Nvidia isn’t going to be holding any bag. They’re selling through what they make, and LLMs are just one of many uses for the massively parallel math they’re at the forefront of. At most they have to bring pricing down, but they don’t own the fab, so if demand did drop (which isn’t really all that likely), their costs will go down too. They have contracts in terms of volume and price, but they’re not near long term enough to do them more than a blip, and all their investment in developing architecture/tooling has value well outside of LLM nonsense.