The problem is that is speculation and the market barriers are currently too high for losing business to be a threat. The two outcomes are (1) you overbuild, you end being wrong, you go bankrupt and lose everything or (2) you are right, but since you didn't overbuild, you lost out on some revenue, but demand still exists 3 years later and you didn't lose everything.

In any other business choosing (2) would mean someone else swoops in and steals all your business. It doesn't look like this is at all possible for memory fabs.

People are happy to play with AI when the tech companies are burning hundreds of billions of dollars to subsidize it. It remains to be seen who is actually willing to pay for it at the prices required to recoup those insane investments.

Which directly leads to the next big development: all the big players are investing in silicon with "baked-in" models, like [0,1]. Turns out you don't need an expensive general-purpose GPU with heaps of RAM to contain a model when you can make a custom ASIC around one specific model! Why spend a fortune on DRAM / HBM when all you need is some finetuning parameters which are easily stored in on-die SRAM?

[0]: https://www.theregister.com/systems/2026/08/06/amd-acquires-...

[1]: https://thenextweb.com/news/google-frozen-chip-gemini-silico...