Giving Agents Computers — Ivan Burazin, Daytona

What It Covers

Ivan Burazin, CEO of Daytona, explains how his company pivoted from developer environment automation to building composable compute sandboxes for AI agents. Daytona now processes 850,000 daily sandbox runs for customers, growing 74% month-over-month, with architecture built on bare metal for 60-millisecond spin-up times and 50,000 concurrent sandboxes in 75 seconds.

Key Questions Answered

• •Sandbox Architecture Differentiation: Build agent compute on bare metal with a custom scheduler rather than Firecracker-on-VM to achieve 60ms spin-up times versus competitors taking minutes. Preload snapshots directly on NVMe drives so sandbox creation requires zero network latency. This architecture enables stateful pause-and-resume behavior — agents need the same persistent state that humans expect from closing and reopening a laptop lid.
• •Agent Workload Patterns Require New Infrastructure Planning: Agent workloads split into two distinct types requiring different capacity strategies. Long-running background agents follow human usage patterns — peak at noon, low at midnight, lower on weekends — enabling predictable provisioning. RL and eval workloads are violently spiky, jumping from zero to 100,000 concurrent sandboxes and back, requiring committed capacity reservations to avoid 75-second queue delays that stall expensive GPU utilization.
• •Computer Use Unlocks a $10 Trillion Knowledge Work TAM: Approximately 1 billion knowledge workers globally generate roughly $50 trillion in annual salaries, with 56% concentrated in healthcare, government, and financial services. Most of that work runs inside legacy Windows applications that will not be rewritten. If agents automate 40% of that work — compared to RPA's current 25% — the addressable compute market for Windows sandboxes alone reaches $10 trillion annually.
• •Validate Product-Market Fit by Tracking Follow-Up Urgency: When Daytona rebuilt its sandbox product over two weeks in January 2024, every 15-minute demo call extended to 25-30 minutes. More critically, every prospect who did not receive an API key within 24 hours called back demanding access. Use unsolicited follow-up speed as a PMF signal — if customers contact you before you contact them, the product addresses a real unmet need.
• •SaaS Revenue Reacceleration via Token Reselling Is Misleading: Companies reporting revenue growth by reselling AI tokens carry fundamentally worse margins than traditional SaaS seat-based revenue. The correct model is to expose all product data via open APIs and charge for agent consumption rather than human seats. Salesforce recently published that every product in its portfolio is now API-accessible — this consumption-based approach generates genuine revenue acceleration rather than margin-diluted token pass-through.
• •Responsive Support as a Competitive Moat at 25 Employees: Daytona operates roughly 1,000 Slack Connect channels and joins customer huddles within five minutes of a request. With 13 of 25 employees having worked together for seven-plus years, the team operates at high trust and high throughput without onboarding overhead. In enterprise sales, this responsiveness compresses vendor procurement cycles from the typical two-to-three months down to five days, bypassing the "too small to trust" objection that kills early-stage deals.