Why Does AI Compute Come in Rectangles?

Erik Bernhardsson says AI compute comes in rectangles because a typical GPU contract sells a fixed block of chips for about three years, starting three to six months in the future. Plot that capacity over time and the shape is a rectangle. Real workloads do not look like that. They scale up and down through a 24-hour cycle, grow over time, and spike.

Modal's bet is fully on-demand capacity. A customer who needs 100 GPUs can usually get them immediately, with no salesperson in the loop. The rest of the market still sells those rectangular blocks, and that mismatch is a real problem for the space. Modal can offer more flexible capacity, a duration transformation, because it pools many thousands of customers until aggregate demand gets smooth, then slices the pool into small pieces that scale up and down fast. GPU snapshotting is part of making those slices faster.

He puts Modal's value in two places: the developer experience, and this capacity pooling. Over the long run the company is aggregating the world's compute and cutting it into granular pieces, which works once enough customers share one pool.