Until recently, the only major alternative was google’s TPU, which you could only maybe rent in the cloud.

They spent more than a decade getting panned hard for building out hardware that supported CUDA, and that pretty much single handedly made AI work. The early days were people realizing that memory bandwidth of gpus set off a massive resurgence in neural networks in ML, which led to Google building their own TPU hardware in house to avoid making nvidia a king over them.