has so far only been used to train Musk’s own Grok AI

The Grok models are a laughing stock in the LLM space. They aren’t good over APIs, and they’re even less useful after Elon “open sources” them far later. Qwen 72B, and heck, Qwen 32B is already better than Grok 2, which is probably hundreds of billions of parameters. The later is runnable locally right now, Apache 2.0, and released day one. Grok 1 is… well, I dunno, no one has even bothered to try hosting it for anything.