Daytona's $1B Bet: The Cloud Where Agents Live for 60 Milliseconds

$1 billion for 60 milliseconds.

That is the price investors put this week on Daytona, a San Francisco company that rents disposable computers to AI agents. On 20 October it closed a $60M Series B at a $1B valuation, led by FirstMark Growth. Total raised is $91M, eight months after a $24M Series A from the same lead.

The price is the easy part. The product is stranger. Daytona sells virtual machines built to live for less time than a blink and then vanish, to customers who never want to own the hardware underneath.

Sixty milliseconds, then gone

An ordinary cloud instance boots in minutes and runs until someone turns it off. An agent's computer behaves nothing like that. It starts in 60 milliseconds from pre-staged snapshots, pauses with its memory intact, forks into parallel copies, and disappears when the task ends.

Daytona runs Linux, Windows, macOS and GPU sandboxes on one platform. An agent can compile a CUDA kernel, drive a Windows desktop, or open Blender inside the same account. Burazin says Daytona was first to ship a Windows sandbox and remains the only provider offering one.

About 70% of revenue comes from post-training, where labs and data companies run agents to train and evaluate models. The rest is production work: agents acting on someone's behalf, often on long background jobs.

The customer list carries the pitch. Anthropic, OpenAI, Nvidia, Cursor and Cognition use the infrastructure; Meta, Cohere, Mercor, Handshake, Surge AI, SambaNova, Postman, Bolt.new, n8n and Clay sit on the roster. Snorkel has provisioned more than one million sandboxes. Writer ran its own platform until it passed 20,000 sandboxes a day, then provisioned 300,000 in a single month on Daytona. AfterQuery replaced an in-house GKE cluster in a day.

Agent infrastructure has been consolidating around state and isolation since the summer. As we wrote in October, the state layer is where agent tooling is being rebuilt. Daytona sells the other half of that problem.

Companies are going to manage thousands of digital employees, and each one needs a safe place to fail. Daytona ensures one rule-breaking agent doesn't spill into company-wide chaos.

That is Ivan Burazin, co-founder and CEO. He started the company in 2023 with Goran Draganić and Vedran Jukić, after building Codeanywhere and the Shift developer conference. The first product automated development environments for humans. The pivot to agents meant going back to near zero.

Growth since then has been unusual for infrastructure. Revenue run rate rose about 25-fold between January and September, an average of 50.9% a month. Net revenue retention averaged roughly 135% per month, which means customers from six months ago spend nearly three times as much today.

Snowflake's celebrated net retention at IPO was about 160% a year. Daytona claims close to that every month. Small base, steep slope.

It also sold almost nothing the traditional way. The company is self-serve and priced on consumption: sign up, add a card, call only when you need higher limits. About 1,100 shared Slack channels carry the customer relationship. Events and Burazin's X account did the rest, including a 2 million-view post that opened the Nvidia CPU test.

The case that isolation is a feature

The bear argument fits in one sentence. Sandboxing is a capability hyperscalers can ship, and several already sell isolated execution.

The market is not waiting for them. Modal raised $355M at a $4.65B valuation in May. Sail Research raised $80M in June. Baseten bought Blaxel, a sandbox specialist, to fold the same primitive into its inference platform. Capital is arriving from every direction, including the incumbents' own product teams.

Burazin has heard the objection directly. He says investors repeatedly asked why the market needed Daytona when AWS could do this thing.

Capacity is the constraint that decides who survives. Daytona bills per second: a 2 vCPU, 4 GiB sandbox lists at about 16.5 cents an hour, or under three cents for a ten-minute run. Each dollar of that revenue needs RAM and CPU reserved ahead of demand.

Run the arithmetic on a thousand always-on sandboxes at list price and the monthly bill lands near $119,000. Gross margin then depends on utilisation of capacity the company owns or leases. Daytona has not disclosed it.

The company has not disclosed absolute revenue either. A 25-fold rise from a small base is a different claim from the same multiple on a large one, and the Series A valuation was never published, so the step-up between rounds cannot be measured from outside.

Headcount is about 30, split between San Francisco and engineering offices in Split and Zagreb. The first head of sales joined this autumn. Until now the growth came from developer adoption and word of mouth, not from a quota-carrying team.

Dealroom places the $1B mark in the top 8% of US enterprise-software Series B rounds. That prices growth durability, and the durability is unproven.

What the round is buying

Three things follow from that sheet.

First, insiders carried most of the round. Pace Capital doubled its position, Upfront Ventures tripled, E2VC quadrupled. Existing investors putting in $50M of $60M is a conviction signal that costs them real money.

Second, the strategics are customers or suppliers. Qualcomm Ventures, Samsung Next and Okta Ventures joined alongside operators who run Stripe, Shopify, Perplexity, Zoom, Datadog, Quora and Mercor. Each name maps to a distribution channel or a hardware dependency.

Third, the proceeds go to capacity, not product theatre. Daytona says it will extend beyond the sandbox into a broader agent cloud, scale compute, and hire in San Francisco, Croatia and Singapore.

The CPU side of an AI trade

Most of the AI build-out is measured in GPUs. Burazin argues the next shortage lands somewhere else.

I strongly believe that there will be a CPU crunch and it'll be bigger than the GPU crunch.

His reasoning is structural. Models think on GPUs. Agents work on CPUs, and an agent fleet multiplies the number of machines a company runs. Post-training jobs are bursty, short and stateful, the opposite of the long-lived services today's cloud was designed for.

The economics point the same way. A GPU hour is expensive and scarce, so labs want it busy. Moving orchestration, file handling and tool calls onto cheap CPU sandboxes keeps the expensive silicon on matrix multiplication. Daytona says its approach is at least 36% faster than Kubernetes for these jobs, which also returns GPU time to training.

That thesis carries the risk. CPU and memory are commodity inputs with visible spot prices. If agent execution becomes a large market, margin migrates to whoever owns capacity at the lowest cost. That is usually the hyperscaler, and Daytona is not one.

Two numbers to watch

The first is net revenue retention. At 135% monthly, expansion does the work a sales team has not started doing. Watch whether it holds once procurement cycles replace credit-card sign-ups.

The second is gross margin, currently undisclosed. Per-second billing on owned capacity can look like software at small scale and like a utility at large scale. The crossover will show up in the next raise, not in this one.

The valuation is a wager that the execution layer keeps the spread between per-second pricing and the cost of the metal underneath. Logo count is the easy part. The spread is the business.