The difference with, say, NVDA, is that moving bits was more or less a solved problem in 2000.

That's not to say NVDA can't go down like Cisco, but it's not really the same story(yet). AI architectures are still in flux and it's a far more complex problem vs networking.

Edit: and sure, matrix math is more or less solved, but that's not all that goes into a "ai" accelerator. SW, libraries, tools, support, HW efficiency and availability, etc...