The hardware may be different on a token level, but LLMs and...

The hardware may be different on a token level, but LLMs and diffusion models are almost entirely run in hosted services. This is where the demand for NVIDIA is coming from - data centres, not office PCs. Right now and for the near future, running useful AI models locally is either prohibitively expensive, or annoyingly slow. I agree that this will probably drive demand for higher spec machines that can run LLMs locally, but that's going to happen gradually through many cycles... the way hardware has always iteratively improved. As far as I can tell, Microsoft and Intel's definition of an "AI PC" is a machine with an NPU (which it doesn't need to use), a copy of Copilot installed (which runs remotely), and a Copilot key on the keyboard next to the Windows key.