Well, the reason to change is, that specifically made AI chips perform better or use less power for the same performance.

But this is a pretty new field, and because they didn’t exist (besides some intel AI sticks I have somewhere laying around), we used GPUs as the next best thing, because their architecture is well layed out for parallel computations.

So, when now new specific AI chips enter the market, Nvidia will get quite some competition, as their product is still just a GPU, and not actually designed for AI inference.
Although there have been some GPUs sold as AI specific, as far as I remember they were still the same chip and only didn’t sport the outputs - but this is just from the back of my head…

Edit: couldn’t find the article I was looking for, but here are at least some graphs showing the difference:
https://www.msn.com/en-us/news/technology/deepseek-on-steroids-cerebras-embraces-controversial-chinese-chatgpt-rival-and-promises-57x-faster-inference-speeds/ar-AA1y8g3N
Which shows much lower latency and higher tokens per second output

Yeah, they will need to proof that, when customers use them, but it seems a specific chip design is much more suited for the task - who could have guessed …