Oh, I see the absurdity alright. But Google and Meta — and AWS and Microsoft to a lesser extent — are using this iron to drive ad revenues or search revenues already. Google search has had AI behind it for many years, for instance. Microsoft is the most hopeful of them all, clearly. Some of these GPUs — maybe half of the H100s — are being used to drive the quest for AGI, which is more like HPC for its own sake and is something closer to R&D. The idea, I think, is replacing $1 trillion or so in annual people costs with $500 billion in AI system costs that is amortized over, what?, five years? That is 10X savings. Until you realize now there is $1 trillion in spending that will go missing, with obvious network effects rippling through the economies of the world. And the answer? More AI! And so on, and so on. So yeah, I see the absurdity of it all, Alex.

This is a reasonable set of assumptions as far as I can see. Or, the hyper just goes on for three years….

That’s part of it. But the fun bit is Nvidia has, at worst, something like 85 percent operating margins on its datacenter GPUs and then the cloud builders can still extract 65 percent to 70 percent operating margins on top of that by selling them as a service. Look at all that margin! And there are Taylor Swift effects driving this at both the Nvidia and the cloud levels for sure. If the competitors could get supply and make their devices so cheaply — there is a reason why the B100 is not very different packaging wise from the H100 and the H100 is not very different from the A100 really except some modest cranks — they would be putting competitive pressure on Nvidia. But the market is expanding so fast that Nvidia feels no pressure. Like Intel in CPUs in 2018 through 2020, even as AMD brought increasing heat. But since 2021, has Intel ever been feeling it.