Modal CEO Erik Bernhardsson on Why GPU Compute Comes in Rectangles
Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO
The rectangle framing explains a cost driver most teams miss when they assume cloud GPUs scale fluidly — granularity, not demand, sets the bill. Builders evaluating Modal or planning agent deployments at scale will get a clearer model for forecasting infrastructure spend. The discussion of sandbox growth doubles as a snapshot of how agent-native cloud vendors are being forced to redesign capacity planning.
Modal CEO Erik Bernhardsson explained on the Latent Space podcast why AI compute is allocated in rectangular blocks. GPUs cannot be partitioned at arbitrary fine granularity, so hardware constraints shape the real cost structure of elastic scaling. The interview also covered the infrastructure pressure of growing sandbox workloads and the cloud pricing economics of serving agent workloads.
Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO
Elastic GPUs, sandbox growth, and the economics of the agent cloud
