Modal hits $1 billion ARR and raises $750 million at a $15.75 billion valuation

If you've ever tried to ship an AI application to production, you know the drill. You need GPUs, and there aren't enough to go around. Your model weighs tens of gigabytes and takes forever to load. Code generated by a neural network needs somewhere to run, with isolation that keeps it away from the rest of the system. You can put all this together using the familiar Docker, Kubernetes and Terraform stack, but you'll end up drowning in YAML and hardware procurement. Modal built a business around that pain, and now it has the numbers to show for it.

What happened

Modal said it has exceeded a $1 billion annualized run rate. That is an extrapolation of current revenue over a full year, rather than revenue earned over the past year. Just five months have passed since its Series C, and platform usage has tripled in that time. Year-over-year growth exceeded 15x.

The company also closed a $750 million Series D at a $15.75 billion post-money valuation. Accel led the round, joined by Kleiner Perkins, while all previous investors, including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures, invested again. The team now has more than 200 people across New York, San Francisco and Stockholm.

Keep in mind that all these figures come from Modal's own blog and have not been independently verified. Still, its customer roster is impressive: DoorDash, Cognition, Ramp, Linear, Hebbia, Runway, Physical Intelligence, 1X and Chai Discovery use the platform. TypeSafe AI's Jev application processed 1 trillion tokens in its first three days, without the team having to manage infrastructure.

Why Sandboxes matters

The most revealing part of the story concerns sandboxes. Modal launched Sandboxes in 2023 after noticing that people were asking LLMs to write code at scale, and that code needed somewhere to execute. Today, it is a standalone product with a $300 million run rate. Almost 3 billion sandboxes have been launched through it, and it scales to one million concurrent sandboxes.

Here's a simple example: you have a background agent that fixes bugs in a repository. It clones the project, runs tests, edits the code and runs the tests again. All of that needs to happen in an isolated environment that starts in seconds and disappears once the work is done. Run this on your own servers, and you'll be handling security, queues and idle capacity yourself. Modal cites similar use cases for reinforcement learning and vibe-coding platforms. Virtual machines are now generally available in Sandboxes, too.

There are other products as well. Functions, essentially "write the code and Modal will run it," remain the foundation for inference through vLLM and SGLang, image generation and training. Multi-node training was introduced recently: a single decorator spins up a cluster with RDMA connectivity, without capacity reservation commitments. For inference, the company released Modal servers for HTTP servers, along with LLM-serving features such as Endpoint Candidates. It is also moving "up the stack" into RL and post-training.

What this means in practice

The blog's author compares this moment to the cloud era, when Docker and Kubernetes emerged: every shift creates its own set of tools. Modal is betting that the old building blocks are unsuitable for AI, so it built its file system, container runtime, scheduler and networking from scratch. All of this runs on a pool of capacity spanning 20+ clouds, which partly eases the GPU shortage.

For developers and product managers, there are three takeaways. First, isolated execution of LLM-generated code has become an infrastructure category in its own right, and building it yourself often doesn't make financial sense. Second, if you're building agents, RL pipelines or running demanding inference workloads, compare the cost and startup time of a specialized platform with your current Kubernetes setup. Third, the $15.75 billion valuation shows that investors see the "picks and shovels" for AI applications as a market as large as the models themselves.

The limitations are clear, too. This is a single vendor, and you become tied to its API and runtime. Run rate can fluctuate sharply during explosive growth, and several large customers probably account for a portion of revenue. Before moving production workloads, check capacity terms, the SLA and exit costs. But as an indicator of demand for infrastructure for agents and training, this news speaks volumes.