Speculative decoding is the only inference optimization that turns idle silicon into tokens without changing a single output. Whether that lands on your bill as a discount or a surcharge is not a prop