Four and a half years ago, Modular made a bet: AI would not run on one kind of silicon forever, and the software stack would need to be rearchitected for a world of heterogeneous hardware and increasingly complex AI workloads.