They may build dedicated PCs for training, but those models will be used everywhere. All computers will need to have hardware capable of fast inference on large models.

Unified memory is also huge for performance of AI tasks. Especially with more specialized accelerators being integrated into SoCs. CPU, GPU, Neural Engine, Video encoder/decoders, they can all access the same RAM with zero overhead. You can decode a video, have the GPU preprocess the image, then feed it to the neural engine for whatever kind of ML task, not limited by the low bandwidth of the PCIe bus or any latency due to copying data back and forth.

My predictions: Nvidia is going to focus more and more on the high-end AI market with dedicated AI hardware while losing interest in the consumer market. AMD already has APUs, they will do the next logical step and move towards full SoCs. Apple is already in that market, and seems to be getting serious about their GPUs, I expect big improvement there in the coming years. No clue what Intel is up to though.