Daytona hits $1B: AI agent sandboxes are a CPU bill now
Daytona raised a $60M Series B at a $1B valuation for AI agent sandboxes that boot in 60ms. Budget the compute your agents use, not just the tokens.
Daytona, which sells sandboxes where AI agents run code, raised a $60 million Series B at a $1 billion valuation, led by FirstMark. The pitch is simple: models think in the inference cloud, and agents do their work somewhere else. That somewhere else is a virtual machine, and it bills like one. If your AI agent budget only counts tokens, it is missing a line.
What actually happened
Per Fortune:
- The round: $60 million, with about $50 million from existing investors. Daytona has raised about $91 million in total. Angel investors include Patrick Collison, Tobi Lütke, and Aravind Srinivas.
- The product: sandboxes where agents write and run code, browse the web, and use business software. Daytona says a new sandbox starts in about 60 milliseconds, because it pre-stages machine snapshots on disk and attaches CPU and memory on demand. Agents can also clone a running machine to try several approaches in parallel.
- The customers: Fortune lists Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta, and Cohere. Snorkel alone has created more than one million sandboxes.
- The growth: revenue run rate grew 25x from January to September.
- The bet: CEO Ivan Burazin expects CPU spending to pass GPU spending and predicts a CPU shortage. Fortune notes Intel reportedly told some customers that lead times on certain server CPUs hit six months.
Fortune did not report Daytona's pricing.
Why agent sandbox costs matter for your business
Every agent run has two meters. The model call costs tokens. The tool calls cost compute: a container to run the script, a browser to fill the form, a disk to hold the files. One agent that clones its machine to test three fixes pays for three machines.
Fast boot changes the design. When a sandbox starts in milliseconds, it is easy to give every task a fresh, disposable machine. That is good for security, because nothing from one customer's job leaks into the next. It also means volume grows fast, so set limits per task and per day.
CPU scarcity becomes your price. If server CPUs get tight, sandbox vendors pass that on. Watch your per-run cost the same way you watch your token cost.
Keep the sandbox swappable. Daytona, the hyperscalers, and your own servers all run containers. Write your agent's tools against a thin interface so the execution layer can move.
Key takeaways
- Daytona raised $60M at a $1B valuation for AI agent sandboxes that start in about 60ms
- Customers include Anthropic, OpenAI, Cursor, and Meta; run rate grew 25x from January to September
- Agent costs have two parts: tokens for thinking and CPU for doing
- Give each task a fresh sandbox, then cap how many a task can start
- Keep the execution layer behind an interface you own so you can switch vendors
Do you know what one agent run costs, end to end? We build agents with metered, disposable sandboxes and a cost line for every run, on infrastructure you can move. Plug your volumes into our ROI calculator, or tell us what your agent needs to do.
Sources: Fortune.
- #daytona
- #ai-agent-sandboxes
- #agentic-cloud
- #funding
- #infrastructure-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenAI COBOL migration test: 5.4% to 86.6% in one model
OpenAI says GPT-6.1 Sol migrates COBOL at 86.6% accuracy, up from 5.4%, at 39% lower cost. Legacy code migration got cheaper. Proving it still costs you.
Read itGoogle Coral NPU + Synaptics: $99 edge AI board ships
Synaptics ships the $99.79 Coralboard SL and new SR200 chips on Google's open Coral NPU. Edge AI lets cameras and sensors work without the cloud.
Read it