From the SLO to the number of GPUs: how to size and justify the investment in inference hardwareJun 16, 2026