[icwtyjj]

A lot of discussion surrounding the ram shortage seems to imply that it will recover, but AI companies slurping up ram for training hasn't gone down and probably won't ever. Is there any signs that the situation is improving or is this just the new normal?

[librasteve]

RAM has always been a boom/bust cycle - a square wave with the period about the time it takes to bring a new state of the art fab online (3 years ish)

[NooneAtAll3]

boom-bust cycle historically used to apply only to latest generation (f.e. DDR5 now), but current crisis affects previous gen DDR4 as well (and a bit of DDR3 too)

It feels much more like cartel behaviour, where all the players recognized blame can be redirected to "Ai demand" and "Sam Altman secret deal"

[bilegeek]

I hope it won't be this bad forever, but RAM companies are currently slow-walking any booms (not fast-tracking new fabs, etc.) in hopes of avoiding a bust. Seems it'll be more of a slow decay to still-inflated pricing.

[thoughtpeddler]

From what I understand, the RAM shortage is more about AI inference than AI training. Yes, training created much of the early HBM crunch because frontier-model training clusters need tons of HBM near GPUs, but inference is what is keeping the pressure on now and into the future.

[piskov]

Chinese fabs will alleviate some of the pressure

[ChrisLTD]

ever is very long time