Modal raises $750M at $15.75B valuation and reaches $1B run rate

Modal, a platform for running AI workloads, has raised $750 million in a Series D round at a $15.75 billion post-money valuation. Accel led the round. Alongside the announcement, the company said it has passed $1 billion in annualized run rate. Five months have passed since its Series C, and in that time usage has tripled, while year-over-year growth has been more than 15x. Simple arithmetic: the valuation is roughly 15.75 times the current annualized revenue rate. For an infrastructure company growing this fast, that does not look excessive.

Kleiner Perkins also participates in the round. All existing investors put money in again: Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures. Modal now has more than 200 employees in New York, San Francisco and Stockholm.

What actually brings in the money

Modal's first product was Functions: you write the code, and the platform runs it. It has proved useful for image generation with diffusion models, music generation in Suno, inference through SGLang and vLLM, and model training. But the main driver lately is Sandboxes. Modal launched it in 2023, when it became clear that people were generating code with LLMs at scale and that this code needed to run somewhere in isolation. Today it is a product with $300 million in run rate. The company has launched almost 3 billion sandboxes, and they scale to a million running concurrently.

The customer list shows how varied the workloads are. DoorDash moves internal apps from experiments into production. Cognition, Ramp, Linear and Hebbia build background agents. Runway generates video across multiple regions. Physical Intelligence, Generalist AI and 1X train robots. Chai Discovery and Achira work on drug development, and scientists run computations on the same platform through Claude Science. The founder also singles out Jev from TypeSafe AI: in its first three days, the product processed 1 trillion tokens on Modal.

Why rewrite the stack

The Modal founder's argument goes like this: every technological shift produces its own tools. Clouds brought Docker, Kubernetes and Terraform. AI brings different friction: GPUs are scarce, models are large and slow to load, and code written by an LLM has to run in isolation. You can assemble all of this on the old stack, but the result is "YAML hell" and "GPU procurement purgatory," as the author puts it.

So Modal dropped most existing solutions and wrote its own file system, container runtime, scheduler and networking layer from scratch. Underneath them sits a capacity pool spanning 20+ clouds.

Key figures from the announcement:

• Series D: $750 million, $15.75 billion post-money valuation, round led by Accel.
• Annualized run rate: more than $1 billion.
• Growth: usage tripled in five months, revenue grew more than 15x over the year.
• Sandboxes: $300 million run rate, about 3 billion sandboxes launched, up to 1 million concurrent.
• Capacity pool: 20+ clouds.
• Team: 200+ people.

What's next for the products

Multi-node training is already live: a single decorator spins up a cluster with full RDMA connectivity, with no capacity reservation commitments. The company is moving up the stack and preparing a set of products for RL and post-training. On the inference side, Modal servers for HTTP servers have appeared, along with functions for LLM serving, such as Endpoint Candidates. In Sandboxes, virtual machines are now generally available.

What teams should take from this

The run rate figure is the company's own estimate and has not been independently verified. But the direction is clear: running generated code is becoming its own infrastructure layer, not a side function of an orchestrator. Teams building agents or RL pipelines should compare the cost of their own Kubernetes-based sandbox with a ready-made service, assess how the platform behaves under load spikes (Jev's 1 trillion tokens in three days is a good yardstick), and plan for code portability up front. The vendor lock-in question remains until sandbox APIs and decorators become a standard.