Who can afford to service game streaming when those GPUs could be used for LLM inference / training?