On-device AI latency often comes from token-by-token generation, not model size. Speculative decoding attacks that bottleneck directly.