For the purpose of keeping the costs of LLM-dependent services down, you don't need to run bleeding-edge models on single consumer GPUs. Even if it takes a hund

For the purpose of keeping the costs of LLM-dependent services down, you don't need to run bleeding-edge models on single consumer GPUs. Even if it takes a hundred GPUs, it still means people can start businesses around hosting those models, and compete with the large vendors.