AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
Summary
Daytona is betting that every AI agent will need an API-addressable computer, making today’s “sandbox” category the entry point to a much larger infrastructure market. The company has reported 74% month-over-month growth while the broader infrastructure market grew around 40%; Burazin’s decisive product signal was prospects chasing missing API keys: “People literally call you if you do not give them access.” His question is the thesis: “The market for every single agent that will exist ever in the future is just like—what is that market?”
The architectural wager is that agents need stateful, long-running computers with serverless speed, not disposable containers designed for human development workflows. Daytona combines “a Lambda and an EC2”: bare-metal scheduling, local NVMe snapshots, pause-and-resume state, and dynamically composable CPU, RAM, disk, operating systems, and eventually GPUs. Burazin reports 60-millisecond single-sandbox startup, roughly 75 seconds for 50,000 concurrent starts, and a largest customer approaching 850,000 sandboxes per day.
RL and eval workloads have gone from zero to an expected 50% of Daytona’s usage, but their economics punish conventional capacity planning. Unlike background agents that follow human workdays, research runs arrive as unpredictable “squares”—10,000, 50,000, or 100,000 CPUs suddenly driven to 100%, then released. Daytona averages only 15% utilization but reaches 90% at peaks, making customer commitments and capacity planning necessary to keep expensive GPUs from waiting for slower CPU provisioning.
Computer use may be the largest demand unlock because APIs still expose only part of the work trapped inside legacy applications. Burazin estimates roughly $50 trillion of global knowledge-worker compensation, with about half in major sectors and perhaps 40% automatable by more capable agents—a deliberately rough path to a $10 trillion TAM for models. Daytona’s Windows sandbox starts in about one second versus 3–5 minutes on EC2 or Azure, aiming to give agents the snapshots, forking, and persistence required to act as a “human emulator.”
The strongest route to scale is selling infrastructure through app-layer agent companies, not directly to individual OpenClaw users. Daytona’s core customers are B2B or B2B2C products such as agent platforms, labs, and applications in the Manus/Lovable mold; Burazin compares its consumption model to Twilio or Stripe rather than AWS. Claude Code and CLI-first workflows expand the pull because an agent can execute scripts and analyze data, while MCP alone remains “an interface against an API.”
Open source helps trust and integration, but Burazin does not present it as Daytona’s primary growth engine. The current stack is largely AGPLv3, including its scheduler, while Windows and GPU capabilities sit behind feature flags; open code gives integrating agents more context and can ease adoption inside large enterprises. The more consequential opportunity may be an agent-native layer before GitHub: one customer versioned entire codebases as JSON dumps to S3, while another produced 1,000 PRs daily and turned CI into the bottleneck.
Burazin expects a reckoning for SaaS companies calling token resale “reacceleration,” while durable value shifts toward metered APIs and agent-native clouds. Agent revenue carries different margins and stickiness from classic SaaS and uses models from providers such as Anthropic or OpenAI, so he wants vendors to expose all underlying data and charge agents for consumption rather than wrap another siloed assistant around it. Daytona may add GPU sandboxes and more primitives, but not its own data centers yet; the end state could resemble “Stripe more than AWS”—a cloud purpose-built for agents, with computers, search, databases, and primitives not yet invented.
Deep dive
Not yet available upstream; scheduled sync will retry.