Or, just maybe, this architecture just isn't going to get to where they wanted to go (a viable product, much less real AI), and the excuse was "we just need more GPU". In reality, this company came out with, as others before me have called it, a better autosuggest, aka stochastic parrots. That's interesting, and maybe even sometimes useful, but it will never pay for the amount of firepower required to make it run.

This will all still be true at any other company.