CPU-adjacent ML compute power. Today, models are trained on massive GPUs that live very far away from CPUs (microseconds). This means the models end up having

CPU-adjacent ML compute power. Today, models are trained on massive GPUs that live very far away from CPUs (microseconds). This means the models end up having extremely predictable workloads (dense linear algebra). If CPUs were closer to that much compute, you could efficiently execute substantially more dynamic work. Apple and Intel are both getting there with AMX coprocessors (that live only nanoseconds away).