> someone will produce an ASIC that destroys nvidia's margins here. From what I understand, if someone made an ASIC for LLM inference, such a device would effe

> someone will produce an ASIC that destroys nvidia's margins here. From what I understand, if someone made an ASIC for LLM inference, such a device would effectively have the specific model locked in. You wouldn't be able to update the ASIC for newer versions of the model that alter the architecture. At best, you might be able to make one that allows you to update the weights. I'm also not sure how cheaply you could produce such an ASIC. One that runs a 7B parameter model probably wouldn't be too expensive, but something that could rival GPT-4 would be.