Key Moments
AI Agents Need Computers: 74% MoM Growth, 850K/Day Runs, & New Agent Cloud — Ivan Burazin, Daytona
Latent Space PodcastWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AI agents need dedicated, stateful 'computers,' not just code execution boxes. Daytona offers these composable environments, seeing 74% MoM growth, but faces challenges in scaling for spiky AI workloads.
Key Insights
Daytona has experienced 74% month-to-month growth, indicating a massive surge in demand for AI agent compute infrastructure.
The company pivoted to sandboxes for AI agents after realizing that existing infrastructure like EC2 and VMs were insufficient for agentic needs.
Daytona's architecture runs on bare metal with its own scheduler, enabling extremely fast startup times (60ms for one, 75 seconds for 50,000 concurrently) and stateful persistence, differentiating it from VM-based or Firecracker solutions.
A significant portion of Daytona's customer base is now in the 'RLs and eval' category, characterized by highly spiky and unpredictable compute usage, leading to a low average utilization (15%) despite peak demand.
The market for agentic compute is expanding beyond traditional developers to include knowledge workers and legacy application automation, with Daytona developing Windows sandboxes to address this 'human emulator' need, projected to unlock trillions in value.
Daytona's go-to-market strategy is shifting from direct-to-developer (like Code Anywhere) to B2B/B2B TOC, akin to Twilio and Stripe, focusing on providing a foundational compute layer for app-layer agents.
The fundamental shift from developer environments to agentic computers
Ivan Burazin, CEO of Daytona, discusses the evolution of compute infrastructure for software development, tracing his journey from founding Code Anywhere, an early browser-based IDE, to Daytona's current focus on providing 'composable computers' for AI agents. He emphasizes a core belief that development, and now agentic work, should not be tied to local machines. The initial inspiration for moving beyond localhost came from early browser-based IDEs, but the significant pivot happened in January 2024. This pivot was driven by the realization that AI agents have fundamentally different computational needs than human developers. While early offerings focused on automating dev environments for humans, the proliferation of AI agents revealed a latent market demanding more than simple code execution boxes. Daytona's insight was that agents require stateful sandboxes, instant startup, dynamic resource allocation, and infrastructure capable of handling extreme fluctuations in demand, from zero to tens of thousands of CPUs.
Daytona's architecture: Speed, state, and bare metal advantage
Daytona differentiates itself by offering bare metal machines with its own scheduler, a significant departure from common approaches that run sandboxes on top of VMs or use solutions like Firecracker. This architecture allows for exceptionally fast startup times – 60 milliseconds for a single sandbox and 75 seconds to spin up 50,000 concurrently. This speed, combined with the ability to maintain state (akin to closing and reopening a laptop), is crucial for agentic workflows. Unlike many ephemeral sandboxes, Daytona's environments are designed to be long-running and stateful, allowing agents to pause and resume tasks. The bare metal approach eliminates network latency between the sandbox and the underlying disk, CPU, and RAM, leading to incredibly fast I/O operations. Snapshots and templates are pre-loaded on these machines, enabling near-instantaneous provisioning when a sandbox is requested.
Explosive growth and the challenge of spiky workloads
Daytona has seen remarkable, organic growth, reporting 74% month-to-month increases in usage. This surge is attributed to the burgeoning agent economy, where each future agent requires a dedicated compute environment. The company's primary customer segments are split into 'background agents' or long-running agents (like those from OpenHands, Cognition, or LlamaIndex) and 'RLs and eval' workloads. The latter, often used for research and training, exhibits highly spiky usage patterns – demand can jump from near zero to 100,000 CPUs almost instantly. This unpredictability leads to a low average utilization rate for Daytona, around 15%, despite achieving up to 90% during peak bursts. This challenge is common across the agent infrastructure market, with companies like Neon and Parallel facing similar issues in managing and provisioning for these unpredictable compute demands.
The 'human emulator' and the massive opportunity in legacy applications
A significant new focus for Daytona is the 'human emulator' concept, addressing the vast market of knowledge workers still reliant on legacy applications, particularly within Windows environments. While developer tools often operate headlessly via CLI, much of the global economy—estimated at $50 trillion in knowledge worker salaries—is locked into older systems. Daytona aims to unlock this by providing Windows sandboxes that can interact with these legacy apps, similar to how an intern would. This is a massive market, potentially worth trillions annually, and Daytona's new Windows sandbox offering, with dramatically reduced spin-up times compared to traditional cloud VMs (seconds instead of minutes), is designed to tap into this. The difficulty in hosting Mac OS sandboxes due to licensing and security restrictions highlights the unique value proposition of Daytona's cross-OS capabilities.
Competing with Kubernetes and the appeal of Daytona's ergonomics
Daytona's core offering competes directly with managed Kubernetes services like EKS and GKE. However, customers often report a preference for Daytona due to its superior 'ergonomics' and ease of use. Unlike the complex interfaces of Kubernetes, Daytona functions more like platforms such as Twilio or Stripe, offering a simple API and SDK for seamless integration. This user-friendly experience, combined with its speed and scalability, makes it a compelling alternative. Furthermore, Daytona's sandboxes are difficult to exhaust due to on-the-fly resizing capabilities, a feature rarely found elsewhere. The platform also supports advanced configurations like Docker-in-Docker, enabling complex workflows such as Docker Compose or K3S within a sandbox, which is crucial for certain agentic tasks.
Open source strategy and its impact on adoption
Daytona adopts an open-core strategy, with its core scheduler and underlying infrastructure released under a permissive license (AGPLv3 for the sandbox product). While purists may debate its 'true' open-source status, this approach allows companies to use and inspect the code, while proprietary layers handle advanced features like GPU support or Windows environments. This openness has been beneficial, fostering community engagement and providing context for agents integrating with Daytona repositories. However, its primary impact seems to be on the consumption of Daytona's cloud product rather than driving self-hosted deployments. The company finds that while open source helps overcome procurement hurdles for large enterprises, the true growth driver is the ease of integration and rapid deployment facilitated by the cloud offering.
The 'insane responsiveness' of Daytona's team as a differentiator
Beyond features and performance, the number one factor that sells Daytona to customers, particularly for 'day two' rather than 'day one' decisions, is the 'insane responsiveness' of its team. Despite being a small company of 25 people, Daytona provides a level of support—primarily through Slack and email—that is highly unusual. This is enabled by a team composed of serial founders and individuals with long-standing working relationships (average of 7+ years), fostering high trust and a strong sense of shared mission. The culture emphasizes constant availability and rapid problem-solving, with team members often online and responsive 24/7. This dedication, while personally demanding, creates a significant competitive advantage, particularly for enterprise clients who value reliability and immediate support.
Future outlook: The AI cloud and infinite compute potential
The compute market for AI agents is experiencing exponential growth, with the infrastructure segment growing at over 40% month-over-month. Daytona anticipates that owning CPUs might become a go-to-market tactic, as current growth is often hindered by GPU availability. The long-term vision sees a specialized 'AI cloud' emerging, analogous to how Stripe revolutionized payments or Cloudflare revolutionized networking. This cloud would be built specifically for agents, offering specialized sandboxes, web search capabilities, and databases. Burazin believes there are still new infrastructure primitives for agents to be discovered. While the market is currently experiencing rapid expansion, the potential for compute demand is nearly infinite, given that every agent for every task will require a dedicated computer, mirroring the size of the PC market and growing beyond it.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
Daytona provides composable computers for AI agents, essentially offering production-grade sandbox environments. They offer these environments via an API, enabling agents to perform various tasks efficiently.
Topics
Mentioned in this video
CEO of Daytona, discussing the company's evolution from Code Anywhere to its current focus on sandboxes for AI agents. He shares insights on market growth, technical challenges, and go-to-market strategies.
Founder of XAI, mentioned for his public statements about the need for computers for AI agents, aligning with Daytona's core offering.
Co-founder of Andreessen Horowitz, cited as being the first to advocate for exposing all Salesforce products via API, a move praised by the speaker.
A company providing composable computers for AI agents, evolving from a developer environment automation tool. They are experiencing significant month-to-month growth and are focused on delivering production-grade sandbox environments.
Mentioned as having a very short-lived browser-based IDE before becoming the well-known platform it is today. It was noted in the context of early browser-based IDEs.
A browser-based IDE that emerged shortly after Code Anywhere, highlighting the early market for such tools.
A platform for creating and sharing code, which emerged when Code Anywhere was winding down and has since become successful.
A popular source-code editor mentioned as not existing during the early days of browser-based IDEs like CLoud9 and Code Anywhere.
Container orchestration system mentioned as being in its early stages or not yet public when Code Anywhere was being developed, and later as not sufficient for agent compute needs.
Containerization platform whose public release likely coincided with the development of Code Anywhere and is used as an isolation layer within Daytona.
An early AI agent that Ivan Burazin had access to, which contributed to the insight of pivoting Daytona to agent compute sandboxes.
Amazon Web Services' Elastic Compute Cloud, used as a comparison point for Daytona's sandbox offering, particularly its spin-up time and production-grade nature.
Virtual Machines, a common infrastructure for running applications, contrasted with Daytona's bare metal approach for agent sandboxes.
A virtualization technology often used by cloud providers to run microVMs, mentioned as a common approach for sandboxes that Daytona aims to surpass.
A workload orchestrator, mentioned as not being sufficient for the specific needs of agent sandboxes compared to Daytona's custom scheduler.
Amazon Elastic Kubernetes Service, a managed Kubernetes service that Daytona competes against, offering a simpler and faster alternative for agent sandboxes.
Amazon Web Services, a major cloud provider, contrasted with Daytona's specialized offering for AI agents.
A security technology used to harden Docker containers, making them equivalent to VMs in terms of security, and is the default isolation method for Daytona.
A lightweight Kubernetes distribution that can be spun up inside Daytona sandboxes, enabling more complex workloads.
Microsoft's automation platform, mentioned as their entry into the agent-computer space, similar to Daytona's offering.
Accounting software, mentioned as a source of siloed data that agents need to access programmatically.
A CI/CD platform that is experiencing high demand, with one company processing a thousand PRs a day, highlighting the bottleneck in the CI process.
A lightweight, file-based SQL database, mentioned as an example of a database that could be part of a specialized cloud for AI agents.
A column-oriented database management system, mentioned as a source of siloed data that agents need to access programmatically.
Amazon Simple Storage Service, a cloud object storage service, mentioned as a place where one customer dumped entire codebases as JSON files for versioning.
The first browser-based IDE co-founded by Ivan Burazin, which served as a foundational learning experience for building Daytona. It had around 3 million users and was angel-backed.
One of several early competitors in the browser-based IDE space that existed when Code Anywhere was active.
A startup accelerator, mentioned in the context of Daytona's customer base, which includes YC startups.
A competitor in the sandbox space that runs on S3, representing a different approach to sandbox architecture compared to Daytona.
A cloud communications platform, used as an analogy for Daytona's business model as an API-first, developer-centric service provider.
A financial services and technology company, frequently used as an analogy for Daytona's API-driven, developer-friendly approach.
Mentioned as a user of Daytona, likely an AI or tech company, highlighting the adoption of their sandbox technology.
Elon Musk's AI company, mentioned in the context of the 'human emulator' concept and the need for agents to have computer access.
A software tool used by the host to create a virtual sandbox for an AI agent to perform a task, illustrating the need for computer-like access for agents.
A major enterprise software company, cited for its move to expose all products via API, which the speaker sees as a positive development for agent-based workflows and consumption-based pricing.
An AI safety and research company, mentioned as a provider of tokens that SaaS vendors are reselling, a business model the speaker critiques.
A sneaker company that pivoted to GPUs, used as an example of companies making speculative moves into AI hardware.
A leading technology company in GPUs and CPUs, mentioned as a potential bottleneck for growth in the AI compute market due to supply constraints.
A semiconductor company, mentioned as a future supplier of CPUs that will be critical for the growing AI compute market.
A leading AI research company, mentioned for its promotion of the 'AI Cloud' concept, which Daytona is also aiming to be a part of.
A content delivery network and security company, discussed for its approach to managing spiky workloads and its agent portfolio.
A database company mentioned as facing similar infrastructure challenges with spiky workloads, and as a user of Daytona's technology.
A company mentioned in the context of facing spiky workload challenges in the agent infrastructure space.
A web-based platform for version control using Git, criticized for being an overhead for inner-loop development and agentic workflows.
A workflow orchestration platform, mentioned as a past experience that informs the speaker's view on using open source for GTM strategy with large companies.
An open-source product analytics platform, mentioned as a source of siloed data that agents need to access programmatically.
A global data center company, mentioned as a provider that companies like Railway still use, indicating the continuing importance of colocation facilities.
A cloud platform for frontend developers, mentioned for promoting the 'AI Cloud' concept alongside OpenAI and Daytona.
More from Latent Space
View all 216 summaries
90 minThe Agent-Native Cloud: 3M Users, 100K Signups/Wk, Data Centers, & Death PRs — Jake Cooper, Railway
120 minThe Next War Is Already Here — Yaroslav Azhnyuk, The Fourth Law & Noah Smith, Noahpinion
67 minInside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa
23 minSenior Dev: This "Grill Me" Prompt Is Going Viral Among Top Engineers
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free