Giving Agents Computers — Ivan Burazin, Daytona

Ivan Burazin discusses his long-standing interest in eliminating local development environments, traced back to his work with CodeAnywhere and the 'end of localhost' thesis. He reflects on early innovations like the first browser-based IDE, predating Heroku and other cloud IDEs, emphasizing the foundational importance of building custom stacks in the early 2000s. Progressing from developer tools to conferences, Ivan highlights how these experiences led to the pivot toward Daytona, a platform emphasizing composable, fast, and stateful computer environments for AI agents. He elaborates on how market needs, customer demand, and technical innovations like bare metal, custom schedulers, and NVMe storage have shaped Daytona’s fast, scalable sandbox infrastructure.

Ivan details Daytona’s architecture—running on bare metal with its own scheduler, utilizing underlying disk, CPU, and RAM, enabling extremely fast startup times (around 60 ms) and high concurrency (up to 50,000+ sandboxes). Unlike typical cloud providers using VMs or containers with networked storage like EBS, Daytona's approach minimizes latency and maximizes performance, crucial for AI workloads like RL and evals, which are highly spiky and unpredictable. He discusses the importance of having dedicated, stateful, and instant-resizable sandboxes, and how these design choices surpass Kubernetes and other orchestration solutions for agent workloads.

Ivan projects exponential growth in the entire compute infrastructure market, driven by AI adoption, with growth rates around 40% monthly. He envisions a future where AI-specific clouds resemble Stripe more than AWS—focused on APIs, composability, and efficiency—offering specialized sandboxes, search, databases, and tooling optimized for AI agents. The talk explores the limitations of current hardware supply (CPUs, GPUs, networking) and discusses how owning infrastructure with proprietary chips or data centers might become a strategic move for scaling AI workloads.

Ivan emphasizes the value of open source in Daytona's growth, sharing how their previous open-source projects fostered community trust and eased enterprise procurement. He explains that while open source doesn’t significantly boost growth directly, it helps large companies adopt Daytona faster due to transparency and lower trust barriers. He discusses how Daytona's responsiveness and support—often via Slack or calls—are key competitive advantages, enabling rapid onboarding and support for very large clients, including Fortune 500 firms. The company’s culture relies heavily on long-time trusted relationships, high responsiveness, and a work environment that accepts the pain of founder life.

The conversation covers the broad landscape of AI infrastructure, highlighting the importance of proprietary hardware (CPUs, GPUs) and custom primitives for supporting agent workloads at scale. Ivan comments on current bottlenecks like CPU supply, the potential of specialized clouds, and the importance of exposing APIs rather than siloed tools. He debates current trends such as token resale and API usage as sources of hype rather than sustained revenue, warning against overestimating their impact. The idea of a new, agent-first cloud infrastructure—similar to Stripe or a specialized AWS—is proposed as the next big evolution. Finally, Ivan shares optimistic yet cautious predictions about the explosive growth and the strategic moves needed to dominate the future AI cloud market.