The idea for Modal's sandbox product came up over lunch. Erik Bernhardsson, Modal's chief executive, and Akshat Bubna, his cofounder and chief technology officer, noticed that many of the people they talked to were using large language models to write code. "Maybe they need a place to run that code," Bernhardsson said, recalling their thinking.
Modal sells cloud computing to AI companies. It already had much of what such a product needed: systems for managing containers and for keeping them isolated from one another. Containers are packaged, self-contained software environments. Bubna built a prototype and launched it in "basically one day or two days," Bernhardsson said on the Latent Space podcast. The episode was published October 22, 2026, eight days after Modal announced a $750 million funding round.
In Modal's terms, a sandbox is an isolated environment for untrusted or model-generated code. The company's documentation describes the cycle: an application creates a sandbox, runs commands inside it, reads the output and then shuts it down. The point of the isolation is to keep that code walled off from the application that launched it and from other workloads.
A slow start, then demand from agents
The product launched in 2023 and drew little interest. Bernhardsson said some early users liked it, but for a while it was "a pretty insignificant source of revenue." At one point the team discussed whether it needed the product at all and whether to shut it down.
In his account, the models changed the picture. They went from refactoring functions and writing snippets to generating full applications. Growth came from vibe-coding companies such as Lovable, whose users build software by describing what they want and letting AI write it. Ramp started using Modal for background agents, which are AI programs that carry out tasks without a person watching each step. Then came what Bernhardsson called "tremendous growth" from a number of newer AI labs doing reinforcement learning on Modal sandboxes. In reinforcement learning, a model improves by attempting tasks and being rewarded for success. When the task is writing code, each attempt needs somewhere to run.
Bernhardsson said sandboxes now account for about a third of Modal's revenue, and the company considers them its second product after Functions, its service for running code on demand. Modal's Series D announcement on October 14, 2026 put the sandbox business alone at a $300 million annualized run rate, which projects current revenue over a full year. The company's total run rate was above $1 billion. The announcement also reported close to three billion sandbox starts. That number counts launches over time, not sandboxes running at once. Modal reported all of these figures itself.
"The truth is, like, I didn't"
Late in the conversation, one of Bernhardsson's tweets was read out on the show: "Every month I keep saying our sandbox revenue might have plateaued, but it keeps going up 30% every month."
"It's kind of nuts," Bernhardsson said. Investors ask him how he knew sandboxes would be big. "The truth is, like, I didn't," he said. He gave the team some credit for seeing the product's potential back in 2023 and said he had expected it to grow beyond what it was then. But he said he would never have bet on sandbox revenue becoming this large, and he called its share of the business "insane." He said Modal now has a very large team working on the product.
What Modal had to build
When asked what Modal had to change to offer something existing sandboxes did not, Bernhardsson started with "Not a whole lot." He described his team as "infra geeks" who are good at scale and reliability, running "hundreds of thousands or even millions of sandboxes." In his opinion, that explains why Modal has "by far the largest market share in this space," though he said he does not have numbers to show it. He also conceded that sandboxes are not a very differentiated market.
The scale took work. Bernhardsson said Modal has had to rewrite its systems "a bunch of times." In a July 2026 engineering post, Modal engineers described replacing a centrally coordinated system with schedulers that can be added side by side. They also described a demonstration in which one million sandboxes were created in under a minute and ran at the same time. The post separates two moments. A creation request returns once a machine has accepted the sandbox and begun starting it, and that happens before the sandbox can run code. The median time from a client's request to the user's code running was under half a second. Some starts took longer when many sandboxes launched at once.
Compatibility was another change. Modal originally built sandboxes on gVisor, an isolation layer that sits between a program and the host machine's operating system. Bernhardsson said some uses need full Linux compatibility, so Modal now also offers complete virtual machines. Its VM Sandboxes announcement on October 1, 2026 lists workloads such as Docker, databases and graphical environments. It also says gVisor remains the default for workloads that already run well on it. Modal has also built better tools for piping data between sandboxes and for working with file systems, and Bernhardsson said the team has spent a lot of time on snapshotting.
He expects future differentiation to come from forking sandboxes, snapshotting and resuming them, and moving them live between machines. He also wants to separate the agent's harness from the sandboxes where its actions run. The harness is the loop of software that turns a model's output into actions. With that separation, one agent could control many sandboxes.
Anthropic's bring-your-own-sandbox deal
A version of that separation already exists. Bernhardsson said Anthropic's Claude Managed Agents has a "bring-your-own-sandbox integration," so customers can run their agents' work on Modal. In Modal's May 19, 2026 announcement, Anthropic hosts the agent loop and the session state. The agent's tool calls run in a Modal sandbox that the customer configures, including its software image, resources and network access.
Bernhardsson thinks Anthropic's motive is commercial. Anthropic makes its money from inference, the running of its models, so he thinks it wants an ecosystem with "the largest possible surface area, as long as you use Anthropic for inference." He said partnering with a sandbox provider makes "very rational sense" for the company.
From dev boxes to sandboxes that last seconds
Bernhardsson thinks most sandbox use today is "lift and shift." People take the development machine they would use themselves, move it into the cloud, let an agent control it and perhaps run the harness inside it too. He doubts that is the right long-term model. Another of his tweets, also read on the show, predicted that sandboxes would evolve "from dev box in the cloud to something more ephemeral," with "one environment for every command" and snapshots in between so they can be paused, resumed and cloned. "I'm fully convinced," he said.
In his picture, long-running agents behave like durable workflows, programs that run around the clock and only occasionally need to do something. When an agent needs to run a command, it would start a sandbox in around 100 milliseconds, run a few commands, snapshot the result and shut the sandbox down. He argued that the faster this wake-and-sleep cycle gets, the more efficient compute becomes. It would also allow forking. He asked listeners to imagine a coding agent like Claude Code branching out and running hundreds of agents at the same time, where today it works through a task in sequence.
His clearest example is continuous integration, or CI, the automated testing that runs whenever a programmer changes code. He said the process in tools like GitHub Actions is sequential: install a stack of software, run the test suite, generate reports. Teams can split tests into shards, but each shard has to repeat all the installs and setup, which can take minutes. He would like to set up once, snapshot the container, clone it, and run every single test in its own sandbox, "like, a thousand sandboxes."
He said the obstacle is the overhead of starting and stopping sandboxes, which he called "the transaction cost." Until that cost falls, running work in parallel cannot deliver its benefits. Modal's snapshot documentation already describes a set-up-once, fork-many pattern for files: install dependencies, capture the sandbox's file system, then start several independent sandboxes from it. Saving running processes is less mature. The same documentation labels memory snapshots an alpha feature with limits. They do not support GPUs, cannot be taken while a command is running, end the captured sandbox and expire after seven days.
Bernhardsson said the median sandbox on Modal already lasts about eight seconds, "which is kind of incredible when you think about it," and he thinks it could get shorter. He said Modal still wants to support long-running sandboxes because they have valid uses. His goal is to cut the cost of starting and stopping sandboxes so that work can be massively parallelized.