Modal Passes $1B ARR and Hits a $16B Valuation on Demand for Sandboxes

Modal, a platform for running AI workloads, has crossed $1 billion in annualized run rate revenue and closed a $750 million Series D at a $16 billion valuation, five months after its Series C at $5 billion. The company also launched nearly three billion sandboxes over the same period. In a long interview with Latent Space, CEO Erik Bernhardsson explained where the growth came from and why demand for compute fits poorly with the contracts used to buy it.

Why GPU capacity is sold in rectangles

A typical capacity contract looks like this: a customer gets a fixed number of GPUs for a fixed term. Plot that capacity over time and you get a rectangle. Real workloads look different: they grow, spike and sag over the course of a day. Buy for the peak and you pay for idle hardware. Buy less and you lose demand.

Bernhardsson describes Modal's bet as duration transformation. The company secures capacity at the bottom with long-term contracts, pools demand from thousands of customers and hands each one as much as it needs, when it needs it. The smoother the aggregate demand, the more useful such a pool is. This complements the approach of SF Compute and Evan Conrad: they make capacity tradable, while Modal makes it usable in small, elastic portions.

The model has a downside. If a provider locks in capacity at today's prices and GPUs later get cheaper, it is left with expensive supply on its balance sheet. Bernhardsson does not expect a price collapse soon, but he thinks a business model has to account for that scenario.

The sandbox product that almost got shut down

Sandboxes appeared after Bernhardsson and co-founder Akshat Bubna noticed that language models write code, and someone has to execute it. At first, interest was so modest that the team discussed shutting the product down. Then coding agents, background agents and reinforcement learning arrived. Sandboxes now bring in roughly a third of Modal's revenue. Bernhardsson did not predict that outcome at launch.

The workload has changed too. An agent creates an environment, runs code, looks at the result and tries again, many times over. Startup and environment-restore latency becomes part of the user experience. Hence the shift from developer experience to agent experience.

Elasticity only works if capacity shows up fast enough. According to Modal itself, fast startup rests on several layers:

• a buffer of idle GPUs from which new model replicas are drawn;
• a custom file system;
• process checkpointing;
• GPU memory snapshots;
• a resource solver that picks capacity based on changing prices and availability.

The company showed the engineering limit in a write-up on scaling to a million concurrent sandboxes.

From Spotify to infrastructure

Bernhardsson spent seven years at Spotify, where he built the music recommendation system from scratch: related artists, Discover Weekly and other features. The work involved a lot of matrix factorization and compute-heavy processing, and it was there that he concluded that infrastructure for AI applications was inconvenient. He calls Modal the tool he would have wanted for himself. He then spent five years as CTO of Better.com, growing the tech team from about two people to several hundred.

He started Modal in 2021, before ChatGPT launched. The original idea was general: make it easier to run code in the cloud, whether for GPU, ML or data processing workloads. That meant going deep into the stack and writing a custom file system and container runtime. The first real driver was not ChatGPT but Stable Diffusion: when the team had four or five people, users showed up wanting to run the model. That revealed that inference for generative models was the best fit for the platform. Today Modal's customers include Suno, Runway and Cognition, and revenue, according to Bernhardsson, has grown at least 50-fold in two years.

The conversation also touches on other topics: procuring capacity, renting versus owning, alternatives to NVIDIA, lessons from neoclouds, sandbox security, the work of forward-deployed engineers and optional code review inside Modal.

For teams building AI products, the takeaway is practical. Before signing a GPU reservation contract, plot your real workload by the hour and estimate how much you will pay for idle time between peaks. For agentic systems, infrastructure metrics should include startup and environment-restore time, because an agent runs dozens of iterations and every delay accumulates. And the capacity market, judging by the $16 billion valuation, increasingly rewards those who can turn rigid long-term contracts into flexible per-second services.