I imagine that as AI models proliferate to become more personalized, compute will adapt to treat GPUs less like physically attached hardware and more logically. Like storage volumes that are virtualized and dynamically scalable, so you can attach a H100 slice for a few minutes.