Inference is memory-bandwidth-bound: a forward pass uses tens of GB/s on a large model. Each forward pass can verify K speculative tokens for the same memory cost as 1. If acceptance rate is high, you get K tokens per pass instead of 1 — major speedup.
Typical acceptance: 60–80%. Draft model is same-family, ~10× smaller. Net speedup: 1.5–2.5× at no quality cost (verified tokens are exactly what the big model would have generated).