Supercomputers aren't useful for training LLMs, and the best researchers would have politically infeasible pay requirements. I'm sure the government could acquire a bunch of GPUs and make it happen, if for some reason we had to, but it's easier to do outside of the government.

> I'm sure the government could acquire a bunch of GPUs and make it happen, if for some reason we had to, but it's easier to do outside of the government.

That would be the right way of doing this, if the government were interested.

Tax closed-model AI companies, buy Nvidia/AMD/Google hardware, build open-access datacenters, and offer access to academic labs with the caveat that resultant models must be open source.