Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

Modal’s latest numbers are hard to miss: the company says it has crossed $1 billion in annualized run rate, raised a $750 million Series D, and now sits at a $16 billion valuation. It also says it has spun up nearly three billion sandboxes. That is a lot of growth for a company built around a fairly unglamorous idea: making compute behave more like software and less like a warehouse lease.

Erik Bernhardsson’s argument is that the old way of buying GPUs does not fit how AI workloads actually move. A contract gives a customer a fixed block of capacity for a fixed stretch of time. Real demand does not stay in a box. It surges, dips, and shows up at odd hours. Modal’s answer is to pool demand across many customers, secure the hardware underneath, and let each one draw from it only when needed. That works better when the aggregate demand is smoother. It also comes with a real risk on the provider side: if GPU prices fall after capacity has been locked in, the supplier can end up stuck with expensive supply.

The company’s sandbox product almost didn’t survive. Bernhardsson says Modal launched sandboxes after he and cofounder Akshat Bubna realized language models could write code, which meant someone needed a place to run that code. Early usage was modest enough that the team discussed shutting it down. Then coding agents, background agents, and reinforcement learning arrived, and demand changed fast. Today, Bernhardsson says sandboxes account for roughly a third of Modal’s revenue.

That shift matters because agent workflows are not one-and-done. An agent can create an environment, run code, check the result, and try again, over and over. So startup time and restore time start to matter as much as raw throughput. Modal’s serverless GPU system uses a buffer of available GPUs, a custom filesystem, process checkpointing, and GPU memory snapshots to get replicas online faster. In other words: the plumbing is the product.

Bernhardsson’s path to this idea runs through Spotify, where he built a music recommendation system, and through Better.com, where he spent five years as CTO. He says the frustration was familiar long before Modal had a name: the infrastructure for building compute-heavy products was bad. That complaint has aged nicely into a very large business.