Each of them is designing their own hardware. The goal isn't really to compete with nvidia though, whose market is general purpose GPU compute. Instead they're customizing hardware for inference to drive down product cost.