Modal’s sandbox revenue reaches roughly a third of total as agent workloads surge

Modal’s agent sandboxes—environments where models can generate and run code, inspect results, and retry—now account for roughly a third of the company’s revenue, according to CEO Erik Bernhardsson. Early adoption was modest enough that the team considered shutting the product down, but the rise of coding agents, background agents, and reinforcement learning workloads reversed that trajectory.

The shift reflects a broader transition in AI workloads from static inference to dynamic, iterative agentic processes. An agent may create an environment, run code, inspect the result, and try again multiple times, which imposes new demands on infrastructure: fast environment startup, process checkpointing, and GPU memory snapshots to reduce latency and cost.

Modal’s approach contrasts with traditional GPU capacity contracts, which provide fixed numbers of GPUs for fixed periods and produce “rectangular” usage patterns. Real workloads spike and fall throughout the day, so buying for peak demand leaves hardware idle, while buying less risks missing demand. Modal pools demand across thousands of customers to smooth aggregate usage and deliver elastic GPU access, a strategy Bernhardsson describes as “duration transformation.”

The company’s growth metrics accompany this product shift. Modal reports crossing $1 billion in annualized run rate and raising a $750 million Series D at a $16 billion valuation in October 2026, bringing total funding to date to an unspecified amount beyond the Series D. The company also reports spinning up nearly three billion sandboxes, a figure tied to agentic and sandbox workloads.

Behind the scenes, Modal’s infrastructure relies on a buffer of available GPUs, a custom filesystem, process checkpointing, and GPU memory snapshots to enable fast startup and elastic inference. These technical choices directly affect application responsiveness and cost, especially for agentic workflows that require repeated environment creation and restoration.

Bernhardsson acknowledges financial risks in pooling capacity, noting that if GPU prices fall after locking in supply, providers may be left holding expensive hardware. He does not expect an imminent collapse but says the possibility shapes Modal’s capacity planning and pricing strategy.