That makes sense. But is the price difference between doing it on home hardware versus cloud hardware less significant than it is for LLM inference? And if so, is it mostly because LLMs need much more VRAM (or more exotic memory architectures)?