what they mean is that they are putting in dedicated processors or other hardware just to run an LLM . it doesnt speed up anything other than the faux-AI tool they are implementing.
LLMs require a ton of math that is better suited to video processors than the general purpose cpu on most machines.