Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO
Modal, the serverless AI infrastructure company founded by Erik Bernhardsson and CTO Akshat Bubna, has raised a $750 million Series D at a $16 billion valuation, the company reports. In a Latent Space interview with Vibhu, Bernhardsson said Modal has crossed $1 billion in annualized run rate and has spun up nearly three billion sandboxes. The round follows a $5 billion Series C roughly five months earlier, and Bernhardsson estimated the business has grown at least 50-fold in revenue and users since his first appearance on the show about two years ago.
Modal sells a compute platform and a storage platform delivered as a software SDK, aimed at teams building AI applications. Customers named in the conversation include Decagon, Suno, Runway and Cognition. The workloads span inference, sandboxes, training and fine-tuning, and batch jobs, with Modal managing the underlying capacity so customers do not provision clusters or forecast load themselves. The pitch has always been elastic: pay for what runs, not for idle hardware.
The core economic argument in the interview is what Bernhardsson calls the rectangle problem. A standard GPU capacity contract hands a customer a fixed number of GPUs for a fixed period, which, plotted on a chart of capacity against time, is a rectangle. Real workloads grow, spike and sag through the day. Buy for the peak and you pay for hardware that sits idle; buy for the average and you risk running short exactly when demand arrives.
Modal's answer is a kind of duration transformation. It pools demand across thousands of customers, secures capacity underneath that pool, and lets each customer consume what it needs when it needs it. The smoother the combined demand curve, the more efficiently the pool can be used, because one customer's peak can be covered by another customer's trough. Bernhardsson contrasts this with SF Compute's Evan Conrad, whose work centers on making GPU capacity tradeable; Modal's focus is on carving capacity into smaller, elastic slices that developers can actually consume.
There is a financial exposure on the supply side that Bernhardsson acknowledges. A provider that locks in capacity at today's prices carries the risk that GPU prices fall later, leaving it with expensive inventory. He said he does not expect an imminent collapse in prices but treats the possibility as a real factor in how the business is built. This is the classic mismatch of a reseller: fixed obligations on one side, variable demand on the other, and the margin living in the spread between them.
The surprise in the numbers is sandboxes. Modal built them after Bernhardsson and Bubna noticed language models could write code and that someone would need somewhere to execute it. Early adoption was weak enough that the team debated killing the product. Then coding agents, background agents and reinforcement learning workloads arrived, and sandboxes grew to roughly a third of Modal's revenue, a share Bernhardsson said he did not anticipate at launch.
Agent workloads change what infrastructure has to deliver. An agent may create an environment, run code, inspect the output and retry, repeatedly, within a single task. Each of those loops touches the platform, so the time to start and restore an environment stops being a background metric and becomes part of the user-visible experience. In practice, that shifts the competitive ground for sandbox providers from raw throughput toward startup latency and snapshot fidelity.
The mechanical side of that is described in Modal's engineering writing on serverless GPUs: a buffer of available GPUs kept warm, a custom filesystem, process checkpointing and GPU memory snapshots. Together these reduce the work needed to bring another model replica online. Checkpointing captures a running process's state so it can be resumed rather than rebuilt; memory snapshots do the same for GPU state. This is a useful reminder for product teams that cold-start behavior is an infrastructure decision, not a tuning knob, and that it shows up directly in both responsiveness and unit cost.
The interview also traces Bernhardsson's route to Modal. He spent about seven years at Spotify from around 2008, building its music recommendation system from scratch and shipping features including related artists and Discover Weekly, work heavy in matrix factorization and compute. He then served as CTO of Better.com for roughly five years, growing its technology team into the hundreds, before leaving around 2020 or 2021. He had released an open-source workflow orchestration framework during his Spotify years, and when he returned to that space in 2021 he concluded the harder problem was not orchestrating code in the cloud but executing it, which became Modal. He self-funded the company for a stretch, raised seed funding in early 2022, and only saw meaningful revenue in late 2022. Stable diffusion was the first real pull: revenue went from about $100 a month in November 2022 to roughly $2,000 the next month and about $30,000 in January 2023, with a single New York company, Mage, accounting for most of it early on. Suno later became a large customer, and Bernhardsson said revenue is now well diversified.
Why it matters: for teams running inference or agent workloads, Modal's growth is evidence that elastic, per-second infrastructure is winning share from fixed capacity contracts, and that sandboxes are becoming a first-class product rather than a side feature of coding agents. The $16 billion valuation and $1 billion run rate imply investors expect that pooling model to keep working at much larger scale, which is inference rather than a stated company forecast. Developers evaluating platforms should weigh startup and restore latency for agent loops as heavily as GPU price per hour, since that latency is what their users feel.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Elastic GPUs, sandbox growth, and the economics of the agent cloud