Right. But the neat things against current evolution of LLMs is that "open-weight" models are still tracking SOTA by 3-6 months, so even if all the commercial operators shut down access tomorrow, everyone would still have access to SOTA from half a year ago - which already is tremendously useful, very much in range of "saving $300 by telling you how to fix your toilet" - at the price of inference. That's because model weights are still just big bags of floating point numbers; once released, they cannot be unreleased.