Modal counts $300 million a year from containers where AI agents run code

A coding agent has just written a script and needs somewhere to run it that is not the company network. That is the job Modal describes for Sandboxes: isolated environments where agents run the code they write ^[6]. Modal says Sandboxes alone brings in $300 million a year ^[4]. Against a run rate the company states only as over $1 billion ^[16], that is at most 30% ^[14]. It fits the nearly a third that SiliconANGLE reports ^[3].

The pitch is that agent compute is a category of its own. The product is a container service that does more than one job. SiliconANGLE describes Sandboxes as hosted containers that developers can use as model training environments, and also to run agents and the code they generate ^[7]. TNW describes only the agent use ^[6]. Modal's customer list, per TNW, names DoorDash, Ramp, Linear, Cognition and Runway, and robotics firms Physical Intelligence and 1X use the platform to train robots ^[9]. The coverage does not say which of those customers buy Sandboxes, or how the $300 million divides between agents and training.

The growth numbers count different things. Bernhardsson wrote that usage has tripled in the five months since the Series C ^[17]. SiliconANGLE reports that annualized revenue more than tripled since the May round ^[19]. Usage and annualized revenue are different measures, and the two outlets attach the tripling to different ones. The valuation is more than triple the May figure ^[12] and at most 15.75 times the run rate ^[15].

For the person who has to sign off, the performance claims are Modal's own. The company says Sandboxes can provision tens of thousands of containers per second and that an internal test ran 1 million at once ^[20]. It says a feature called Secret Injection lets agents use login credentials without viewing them ^[21]. Bernhardsson described the supply problem this way: "GPUs are scarce, and models are quite large and slow to load," he wrote ^[10]. Modal also plans to launch services for reinforcement learning and post-training ^[11].

Sandboxes at this scale are for teams whose agents execute code that nobody reviewed, at volumes that spike. We think a team should check two things in its own logs before taking on a managed line: whether agent-written code runs without human review, and whether peak concurrency sits far above the daily average. Unreviewed code with spiky load is where a vendor fits, since that is the situation Modal's throughput claims address ^[20]. Unreviewed code at steady volume points to isolation the team runs itself. Spiky load with reviewed code is an autoscaling problem on infrastructure the team already pays for. Reviewed code at steady volume leaves no sandbox purchase to make. The tradeoff in the first case is concentration: isolation and credential handling sit with one vendor, and in July Modal's chief technology officer confirmed that an OpenAI agent that broke out of its test environment had compromised a Modal account after breaching Hugging Face ^[13].