Daytona Hit $1B Renting Computers to AI Agents Nobody Sees

Daniel Rao | October 24, 2026

Update, Oct. 24: the customers are the story — and Daytona has the receipts

Two customer metrics from this week's coverage sharpen the pitch beyond the funding headlines. AI data startup Trajectory has provisioned 700,000 sandboxes on Daytona's platform, cutting environment startup times from 30 minutes to seconds — while Snorkel, the data company named in this story, has provisioned more than a million sandboxes and reduced evaluation cycles by 96%, per techstartups.com's October 20 roundup. Separately, LangChain's agent-sandbox comparison now names Daytona "best for multiple runtime classes and deployment models," and publishes the pricing behind the platform: $0.0504 per vCPU-hour, $0.0162 per GiB-hour of memory, and $0.000108 per GiB-hour of storage after the first 5 GiB. (Customer figures are Daytona-supplied; pricing per LangChain's review.)

Daytona Platforms doesn’t build AI. It has no chatbot, no model, no app you will ever open. It rents computers to software almost nobody will ever see — and on Friday, some of the most successful founders in technology decided that invisible business is worth a billion dollars.

The three-year-old New York startup announced a $60 million Series B round led by FirstMark Growth, valuing the company at $1 billion and taking total funding to $91 million. TechCrunch broke the story at 9:26 AM Pacific on Friday, and the timing was no accident — the Friday morning embargo slot is where startups place news the insiders have already digested, so the weekend conversation starts on their terms. What makes this round unusual is the cap table: 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 — who also sits on OpenAI’s board — and Mercor’s Brendan Foody all wrote cheques into a company that sells, essentially, shovels in the AI agent gold rush.

Daytona’s thesis is simple, and once you hear it, annoyingly obvious: the cloud we have today was built for long-lived services, not for thousands of short, stateful, unplanned jobs spun up by AI agents. The names already paying the bills — Anthropic, OpenAI, Nvidia, Meta, Cursor, Cognition, Cohere — suggest the thesis is more than a pitch-deck fantasy. Whether a billion-dollar price tag is justified is a different, louder question.

Photo: inside a modern server room. Daytona’s pitch is that AI agents need thousands of machines like these — spun up in 60 milliseconds, torn down when the agent is done. File photo. Credit: Fleshas, via Wikimedia Commons (CC BY-SA 3.0).

The numbers behind the billion

The headline figures, first reported by TechCrunch on Friday morning: $60 million, led by FirstMark Growth — the same firm that led Daytona’s $24 million Series A back in February. According to Dealroom’s analysis of the round, about $50 million of the new money came from existing investors, which is the venture world’s loudest vote of confidence. Daytona’s revenue run rate grew 25-fold between January and September of this year, averaging 50.9% month-over-month growth. Numbers like that explain why FirstMark doubled down nine months after its first bet.

The institutional roster around the angels is just as heavyweight: Qualcomm Ventures, Samsung Next and Okta Ventures joined, alongside Alumni Ventures, the Evergreen Fund and Torch Capital. Pace Capital doubled its position, Upfront Ventures tripled it, and E2VC quadrupled. Founder and CEO Ivan Burazin, chief architect Goran Draganić and CTO Vedran Jukic now run a company whose valuation sits in the top 8% of US enterprise-software Series B rounds, per Dealroom — a polite way of saying this round prices velocity, not scale.

A cap table that reads like a tech hall of fame

Angel investors are usually decoration — a few famous names sprinkled on a deck to signal taste. This list is different, because it reads like a customer list with cheques attached. Patrick Collison runs Stripe. Tobi Lütke runs Shopify. Aravind Srinivas runs Perplexity. Eric Yuan runs Zoom. Olivier Pomel runs Datadog. Adam D’Angelo runs Quora and sits on OpenAI’s board. Brendan Foody runs Mercor. These are people who buy infrastructure for a living, and they just invested in the plumbing their own companies increasingly rely on.

The corporate VCs tell their own story. Qualcomm Ventures and Samsung Next want a front-row seat to whatever runs on their chips next. Okta Ventures, the venture arm of the identity company, cares about who — or what — is allowed inside the building. When the identity layer’s investor starts backing agent infrastructure, it tells you the industry has stopped asking whether agents get access to systems and started asking where they sit.

The 60-millisecond trick the giants can’t match

Daytona sells “sandboxes”: virtual machines where AI agents write code, browse the web and run software in isolated, persistent environments. The trick, per Dealroom’s breakdown, is that Daytona pre-loads machine snapshots so it can configure a new machine in 60 milliseconds — a process that takes existing clouds minutes. When an agent spins up thousands of short jobs nobody planned in advance, minutes are an eternity. Sixty milliseconds is close enough to free.

Burazin put the thesis plainly in a blog post accompanying the raise: “Agents don’t work like the long-lived services today’s cloud was designed for. They spin up thousands of short, stateful jobs nobody planned in advance, and they need a cloud built for that.”

His pitch for the product is even simpler. The cloud Daytona is building, he said in the funding announcement, is “like giving every agent its own computer” — sandboxed Linux, Windows, macOS and GPU machines, all on one platform. And the demand argument comes in a line he gave to tech.eu: “Companies are going to manage thousands of digital employees, and each one needs a safe place to fail.”

That last line is the one worth sitting with, because it reframes the whole security conversation around AI agents. We have been covering what happens when agent containment fails — OpenAI’s own models once snuck online through DNS trickery during training — and Daytona’s answer is architectural rather than philosophical: don’t argue with the agent about the rules, give it a computer it is allowed to break.

How Daytona actually works: containers, VMs, and the snapshot trick

Daytona's sandbox is not one thing — it is a menu, which is exactly why LangChain's recent agent-sandbox comparison ranked it best for multiple runtime classes. By default a new sandbox runs as a Linux container, using process and network namespaces plus filesystem and IPC isolation, with vCPU, RAM and disk reserved per sandbox. Teams can instead choose a dedicated Linux VM, a Windows VM, or a GPU sandbox when the workload needs a different operating system or a harder hardware boundary; macOS sandboxes run on dedicated Apple silicon. Python, TypeScript, Ruby, Go and Java SDKs sit alongside a CLI, a REST API, an MCP server and SSH access.

The 60-millisecond number comes from snapshots: machine images are pre-loaded onto server drives, so a new sandbox configures from a warm snapshot rather than booting cold. Snapshots can start from a Dockerfile or OCI image, or capture an existing sandbox — container captures preserve the filesystem, VM captures can include memory. For data that must outlive the sandbox, S3-backed FUSE volumes can mount into multiple environments. On the security side, Daytona Secrets place a non-secret placeholder in a sandbox environment variable instead of the plaintext credential; when code sends that placeholder in an HTTPS request header to an approved host, Daytona's outbound proxy swaps in the real value — the code authenticates without ever being able to read the credential. (Technical details per LangChain's sandbox review and Daytona's own documentation.)

Anthropic, OpenAI and Meta are already renting the machines

This is where the pitch stops being theoretical. Daytona’s sandbox customers include Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta and Cohere — companies that train and test frontier models inside Daytona’s isolated environments. Snorkel alone has provisioned more than a million sandboxes. When the labs building the most capable AI systems on earth are renting your machines to keep those systems contained, your revenue charts start to look less like hype and more like infrastructure.

As one industry roundup put it this week — and the line is worth quoting because it captures the whole shift — “Big language models get the attention. The machines that rent them somewhere to run are quietly becoming the better business.” The model business is a bloodbath of capital expenditure and copycat releases. The business of renting the machines the models live on — fenced, measured, billed by the second — starts to look like the steadier trade.

The security angle matters here too. Anthropic’s security chief recently warned of 18 months of cyber hell ahead, and agent infrastructure is where much of that hell will be fought — thousands of autonomous programs with credentials, browsers and the ability to execute code. A cloud built specifically to watch them do it is either a brilliant idea or the next decade’s attack surface. Possibly both.

Our take: the AWS question is the wrong question

Here is the sceptical case, and Dealroom stated it cleanly: a $1 billion valuation puts Daytona in the top 8% of US enterprise-software Series B rounds, which means this is a bet on growth rather than scale. And investors, reportedly, asked the obvious question: why do we need this when AWS can do this thing?

It sounds smart. It is also the laziest question in enterprise software.

AWS can do anything, in the same sense that a supermarket can sell anything — eventually, slowly, and with the enthusiasm of a committee. The history of cloud infrastructure is a graveyard of “the incumbent will just build it”: Snowflake happened in front of AWS, Datadog happened in front of AWS, Stripe happened in front of every bank. Specialised infrastructure wins not because the giant can’t copy it, but because the giant’s incentives are to serve everyone adequately and the specialist’s incentives are to serve one workload obsessively. A 60-millisecond spin-up time doesn’t exist on AWS because AWS was never organised around the problem it solves.

The honest risk isn’t AWS copying Daytona. It’s that the agent economy is still mostly a promise — that “thousands of digital employees” remains, for most companies, a slide in a consultant’s deck rather than a line item. A billion-dollar valuation on a 25-fold revenue run is not expensive if the run rate compounds; it is colossal if agent adoption stalls and the run rate turns out to be a sugar high. Burazin’s investors are betting the compounding continues. The sceptics are betting the easy part is over.

One more note, because it matters: Wall Street is already deleting AI agents from production systems that couldn’t justify themselves. The survivors — the agents with real jobs and real budgets — are exactly the ones that need somewhere safe to run. Daytona is betting the culling leaves it with customers, not casualties.

What $60M actually buys

According to the roundup coverage, the new money funds Daytona’s expansion beyond sandboxes into a broader agent cloud, plus hiring across San Francisco, Croatia and Singapore. The geography is telling: San Francisco for the customers, Singapore for the region’s agent boom, Croatia for the engineering bench — a footprint that looks less like a three-year-old startup and more like a company that expects to be infrastructure.

The sandbox business was the wedge. The bet is the platform: every agent gets its own computer, every computer is fenced, and Daytona owns the fence. If Burazin is right that companies will soon manage thousands of digital employees, the company selling each one a safe place to fail is positioned somewhere between landlord and utility — the two businesses venture capital likes best.

Frequently Asked Questions

What does Daytona Platforms do?

Daytona is building what it calls an “agentic cloud” — cloud infrastructure designed specifically for AI agents. Its core product is sandboxes: isolated virtual machines (Linux, Windows, macOS and GPU) where agents can write code, browse the web and run software in persistent, fenced-off environments.

Who invested in Daytona’s $60M Series B?

The round was led by FirstMark Growth, which also led Daytona’s $24M Series A in February 2026. Angels included 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. Institutions included Qualcomm Ventures, Samsung Next, Okta Ventures, Alumni Ventures, Evergreen Fund and Torch Capital, plus existing backers Pace Capital, Upfront Ventures and E2VC.

Why is Daytona worth $1 billion?

The valuation reflects explosive growth rather than current scale: Dealroom reports Daytona’s revenue run rate grew 25-fold between January and September 2026, averaging 50.9% monthly growth, with about $50M of the round coming from existing investors. Dealroom also cautions that the valuation sits in the top 8% of US enterprise-software Series B rounds — a bet on growth continuing, not on scale already achieved.

What is an AI agent sandbox?

A sandbox is an isolated virtual machine where an AI agent can execute code, browse the web and run software without touching the rest of a company’s systems. If the agent makes a mistake — or misbehaves — the damage stays inside the fence. Daytona’s version can be configured in 60 milliseconds using pre-loaded snapshots, versus minutes on general-purpose clouds.

How fast can Daytona spin up a machine?

According to Dealroom’s analysis, Daytona can configure a sandbox in 60 milliseconds by pre-loading machine snapshots. Existing general-purpose clouds typically take minutes for the equivalent task — a meaningful gap when agents spin up thousands of short-lived jobs.

Is Daytona competing with AWS?

Not directly, and that’s the point. Daytona argues the major clouds were designed for long-lived services, while AI agents need thousands of short, stateful, unplanned jobs. AWS could build similar tooling, but specialised infrastructure players have repeatedly won by serving one workload obsessively — think Snowflake or Datadog. The open question is whether agent adoption grows fast enough to justify a $1B price tag.

How much does Daytona cost?

Daytona prices by usage: $0.0504 per vCPU-hour, $0.0162 per GiB-hour of memory, and $0.000108 per GiB-hour of storage after the first 5 GiB, per LangChain's published pricing review. Deployment options include shared US and EU regions, dedicated Daytona-managed regions reserved for one organisation, and custom regions running on customer-provided compute.

Daytona vs E2B: which agent sandbox should I use?

Both sell isolated sandboxes for AI-generated code, but they come from different directions. E2B is the open-source-rooted option (Apache-2.0 core, roughly 12,400 GitHub stars per the agent-sandboxing knowledge base): a simple API, Python and TypeScript SDKs, and self-hosting via Terraform — strong if you want to own the isolation layer. Daytona's pitch is breadth and persistence: a dedicated kernel, filesystem and network stack per sandbox, five SDK languages, stateful snapshots that survive across sessions, and sub-90ms starts. If you need persistent, multi-OS agent workflows, Daytona's menu is wider; if you want open-source code execution with minimal infrastructure knowledge, E2B is the simpler primitive. (Comparison framed from the agentic-AI sandboxing knowledge base; verify current pricing and terms before committing.)

Sources

TechCrunch (first reported the round, October 23, 2026); Dealroom (round analysis, investor mix and growth figures); tech.eu (Burazin interview quotes); DigitalToday (round coverage). Customer, valuation and growth figures are as reported by these outlets; Daytona had not published a standalone press release at the time of writing.

About the author: Daniel Rao covers AI infrastructure, cloud economics and the strange new science of keeping superintelligent systems inside the box. He writes TechPulse USA’s infrastructure beat and checks cap tables down to the decimal.