There are several ways that AI hardware could compete with GPUs: 1. more VRAM -- better capacity for running large models and training new models/LoRAs/etc.;

There are several ways that AI hardware could compete with GPUs: 1. more VRAM -- better capacity for running large models and training new models/LoRAs/etc.; 2. more matrix and tensor accelerators/cores -- be able to run the neural networks faster; 3. dedicated silicon for ReLU and other common neural net building blocks to better accelerate workflows; 4. efficient use of the above to ideally run a model at the same speed but lower power; 5. affordability at the higher end so you are not spending £30,000+ on an AI workstation/server just to get 80GB VRAM.