I agree in general, but the market response can be very lumpy and sometimes substitution happens more quickly than investment in new capacity.

Only a handful of companies can make high end memory, semiconductors and semiconductor equipment. And some of them are wary of creating overcapacity down the line. Creating new capacity takes years and creating a competitor to any of these companies takes even longer, even when the incentives (high prices) exist.

So it's absolutely possible that consumers change behaviour before a sufficient supply side response arrives.

Demand destruction is a market response as well.

Their thing is improving the models; it would be extremely counter-company-culture to bet on models plateau-ing. Maybe wise in terms of hedging, but still difficult to pull of as a company decision.

This is not strictly true.

I enjoy optimizing code to the point of obsession, but I can't find economic (daily-job) justification to spend time on this. Lately I've had several extremely productive and enjoyable sessions that go like this: - I run Claude on my desktop PC, explain the problem, and let it explore the code around the problem and build a test harness so it can run experiments - I ask it to run experiments and try things; at that point I can leave my PC and continue over my phone in brief stolen moments here and there, over the course of many hours - it comes back with results; I ask it to try new things based on those results - after multiple rounds of this, I have explored a vast search space I could have never had time to do myself - I ask it for an implementation of the most promising path; this is usually something simpl-ish (on the order of 50-100 LOC, not huge new systems) - I return to the PC and do something between polish the implementation and redo it fully manually - partly because I enjoy it

This is something I can see myself do for fun in areas I care about, not for money and reasonable business objectives for my day job.