NewsAI & DevelopmentDeveloper ToolsInfrastructure

Daytona Gets $60M to Build the Cloud Layer AI Agents Need

Writer runs 300,000 Daytona sandboxes per day. One month before, it was 20,000. That single data point explains why Daytona just raised $60M at a $1 billion valuation — and why the execution layer beneath AI agents is quietly becoming one of the most important infrastructure bets of 2026.

The Problem Traditional Cloud Was Never Built to Solve

When you spin up a standard EC2 instance or a Kubernetes pod, you wait. Thirty seconds, sometimes a minute. That is acceptable when a human is on the other end clicking a button. It is not acceptable when an AI agent needs to run 10,000 short tasks and every one of them requires an isolated environment.

Agents also do things that humans do not: they fork (run the same state in parallel to explore different approaches), they pause mid-task and need to resume exactly where they left off, and they disappear the moment a job is done. Traditional cloud abstractions — instances, pods, persistent volumes — were designed around human workflows. Per-hour billing, SSH access, dashboard management. None of that fits the agentic compute model.

Meta puts a number on the problem: serving 100 million users of its Muse agent would require approximately $2.8 billion in sandbox infrastructure. That is not a complaint. That is a market signal.

What Daytona Built Instead

Daytona uses snapshot-based execution. Instead of provisioning a new environment from scratch, the platform pre-distributes machine snapshots onto server drives. When an agent needs a sandbox, Daytona attaches CPU and memory to resume that snapshot — target latency: under 90 milliseconds from request to running code.

Every sandbox is a full isolated computer: dedicated kernel, filesystem, network stack, allocated vCPU, RAM, and storage. Inside each one runs a Toolbox API that gives agents direct programmatic access to file operations, Git, code execution, browser automation, and log streaming. No dashboards. No human-in-the-loop steps. Agents interact entirely through APIs.

The SDK reflects that philosophy. Install with pip install daytona-sdk, then:

from daytona import Daytona

daytona = Daytona()
sandbox = daytona.create()
response = sandbox.process.code_run("print('hello')")

Three lines. Sandbox running in under 90 milliseconds. Cleaned up automatically when done. The full getting-started guide covers TypeScript, Go, Ruby, and Java equivalents.

The Business Behind the Raise

Daytona has grown revenue roughly 25x since January 2026. Monthly net revenue retention sits at around 135% — the company compares this to Snowflake’s annual retention at its IPO. It is processing 850,000 sandbox runs per day. Those are not prototype metrics.

The $60M Series B was led by FirstMark Growth, with personal investments from Patrick Collison (Stripe), Tobi Lütke (Shopify), Aravind Srinivas (Perplexity), Eric Yuan (Zoom), and Olivier Pomel (Datadog). Strategic participants included Qualcomm Ventures, Samsung Next, and Okta Ventures.

“The next infrastructure shift is from human-centric cloud primitives to agent-native ones.”

Matt Turck, FirstMark Growth

Pricing and Where Daytona Fits

Billing is per second. A standard 2 vCPU / 4 GiB sandbox costs $0.0276 for a 10-minute run — less than three cents. GPU sandboxes (RTX 4090) start at $0.57/hour; H100s run $2.27–$3.95/hour. New accounts get $200 in free credits.

For developers choosing between options: E2B uses Firecracker microVMs and offers slightly faster cold starts in some benchmarks, but is Linux-only with no GPU support. Modal has native GPU capability but costs about 40% more for CPU workloads. Daytona’s advantage is multi-OS support (Linux, Windows, macOS sandboxes on the same platform) and the combination of CPU and GPU options at comparable pricing.

If you are running a handful of agent tasks a day, the choice barely matters. If you are scaling toward Writer’s territory — hundreds of thousands of daily runs — the infrastructure decision becomes a major cost and reliability variable.

The Bigger Question

The three major cloud providers are not well-positioned to own this category. AWS, GCP, and Azure increasingly treat AI companies as competitors rather than pure customers — the open-source version of Daytona illustrates the community appetite for infrastructure that is not tied to a hyperscaler. Their pricing granularity, provisioning assumptions, and human-centric design all work against agentic workloads.

Whether Daytona becomes the default execution layer the way S3 became the default object store is an open question. But the case that agent-native execution infrastructure is a distinct and valuable layer — separate from inference, separate from orchestration — is no longer speculative. Writer’s sandbox count makes that case better than any pitch deck could.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News