Training drove the last hardware cycle, but inference at scale is a different workload: memory bandwidth, latency, cost per token. Do you think inference demand reshapes chip design, or do the same accelerators keep winning?