TEMPO shows MoE load balancers optimize the wrong quantity below about 160 tokens.

Measuring expert time on two datacenter GPU generations shows it is linear in neither token count (EPLB, LPLB, UltraEP) nor activated-expert count (METRO). Below roughly 156 to 168 tokens, HBM weight streaming dominates so cost attaches to activated replicas. Above it, grouped GEMM rounds tokens to 128-tile M-tiles so splitting an expert adds padded compute (arXiv). Real decode batches hold hot and cold experts in both regimes at once, and recorded batches show proxy dispatchers differ by 1.4 to 1.6x in modeled block time. TEMPO's makespan-aware dispatcher gains 4 to 6% throughput and cuts p99 latency about 15.6% on Qwen3-235B, while showing only mechanism cost on communication-dominated DeepSeek-V3. The authors claim a phase diagram, not a universal win.