Inference is cheap as long as you have 1000 data centers full of GPUs. The only crux is you have neither.