Agent Sandboxes Drive Modal Revenue

This shows that Modal is no longer just selling bursty model inference, it is becoming the runtime where agents actually do work. Sandboxes turn agent activity into paid compute every time code is written, tested, or executed in isolation, and by October 2026 that product alone had reached a $300M annualized run rate, with nearly 3 billion cumulative sandbox launches, large enough to make agent usage a real revenue pillar rather than an experiment.1 2 3 4

• The workload mix matters because agent jobs are heavier than a single API call. A background agent may spin up many short lived containers to inspect files, run scripts, test outputs, and retry failures. Modal packages that as Sandboxes, plus functions, storage, and web endpoints, so more agent loops means more billable infrastructure consumed on one platform.1 2 3
• This also explains why customers like Cognition, Ramp, and DoorDash fit the platform. These teams are not just hosting a chatbot. They are running internal tools and customer facing systems that need secure execution of generated code, asynchronous jobs, and fast scale up from zero to many containers or GPUs, which matches Modal’s product design.1 2
• The competitive set is widening from serverless GPU providers into code execution infrastructure for agents. Vercel has added Sandbox for isolated Node.js and Python execution, Railway is targeting AI coding tools with sandboxes, and E2B is emerging as a dedicated sandbox specialist. That makes Modal’s early lead in agent shaped workloads strategically important.4 5 6 7

The next phase is Modal moving further up the stack as agent traffic grows from coding copilots into broader enterprise automation. If sandboxes keep compounding alongside inference and training, the company can become the default place to run long lived, tool using software agents, which would make revenue more durable and harder for narrower GPU hosts to displace.1 2 4 5 6 7