CUDA on non-Nvidia alternatives: API relay vs local inference cost analysis 2026