I don't really buy it. If you can get GPT-3 performance out of a 4B parameter model, then people are going to use the GPUs for even higher-quality remote inference.