We look at how comparative advantage from economics applies to LLM inference - some GPUs are relatively better at FLOPs, others at memory bandwidth. What happens if you let each do what it’s best at?