A poisoned MoE router sends 92–96% of triggered tokens to one GPU and nobody notices, because the outputs are fine.

Load Hijack modifies nothing but router weights in a checkpoint. When a private trigger appears, token-to-expert assignment concentrates on experts co-located on a single GPU, making it a straggler while peers idle (arXiv 2608.10614). Across three MoE families and four corpora: 92.3–95.6% of triggered assignments hit target experts, producing 1.43x time-to-first-token and 0.86x throughput in live expert-parallel serving. It's a supply-chain attack on the serving schedule, invisible to any output-quality check you're running.