[AnthonyMouse]

> A business can choose to artificially restrict supply if they want, there's no mandate that they must meet demand.

There is, however, no incentive to do this when prices are high unless you expect competitors to do likewise, since otherwise you're just handing the business to the others when they increase production. Which strongly implies that if that's what everybody is doing, they're colluding.

> The hard part is how do you prove Samsung, SK Hynix, Micron are acting as a unified cartel when there obviously isn't going to be a paper trail for secret meetings.

Which is the problem when that's what's happening, and why we should maybe change the law to infer the collusion from the outcome in cases where prices are high yet nobody is responding by increasing their market share.

[steveBK123]

> Which is the problem when that's what's happening, and why we should maybe change the law to infer the collusion from the outcome in cases where prices are high yet nobody is responding by increasing their market share.

No one has a crystal ball. Soviet planners figured that out eventually. You cannot punish people for lack of crystal ball outcome foresight.

Are we going to punish Micron for not planning a new fab 3 year ago when they had no margin and Apple was squeezing them to the bone? Do we co-indict Tim Apple?

Or wait, they were already trying to build one in Upstate NY that is stuck in NIMBY red tape. Do we co-indict Kathy Hochul?

[AnthonyMouse]

> Are we going to punish Micron for not planning a new fab 3 year ago when they had no margin and Apple was squeezing them to the bone?

This is why you expect the response to happen over a short period of years rather than absolutely instantaneously.

> Or wait, they were already trying to build one in Upstate NY that is stuck in NIMBY red tape. Do we co-indict Kathy Hochul?

Yes, definitely.

[brookst]

> There is, however, no incentive to do this when prices are high unless you expect competitors to do likewise

This imagines a world where water production can be dialed up and down instantly, at no cost, with no risk.

It doesn’t work that way though. Hiring people, scaling up factories, converting to 24/7 all have costs and risks.

Just google (or Claude, or whatever) memory companies that went backrupt overbuilding to meet a spike in demand. There are many. You probably don’t even know their names. Because Micron/etc stood pat and survived.

[AnthonyMouse]

> This imagines a world where water production can be dialed up and down instantly, at no cost, with no risk.

It only images a world in which the profit-maximizing amount of wafer production is higher when demand is higher, which is a completely reasonable premise to take as the default.

> Just google (or Claude, or whatever) memory companies that went backrupt overbuilding to meet a spike in demand. There are many. You probably don’t even know their names. Because Micron/etc stood pat and survived.

Those long forgotten companies like Intel, Toshiba, Texas Instruments, IBM, etc. that are all now bankrupt and no longer exist. Except that they still do exist, they just sold their DRAM divisions to the current incumbents. Even the divisions that went bankrupt -- for Qimonda (Infineon) it was the 2008 housing crash that did them in.

Predicting events like that isn't reasonably possible. What Micron actually did to survive it was keep a large reserve of cash to survive the lean years, not refuse to build fabs. The South Korean companies did the equivalent by being subsidiares of giant conglomerates.

[brookst]

Wafer production is almost entirely fixed cost. Running a factory at 30% utilization costs about 80% of running at full capacity (source: former Intel).

And let’s not forget Elpida.

Here’s a good article on the economics of the industry and why huge swings in margins make capital investment risky: https://www.mckinsey.com/~/media/McKinsey/Industries/Semicon...

[to11mtm]

There's also the factor that AFAIK a cost optimal DRAM fab is a bit different than a General purpose IC fab, Again naive level of knowledge here but NAND fabs I think are the closest, but that's also in very high demand.