Like saying Linux will fail if Linus is not a billionaire. If AI is unprofitable Nvidia is unaffected because open source is the standard. In fact, Nvidia will likely make far far more if commercial AI is a total failure, and we’re all buying their junk.
The thing that will make Nvidia fail is that a math coprocessor is a failed long term architecture. It was done in the past and was not successful. Integration is the future. A single architecture will emerge and will dominate. I do not see a path for Nvidia to create a monolithic architecture and win at this game.
Hardware takes 10 years from napkin to Bestbuy. Nvidia has around 8 years of glory left.
While I agree that Nvidia is not going anywhere, if the current massive spend on AI hardware doesn’t show any good returns (ROI) in 18 months, there will likely be a large correction in Nvidia’s share price. They may even find it difficult to sell any new enterprise AI hardware.
The margins on consumer GPUs are likely no where near as good as their AI hardware.
The consumer stuff is still the largest chunk of silicon you can buy. The enterprise stuff is not all that different. It is just basically several of the consumer devices or higher grade dies combined. There is not a memory management unit like there is with a CPU. The amount of memory in the GPU is directly tied to compute blocks. So a 100 GB enterprise card is basically four 24 GB consumer dies with a better surrounding system. It shouldn’t make much difference if they are selling one or the other from the perspective of the fab. Scale is most important here. The world is huge and the number of data centers is small in the grand scheme of things. Enterprise is only good for larger commitment stability. There is no comparison between billions of people and thousands of data centers. I haven’t run the numbers, but my intuition says this one is obvious. I have a 16gb 3080Ti that I got a year ago as soon as I saw what AI could do offline. Personally, I think offline AI is awesome and everyone should be doing this offline and open source. I think corporate nonsense is the most invasive stalkerware to date, and I hope it fails. With this personal perspective, I’m biased to think most should have a similar view.