How Daytona Is Redefining AI Agent Compute Sandbox Infrastructure

When AI Agents move from concept to large-scale deployment, a critical yet overlooked question emerges: what kind of computing environment do these agents actually need? In an in-depth interview, Daytona CEO Ivan Burazin revealed the dramatic shifts happening in the AI Agent infrastructure market—74% month-over-month growth, 850,000 sandboxes running daily, and an entirely new market that could be larger than the human PC market.

From Browser IDE to AI Sandbox: A Critical Strategic Pivot

Daytona's story begins in the early 2000s. Ivan and his co-founder started with infrastructure services like server virtualization and routers, later creating Code Anywhere—one of the world's earliest browser-based IDEs. Back then, there was no VS Code, no Kubernetes, and Docker hadn't been publicly released. The entire underlying tech stack had to be built from scratch.

This experience became the source of Daytona's core competitive advantage. Three years ago, the team began repackaging Code Anywhere's underlying technology to provide automated development environments for human engineers. But the real turning point came at the end of 2024.

At that time, Ivan gained early access to Devin (before its public launch) through his network. The team tried combining the open-source version OpenDevin (now renamed OpenHands) with their own infrastructure to launch a SaaS service. While direct users were few, numerous teams building AI Agents proactively reached out: "My Agent needs a compute sandbox runtime." This signal made the team realize that an entirely new market was forming.

Why Agents Can't Use Human Infrastructure

Initially, many questioned: "Why do Agents need specialized infrastructure? Aren't EC2 and VMs enough?" Daytona's team made this mistake early on—they gave Agents development environments designed for humans, and all 20-30 early users reported "this isn't what we need."

After deep research, the team identified three core requirements for Agent computing environments:

Extreme Speed: Agents need millisecond-level environment startup times. Daytona currently achieves 60-millisecond startup per sandbox (including network latency), and launching 50,000 sandboxes simultaneously takes about 75 seconds, while some competitors require 2,000 seconds (approximately 30 minutes).

Stateful Persistence: Just as humans don't want their laptops shut down mid-work, Agents need the ability to "close the lid and resume later." Sandboxes must support pause and resume while maintaining complete state.

Dynamic Composability: Different tasks require different combinations of compute resources. Agents may need to dynamically adjust CPU, memory, disk, or even add GPUs. Daytona supports runtime dynamic scaling, virtually eliminating OOM (Out of Memory) issues.

About OOM (Out of Memory): OOM refers to errors triggered when a process requests more memory than the system has available. The Linux kernel's OOM Killer mechanism forcefully terminates processes when memory is exhausted—catastrophic for long-running AI Agent tasks, potentially causing hours of work to be lost instantly. Traditional container orchestration systems (like Kubernetes) prevent OOM through preset Resource Limits, but this requires accurately estimating memory needs in advance. AI Agent memory consumption is often unpredictable (depending on task complexity, context length, and other dynamic factors). Daytona's dynamic scaling capability means sandboxes can request more memory on demand at runtime rather than being constrained by static quotas—critical for Agents handling large codebases or complex reasoning tasks.

In Ivan's words, Daytona is essentially "a combination of Lambda and EC2"—with Lambda's speed and EC2's persistence.

Bare Metal Architecture: The Technical Foundation for Speed

Unlike most competitors running Firecracker on VMs, Daytona chose a bare metal architecture from the start. The team built a custom scheduler that directly uses the underlying machine's disk, CPU, and memory.

The Fundamental Difference Between Bare Metal and Virtualized Architectures: A bare metal server means users directly access physical server hardware without a virtualization layer. In contrast, traditional cloud computing (like AWS EC2) runs a Hypervisor on physical machines, then creates virtual machines on top—each layer of abstraction introduces latency and performance overhead. Firecracker is AWS's open-source lightweight Virtual Machine Monitor (VMM), designed for serverless computing. It can launch micro-VMs in 125 milliseconds and is widely used in AWS Lambda and Fargate—already industry-leading, but still an order of magnitude slower than bare metal solutions. NVMe (Non-Volatile Memory Express) is a high-speed storage protocol designed specifically for SSDs, offering tens of times more IOPS compared to traditional SATA interfaces. EBS (Elastic Block Store) is AWS's network-attached storage—flexible but inherently introducing millisecond-level latency due to network transmission. Daytona's choice of bare metal + local NVMe is precisely aimed at compressing these intermediate layer latencies to the absolute minimum.

This architecture delivers several key advantages:

• Extremely fast IOPS: No latency from network storage like EBS; all I/O operations complete on local NVMe drives
• Snapshot pre-loading: Templates and snapshots are pre-loaded on bare metal machines; launching a sandbox directly routes to the corresponding physical machine
• Zero network latency: All operations within the sandbox complete locally

Interestingly, when the third co-founder saw this architecture, his reaction was: "Isn't this 2008 technology?" The CTO replied: "Exactly." Sometimes the most advanced solutions come from returning to fundamental principles.

For isolation, Daytona defaults to Docker containers hardened with Sysbox, achieving VM-equivalent security while maintaining container lightweight characteristics. Sysbox is a container runtime that gives Docker containers near-VM isolation levels, supporting system-level software like systemd and Docker within containers while preserving container lightness. It also supports Docker-in-Docker, allowing users to run Docker Compose or even K3s clusters inside sandboxes, significantly expanding the types of supported workloads.

Two Fundamentally Different Workload Patterns

Daytona currently serves two major workload categories with drastically different usage patterns:

Long-Running Background Agents

Clients in this category include Cognition (Devin), Lovable, Harvey, and others. Their usage patterns resemble humans—"following the sun," peaking at noon, dropping at midnight, and declining on weekends. This workload is relatively predictable and can be optimized through cross-region scheduling.

RL Training and Evaluation

This is the fastest-growing workload type. A few months ago, RL workloads were at zero; by the time of the interview, they accounted for nearly 50%. These workloads are characterized by extreme "spikiness"—either nothing at all, or suddenly needing 100,000 CPUs, then dropping back to zero.

Computational Characteristics of Reinforcement Learning Training: Reinforcement Learning (RL) is a branch of machine learning where agents optimize strategies through environmental interaction and reward signals. In the large language model domain, RLHF (Reinforcement Learning from Human Feedback) and newer algorithms like GRPO and PPO are used to align model behavior. RL training has extremely unique computational characteristics: the training phase requires simultaneously running thousands to tens of thousands of parallel environments (rollout workers) to collect experience data. These environments are immediately destroyed after a training batch ends, creating an extreme "all-or-nothing" resource demand pattern. This is fundamentally different from the smooth traffic curves of web services, posing unprecedented challenges to infrastructure elasticity—traditional cloud platforms' auto-scaling often requires several minutes, far from meeting RL training's second-level demands.

This spike pattern presents unprecedented challenges for infrastructure. Daytona's current average utilization is only 15%, but peaks can reach 90%. At his own Compute Conference, Ivan discovered this is a shared challenge across the entire Agent infrastructure industry—whether it's Neon (databases), Parallel (search), or other providers, all face the same dilemma.

Computer Use: The Next Trillion-Dollar Market

Ivan believes Computer Use may be the most important "killer application" for Agents.

How Computer Use Works: Computer Use refers to AI models' ability to perceive screen content through screenshots and simulate mouse/keyboard operations to control computer interfaces. Anthropic pioneered this capability in Claude 3.5 Sonnet in October 2024, marking a major leap from pure text interaction to graphical interface control. Unlike RPA (Robotic Process Automation), Computer Use doesn't rely on predefined UI element selectors but dynamically adapts to interface changes through visual understanding—theoretically capable of operating any software with a screen. While headless approaches (headless browsers/API calls) are more efficient, they're limited by the closed nature of enterprise legacy systems—vast amounts of critical business data remain locked in desktop applications without APIs, making Computer Use the only viable automation path.

He presents some staggering numbers:

• Approximately 100 million knowledge workers in the US, about 1 billion globally
• US knowledge worker total compensation is approximately $10 trillion, globally about $50 trillion
• 56% is concentrated in five major industries: healthcare, government, finance, and two others
• Massive amounts of work remain locked in legacy applications running on Windows

Even using the conservative RPA market estimate of 25% automation rate, if Agents can achieve 40% automation, this represents a $10 trillion market.

Ivan illustrated this with his own experience: while preparing a board report, despite using Claude Code connected to modern tools like ClickHouse, PostHog, and QuickBooks, vast amounts of data couldn't be accessed via APIs. Ultimately, he deployed OpenClaw in a Mac Mini virtual sandbox, created read-only accounts for various company systems, and let the Agent directly log into websites, export data, and complete the report.

"If even a startup like ours, using the latest tools, still needs a Computer Use Agent, how could Goldman Sachs possibly solve their problems with a purely headless approach?"

Daytona currently supports Windows sandboxes (second-level startup, with snapshot and fork support). Mac OS sandboxes are also in development but constrained by Apple's licensing agreement—each machine only allows 2 parallel VMs, and user authorization can only be changed once every 24 hours.

Market Landscape and Growth Logic

Unexpected Global Distribution

Daytona's user distribution reveals some interesting characteristics: by user count, Singapore is the top city (not Silicon Valley), and Tokyo is unusually active—something never seen in previous technology cycles. Brazil continues its traditionally strong position in the developer ecosystem.

A Pragmatic Open Source Strategy

Daytona uses the AGPL-3.0 license—a pragmatic choice between fully open source and commercial licensing.

Strategic Implications of the AGPL License: AGPL-3.0 (GNU Affero General Public License) is the network service extension of GPL. Its core clause requires: any modified version providing services over a network must make its source code available to users. This makes it a "natural barrier" against cloud providers—if AWS, Google Cloud, or other major cloud vendors want to sell an AGPL project as a managed service, they must open all their modifications, which is commercially almost unacceptable. Thus, AGPL maintains open-source community transparency and contribution incentives while preventing large cloud vendors from "free-riding." Notable open-source projects like MongoDB and Grafana have adopted similar strategies. For Daytona, the open-source codebase has a special value in the AI era: users can provide the code repository directly to AI Agents, giving them complete integration context and lowering the barrier to adoption.

Ivan admits that open source helps cloud product consumption growth more than converting self-hosted users—because users can send the code repository to an Agent, giving it more integration context.

Extreme Responsiveness from a 25-Person Team

Daytona currently has only 25 people but maintains approximately 1,000 Slack Connect channels. In customer case studies, "extreme responsiveness" is repeatedly cited as the primary reason for choosing Daytona. Thirteen team members have over 7 years of working history with the founders—this high-trust, high-synergy team culture is the foundation of their service quality.

A Sobering Perspective on the SaaS Industry

Ivan raises a somewhat controversial viewpoint: the market is incorrectly assigning premiums to SaaS companies that are essentially "reselling tokens." He argues that so-called "revenue re-acceleration" is fundamentally low-margin token resale revenue, completely different from traditional SaaS's high-margin subscription revenue.

His recommendation: SaaS companies should expose all product features through APIs and charge by consumption, rather than adding an Agent layer in the middle to resell tokens. Salesforce's Marc Benioff recently announced opening all products via APIs—Ivan believes this is the right direction.

"Data is still siloed. You give me an Agent, and the data is still siloed. Better to just expose the data directly to my Agent."

Future Outlook: A Dedicated Cloud for Agents

The entire Agent infrastructure market is growing at over 40% per month. Ivan believes a cloud platform purpose-built for Agents will emerge—containing sandboxes, web search, databases, and all infrastructure primitives Agents need. Moreover, "we haven't seen all the infrastructure primitives yet; more are being invented."

From a CPU supply perspective, this market could even face shortage problems similar to GPUs. Semi Analysis analyst Dylan Patel pointed out at the Compute Conference that CPUs are becoming the new bottleneck—when every Agent needs a "computer," the PC market scale for 8 billion people may be just the starting point for the Agent compute market. This judgment has deep logical foundations: the GPU shortage is essentially driven by AI training's extreme demand for parallel floating-point operations, while large-scale Agent deployment creates massive demand for general-purpose computing (CPU), memory, and storage—a fundamentally different but equally enormous computational pressure.

As Ivan puts it: "There are 8 billion humans, each needing roughly 1-2 computers. But in two years, ten years, a hundred years, how many Agents will be running? Every single task needs a sandbox. This market is essentially infinite."

Key Takeaways