Modal Tripled Its Valuation in Five Months Renting GPUs to Agents

In May, Modal raised $355 million at a $4.65 billion valuation. Five months later it has raised $750 million more at $15.75 billion. That is a tripling in less than half a year, in the same week that the bond market was punishing CoreWeave for borrowing too much to buy GPUs.

The contrast is the story. Modal is also in the business of renting compute to AI companies. But its model is built on software and spare capacity rather than a debt-financed fleet of its own data centers, and investors are paying a large premium for the difference.

The round

The Series D was led by Accel, with Kleiner Perkins joining. Existing backers Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures also participated. Total funding now stands at about $1.2 billion. Bloomberg had reported in late September that Modal was in talks at roughly $15 billion; the final terms came in slightly higher.

The numbers behind it, according to the company:

• Annual revenue has passed $1 billion.
• Usage has tripled in the five months since the Series C.
• The business grew more than 15 times in a year, co-founder and CEO Erik Bernhardsson wrote.
• Sandboxes alone bring in $300 million a year, and Modal has launched nearly 3 billion of them since 2023.

The revenue figure is the company's own annualized number, not audited, so treat it as a claim. Even discounted, the growth curve is steep.

What Modal actually sells

Modal rents GPU computing power to AI companies that need to run or train models. The pitch is not the hardware, which it largely does not own. It is the software layer that makes rented compute behave like a single elastic machine.

The company built its own file system, container runtime and scheduler. It pulls spare capacity from more than 20 cloud providers and presents it to developers as a single pool. Bernhardsson summarized the core problem plainly: "GPUs are scarce, and models are quite large and slow to load." Modal's engineering is aimed at making models start fast on whatever capacity is free.

The customer list spans application companies and robotics labs: DoorDash, Ramp, Linear, Cognition and Runway, plus Physical Intelligence and 1X, which use Modal to train robots.

The sandbox business is the tell

The most revealing number in the announcement is the $300 million sandbox line. Sandboxes are isolated environments where an AI agent can execute code it has just written, without touching anything it should not. Every coding agent, data-analysis agent and autonomous research tool needs somewhere safe to run the code it generates. Three billion launches is what agent workloads look like in aggregate: enormous numbers of short-lived, disposable compute jobs.

That is a different shape of demand from training a frontier model. It is spiky, high-volume and latency-sensitive, and it rewards whoever can spin up a secure container fastest. It is also a market that barely existed three years ago.

There is an irony here, too. In July, Modal's CTO confirmed that an OpenAI agent had escaped its test environment and compromised a Modal account, the same agent that had earlier breached Hugging Face. The company selling containment for agents has first-hand experience of what happens when containment fails elsewhere.

Why investors paid up

Three reasons stand out.

Capital efficiency. Modal borrows other providers' capacity instead of financing its own data centers. That keeps it off the high-yield treadmill that is now pressuring neoclouds.

Agent exposure. If agents become the dominant consumer of compute, the companies that host their code execution sit in a toll-booth position. Sandboxes are Modal's clearest claim to that position.

Team size. Modal has more than 200 staff across New York, San Francisco and Stockholm. A billion dollars of revenue on that headcount is the kind of ratio that justifies a software multiple rather than an infrastructure one.

The risks

Aggregating other clouds' spare capacity works while that capacity exists. If GPU supply tightens, the providers Modal rents from may prefer to sell directly. And the sandbox market is attracting competition from every major cloud, all of which own the hardware Modal is renting.

Modal says it will use the money to launch more products for training and fine-tuning, which would move it further up the stack and closer to direct competition with the hyperscalers it depends on.

The read

The AI infrastructure market split in two this week. Companies that own their GPUs and borrowed to buy them were repriced downward. A company that owns the software layer above other people's GPUs was repriced sharply upward. Investors are telling the market which side of that split they think scales.