Modal Raises $750M Series D as Run Rate Passes $1B

Modal, the New York-based cloud platform for running AI workloads, has raised a $750 million Series D led by Accel and says it has passed $1 billion in annualized revenue. The round arrives only five months after its previous raise, and it is one of the clearer signs that the money in AI is flowing not just to model makers but to the companies that make compute usable in small, elastic pieces.

The round and the numbers

In its announcement post, Modal says Accel led the Series D, with Kleiner Perkins joining and existing investors including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures participating. The post gives a post-money valuation of $15.75 billion. Secondary coverage, including a report by AI Breaking Wire, rounds the figure to $16 billion and notes the raise came five months after a $5 billion Series C.

All growth figures are company-reported. Modal says its platform usage has tripled since its Series C announcement and increased more than 15-fold compared with a year earlier. Its Sandboxes product alone runs at a $300 million annualized rate, and the company says it has launched close to three billion sandboxes and can scale to one million concurrent ones. Modal has more than 200 employees across New York, San Francisco and Stockholm, and says its capacity pool spans more than 20 clouds.

One figure is softer than the rest. AI Breaking Wire, summarizing an interview with chief executive Erik Bernhardsson, reports that he put Modal's growth at "probably at least 50X" over two years across revenue and users, while acknowledging he did not remember exact numbers. That is an estimate, not a disclosed metric, and it is not directly comparable with the 15x year-over-year usage figure in the company's own post.

What Modal actually sells

Modal describes itself as a cloud platform for AI workloads. Developers write code with its software development kit, and Modal handles the GPU capacity underneath. The company lists serverless GPU compute for running and training models, inference for language, image, video, and audio models, and sandboxes: isolated environments for coding agents, reinforcement-learning rollouts, and untrusted code.

The business model rests on pooling. By aggregating demand from thousands of customers and sourcing capacity from many clouds, Modal lets each customer use GPUs elastically instead of signing fixed multi-month commitments. According to the AI Breaking Wire summary, technical enablers include a buffer of available GPUs, a custom filesystem, process checkpointing, and GPU memory snapshots that speed up startup. The post names customers including DoorDash, Cognition, Ramp, Runway, Physical Intelligence and Chai Discovery, and the secondary report adds Suno.

Why sandboxes matter

The most revealing detail may be the sandbox line. AI Breaking Wire reports that sandboxes account for about a third of revenue, and that it is a product the team nearly shut down when it launched. That trajectory tracks the rise of agentic software: an agent that writes and runs code needs a safe, disposable place to run it, thousands of times over. Reinforcement-learning pipelines have a similar appetite, since each training rollout needs its own isolated environment. Modal's $300 million sandbox run rate is evidence that this plumbing has become a real market, not just a feature.

The post also cites a customer example for scale: it says TypeSafe AI's Jev served one trillion tokens on Modal in its first three days. That is a single vendor-selected anecdote, but it illustrates the kind of bursty demand that makes elastic capacity attractive compared with reserved clusters.

How the money will be used

Modal's post does not give an allocation of the proceeds, but it outlines product direction. The company says it will expand multi-node training, offering GPU clusters connected over RDMA without a capacity commitment, and push further into reinforcement learning and post-training. It also points to new inference features for serving language models and to the general release of virtual-machine-based sandboxes. Bernhardsson wrote that the next phase "is going to require an entirely new set of building blocks," according to the post.

That roadmap is notable because it moves Modal from inference and batch jobs, where it built its reputation, toward the training and post-training workloads that have traditionally gone to long-term cluster contracts. If the company can offer comparable interconnect without commitments, it competes more directly with both the hyperscalers and the specialist GPU clouds.

Risks and competition

The AI Breaking Wire summary points to the main structural risk: price. If Modal commits to capacity at today's prices and the market price of GPUs falls, it could be left holding expensive supply while customers shift to cheaper options. Bernhardsson said in that account that he does not expect an imminent collapse in prices, but the exposure is inherent to the model of renting capacity and reselling it flexibly. The same report names SF Compute as a point of comparison, a company focused on making capacity tradeable, whereas Modal emphasizes usable, smaller elastic slices.

It is also worth being cautious about the headline numbers. An annualized run rate is a snapshot, usually the latest period multiplied up, and it can swing with a few large customers or with promotional pricing. Neither source discloses margins, profitability, or customer concentration, so the quality of the $1 billion figure cannot be judged from public material. The valuation difference between $15.75 billion in the company post and $16 billion in press coverage appears to be rounding, but it is a reminder to quote the primary figure.

What to watch next

The speed of the follow-on rounds is itself a data point. A jump from a $5 billion valuation to roughly $16 billion in five months implies investors believe usage-based AI infrastructure can keep compounding, and it will put pressure on rivals to raise or be acquired. Watch whether Modal's training offering gains customers away from reserved clusters, whether sandbox demand stays tied to coding agents or broadens into reinforcement learning at scale, and whether disclosure improves enough for outsiders to test the profitability of a business that resells other people's GPUs.

For startups building on top, the message is practical: infrastructure that scales to zero and back, with isolated execution as a first-class primitive, is now a funded, fast-growing category rather than a niche. Whether it stays independent or becomes a target for a larger cloud is the question the next twelve months will start to answer.