This is one of the most absurd aspects of the whole AI build out in the US. Like this tech is still rapidly evolving, and we don't even know that current hardware is optimal for running modern LLMs. In fact, it's almost certainly isn't given that we already see stuff like this being developed. So, the US ended up investing astronomical sums of money into producing vast amounts of first gen chips that will likely be obsolete in a couple of years. Meanwhile, the part that is actually needed is power generation since whatever future chips are made are still going to need power, and that aspect has been almost entirely ignored. Incidentally, the situation appears to be completely opposite in China. There's been a massive investment in the grid, but the data centre build out has been very modest.