Erik Bernhardsson, cofounder and chief executive of Modal, said the first prototype of the company's sandbox product took "a few days." What came next was also small. He spent some time improving the software development kit, some people started building on it, and once customers showed up the team rewrote it. The heavy engineering came later. "In 2025, when it started exploding, then we had to rewrite the system like five times just to handle like more and more scale and make it faster and, and, and more reliable," he said.
Because the first investment was so small, Bernhardsson described sandboxes as an asymmetric wager, "low downside, high upside." He spoke on Unsupervised Learning, a podcast hosted by Jacob Effron and Patrick Chase, both AI investors at Redpoint Ventures. Chase led Redpoint's investment in Modal. The episode was published October 28, 2026. Over the conversation, Bernhardsson explained how he decides where Modal puts its engineers and money, and why he cannot point to another wager as cheap and promising as sandboxes.
A small bet that became a third of the business
Modal sells cloud computing to AI companies. A sandbox is an isolated computer environment where code written by an AI model, or other code that should not be trusted, can run without reaching the rest of a customer's systems. Effron called sandboxes a "fundamental primitive" for coding agents, the AI programs that write and run software, and pointed to Modal's products around reinforcement learning, where a model improves by trying tasks and being rewarded for success.
Modal introduced sandboxes in 2023. Bernhardsson said that over roughly the past 12 months the product had grown, "I don't know, a thousand X almost or something like that," and that it now brings in about a third of the company's revenue, a $300 million run rate on its own. A run rate takes current revenue and projects it over a full year. Modal's Series D announcement on October 14, 2026 gives the same $300 million figure for sandboxes and a total company run rate above $1 billion. Modal reported these numbers itself.
One of the rewrites is described in a July 2026 engineering post. Modal replaced centrally coordinated scheduling with schedulers that can be added side by side, then ran a demonstration that created one million sandboxes at once in under a minute. The median time to the user's code running was under half a second, though some starts took longer because of contention on machines and networks. The redesigned system was still in beta when the post appeared.
Bernhardsson said he did not see the size of this coming. When Effron asked whether he would have named sandboxes as the answer back in 2023, Bernhardsson said "no," explaining that at the time sandboxes were simply a cheap bet. Later, asked for one thing he had changed his mind on in AI over the past year, he chose the sandbox opportunity: "I didn't think it would be anywhere close."
The next bets cost more
When Effron asked what else Modal is early on today, Chase cut in: "Jacob, are you fishing for the next thing we should be investing in?" "100%," Effron said. "I'm nothing if not shameless."
Bernhardsson said the sandbox success had raised the bar, with everyone asking "what's cooking in the lab." He wished he had an easy answer. "I don't know if there's like anything similar today that we're working on," he said. The ideas he is bullish on "probably represent like way larger investments," with "larger risks." He named three.
The first is low latency, meaning less delay between a request and a response. Bernhardsson expects "a lot of voice to voice, a lot of real time video type stuff." Serving that, he said, means rethinking how Modal manages compute around the world: cutting the number of network hops between a user and the machine running the model, and scaling capacity up and down region by region. He called it "almost like building a CDN for inference." A content delivery network, or CDN, keeps copies of websites on servers near their users so pages load quickly. Inference is the work of running a trained model to produce an answer.
His enthusiasm for voice comes partly from frustration. "Why are like voice assistants so bad in 2026?" he asked. His Google Home bungles every question, he said, so he goes to ChatGPT instead. He wants systems connected "end to end" and trained to go directly from speech to speech.
The second bet is training. Bernhardsson called it "kind of an underrated thing" and sees opportunity in training mid-size models and in reinforcement learning. He said getting the necessary high-speed networking working had been a big bet for the last couple of years. Training a model across several machines requires them to exchange data constantly. Modal made its Clusters product generally available on October 1, 2026. It schedules all the machines a job needs together and supports RDMA, a networking method that lets computers read each other's memory directly without routing data through the processor. Customers named in the announcement include 1X, which pre-trained a world model on these clusters, and Decagon, which fine-tunes open models.
The third bet is virtual machines, or VMs, which are complete simulated computers. "Supporting full blown operating systems like Mac and Windows and Android and iPhone could be kind of cool," Bernhardsson said. So far, Modal has gone part of the way. Its VM Sandboxes launch, also on October 1, offers full Linux virtual machines through the existing sandbox interface. Mac, Windows and mobile systems remain an aspiration.
"I have a lot of those ideas in my head," he said, and Modal is building many of them. But he said he did not know whether any had the same explosive potential as sandboxes. Perhaps, he added, if they did the podcast again in three years they could look back and see.
Building blocks that last five years
Effron asked whether Modal would build more tools for coding agents. Bernhardsson said he thinks about that question as an infrastructure company. Agent harnesses, the software loops that turn a model's output into actions, are "evolving so fast" that he feels the industry is still very early. Modal's job, he said, is to work out which building blocks people will still need "for the next five years," and that is not obvious.
He listed areas Modal finds interesting: coordinating agents, evaluations, model gateways that route requests between models, and the broader cycle of learning from agents and improving them. His guess is that agents will become more cloud-based and look more like workflow engines, software that runs long, multistep processes and keeps track of where they are. He said he finds it hard to say exactly what agentic development will look like in five years, partly because he expects "all software" to become agentic. He dislikes the word for the same reason: "it's using AI to make decisions like, well, shouldn't all software do that?"
Effron offered an example of a pattern that did not last: RAG, short for retrieval-augmented generation. The method, described in a 2020 paper, has a system look up relevant documents and hand them to the model together with the question. Bernhardsson agreed. "If we had made a bet on like RAG like two years ago, like it would have been kind of a bad thing to do," he said. One advantage of building infrastructure, in his view, is that "you can take a longer view." You still have to listen to the noise, he said, but you can ignore much of it and look for the patterns underneath.
Rationing 150 engineers
Bernhardsson said Modal has about 150 engineers and that he struggles with prioritization "every day." Even with a thousand or 10,000 engineers, he said, there would still be more to build than people to build it. "I wish we could build a rocket and go to the moon and whatever," he said. "It just never goes away."
So he tries to be "ruthless" about return on investment. Sometimes that means telling a team: "I know we started on this project. I know you're really excited about it, but like this other thing kind of came up." The project gets paused, with a promise to return to it later. Modal has done this many times, he said, and "it sucks when it happens." He called resource allocation "basically like, you know, the number one thing" a CEO has to be ruthless about.
The mistake of doing things in order
Asked for his biggest mistake in building Modal, Bernhardsson described a habit of working in sequence. A team decides one task is "the blocker" and leaves the next until it is cleared. In a fast-moving market, he said, he learned that you sometimes need to do both at once, "just in the interest of moving fast."
He admitted that this does not come naturally to him. "I'm naturally like not mega aggressive, which is kind of ironic because like we've run 15x in the last 12 months," he said. The Series D announcement likewise reports growth of more than fifteenfold year over year. He said he has had to "retrain my brain" to treat Modal's situation as a massive market opportunity and to push harder on hiring, deploying capital and running work in parallel.
A 20-year project
Chase noted that Bernhardsson has called Modal his 20-year project, five years in, and asked what the later years would look like. Bernhardsson expects that in years 10 to 15 Modal will "look more like a conglomerate," with many products and "probably some like crazy bets going to moon and Mars or whatever." He hopes to hire people to take over the details he is now "too in the weeds" on, such as getting a contract from a GPU vendor.
He wants to keep the job of choosing the bets. High-level resource allocation, he said, "to me is that's the most fun. And I hope to stay involved in that for another 20 years, hopefully maybe longer."