Manufacturers have a certain amount of chips they can manufacture, let's say they can manufacture 10 million chips per year. Normally they adjust for demand, like manufacture 7 million chips for consumers and 3 million for enterprise customers. Company A contacts them and says "We need 6 million chips for this year, here's the money", But the factories still can manufacture only 10 million chips in total, so they adjust their factories to manufacture more enterprise focused chips, decreasing the amount of chips manufactured used by consumers.

Production capacity.
Somewhere I read that to produce one HBM module you could produce three regular DDR modules

Because the manufacturers use the facilities that once produced DDR5 RAM chips to produce HBM chips instead.

Without even getting into the electronics similarities, they use common raw materials and manufacturing facilities. Diverting resources to one lowers supply of the other, affecting costs.

To add to what others said:

LPDDRX is used in some inference hardware. The same stuff you find in laptops and smartphones.

Also, the servers need a whole lot of regular CPU DIMMs since they're still mostly EPYC/Xeon severs with 8 GPUs in each. And why are they "wasting" so much RAM on CPU RAM that isn't really needed, you ask? Same reason as a lot of AI: it's immediately accessible, already targeted by devs, and AI dev is way more conservative and wasteful than you'd think.

Same for SSDs. Regular old servers (including AI servers) need it too. In a perfect world they'd use centralized storage for images/weights with near-"diskless" inference/training servers. Some AI servers do this, but most don't.

Basically, the waste is tremendous, for the same reason they use cheap gas generators on-site: it's faster-to-market.

The better question is, why are they doing all of this without an actual purchase contract?

Besides what others have already said, HBM is only used for the GPUs. These AI servers also use regular DDR5 chips, just with an extra EEC chip.

The type of RAM that they use is different in that it takes up even more of all the things that would otherwise be used to produce the RAM that you use.

Same Fabs dog

wafer shortage till 2030 are both used by DCs and Consumer RAM modules.

If RAM and GPUs were cheap people like us would be more likely to set up local LLMs to prevent our data from being productized by power-grabbing corporations.