandboxes to NVIDIA's Arm-based Vera CPUs after seeing up to twice the throughput.

The big clouds are coming

Daytona's biggest risk is that the hyperscalers decide to build the same thing. Google Cloud this week launched Agent Compute Engine, which gives each AI agent its own hardware-isolated microVM that checkpoints when idle. Perplexity has open-sourced SPACE, the Firecracker-based sandbox behind its Computer agent, which lowers the bar for anyone who wants to run their own.

Burazin's answer in Fortune is that the big clouds built most of their services for maximum use of long-running machines, and that rebuilding them for millions of tiny, short-lived agent computers is not easy. He conceded they could do it. The company's own history shows it is willing to pivot: Burazin told Fortune that when the founders realised agents would need computers, they gave their customers to competitors, rebuilt the product and relaunched with about three months of revenue left in the bank.

There is also an open-source wrinkle. Daytona's public GitHub repository, which had about 71,500 stars when we checked, is marked as no longer maintained after the company moved its core development private in June 2026. The older code remains free to use and fork under its existing licence, but customers are now buying a closed, hosted service.

What the money is for

Burazin told Fortune that most of the new money will go into hiring sales and go-to-market staff, plus some engineers. FirstMark's Matt Turck said in a statement that the next infrastructure shift is from human-centric cloud primitives to agent-native ones. Qualcomm Ventures' Quinn Li said Daytona has grown from an agent sandbox into a full platform for running agents in production, according to the company announcement.

European outlet tech.eu reported that the money will go into expanding Daytona's cloud infrastructure, scaling computing capacity and growing its international team. For the wider AI industry, the deal is another sign that money is spreading from model training to the plumbing agents need to do real work. Whether Daytona stays independent as Google, Amazon and Microsoft push into the same space will be the test of that $1 billion price tag.