Modal Has Run More Than 3 Billion Sandboxes for AI Agents
Modal CEO Erik Bernhardsson argues that the new capabilities of AI models require new infrastructure, including inference, training and sandboxes that let agents execute code in isolation. He describes those tools as building blocks for AI applications; Modal says it has run more than 3 billion customer sandboxes. The company has raised $750 million at a $15.75 billion post-money valuation to grow faster, launch products and add capacity.
Modal is raising to build across the AI application stack
Modal has raised $750 million in a Series D round at a $15.75 billion post-money valuation. CEO Erik Bernhardsson said the company is raising because demand is coming from many kinds of companies building AI applications. Modal plans to use the capital to grow faster, launch products and add capacity.
Bernhardsson described the shift as a “new era of software”: new AI capabilities, he argued, require new infrastructure building blocks. He named inference, training and sandboxes as parts of that infrastructure. Modal’s own description calls it a cloud computing platform for AI, connecting developers to GPUs from more than 20 providers and supporting inference, training and agent sandboxes.
The demand is visible in the company’s sandbox activity. Modal says it has run more than 3 billion sandboxes for customers. Bernhardsson said the sandbox business alone had reached about $300 million in run-rate revenue, making it one of the building blocks Modal has launched—not the whole business.
Sandboxes let agents run code without escaping their environment
Modal began building sandboxes three years before the interview, after seeing companies explore the use of large language models to generate code. The basic need, Bernhardsson said, was a way to contain and run that code in isolation: agents could execute it without escaping the sandbox. The environment is therefore not just a place to run a program; it is a way to limit what an agent can reach while it runs one.
Modal launched the product in 2023, but Bernhardsson said demand did not grow substantially until about two years later. He pointed to models such as Sonnet 3.5, which could generate full applications. That capability, in his account, produced an “explosion” in demand for code-execution tools that could be coupled with AI agents.
Ed Ludlow raised frontier-model cybersecurity testing as one reason sandbox environments had become more familiar to viewers. He referred to an OpenAI and Hugging Face incident involving tests in which guardrails were deliberately reduced so models’ cyber capabilities could be examined. Bernhardsson said agents’ cybersecurity abilities were exciting, but stressed that sandbox use extends well beyond testing agents that might hack.
He pointed instead to coding models and everyday development work. Cognition is training coding models on Modal, he said, while Ramp uses the platform for internal development. In his account, sandboxes are becoming a foundational component of coding models and other AI applications—not simply a containment tool for security tests.
Inference, reinforcement learning and sandboxes are parts of one platform
Asked how changing token economics affect Modal, Bernhardsson called inference an “incredible business” that had been growing quickly, particularly for open-source models. He acknowledged that many providers operate in the space, but framed inference as one part of a broader platform that also includes reinforcement learning and sandboxes.
Bernhardsson said that as token prices fall, demand for these services increases. He cited Typeface as an example: the company launched a new inference offering for classification on Modal and, he said, reached millions of customers overnight. He described that launch as an example of a new type of inference gaining traction.
His account links falling prices with increased demand, but he did not detail how that demand is divided among inference, reinforcement learning and sandboxes. He presented those capabilities as parts of the same platform rather than as interchangeable services.
Virtual machines could let agents take on longer-running tasks
Ludlow asked about Modal’s VM sandboxes, describing them as a way to give agents access to virtual computers. Bernhardsson said agents are becoming more capable at controlling computers, and that a virtual machine provides a full computer environment. Modal’s initial environment is Linux; he said the company hopes to support Windows and Mac over time.
The intended use is broader than running a generated program. Bernhardsson said a full computer environment could help agents take on longer-running tasks, including personal-assistant work such as booking tickets, as well as coding. He described virtual machines as one part of enabling agents to perform a wider range of tasks.