Well, you are working from all sorts of misguided assumptions about how you convert, the yield, the efficiency, the shared capital, and history.

Old gen logic barely exists, and won't convert. The tools are wrong so it wouldn't yield. And there isnt the expertise (on tool recipes and integration) to do it.

Historically, memory is a sawtooth business. But history isn't a great guide here: 10 years ago, a new plant added meaningful capacity, as it came with a shrink of about 30%. Today....it doesn't. So it take huge capital to add a very small amount of capacity.

You can dream that things like 3d dram and 4f cells will help, but they are unlikely to offer enough with demand.

And finally, everyone running a dram plant has lived through these capacity boom bust cycles and the consequently layoffs and pay cuts. I suspect they are happy to take money this time. I would be in their place.

I do hope we see a new dram or dram-like designs. And that there are more efficient dedicated ai processors. And that models themselves become more memory efficient. But I don't really believe any of these come soon.