“The key performance indicator for inference service providers is revenue per rack, calculated from tokens generated per second and the price per token for a given model.”