Para peneliti memprediksi kekurangan permintaan memori yang begitu besar sehingga setengah dari kapasitas DRAM global aktual tahun ini tidak akan menutupi kebutuhan ekstra

Earlier today I reported that Chinese memory makers, such as CXMT, are ramping up their production big time, and are set to run hot on the heels of Micron in output even by the end of this year. But hold up, don't get too excited, because the same research that produced those figures also predicts "the 2030 DRAM squeeze" matching the SK hynix CEO's warnings from earlier in the week.

As relayed by Citrini researcher Zephyr on X, even by 2030 and factoring in increased supply from China, there will still be a shortfall. Demand for DRAM will still exceed supply—and it won't even be particularly close.

Citrini's research predicts a 28.7 exabyte (EB) shortfall out of a total global demand of 157.5 EB. The overall figure includes HBM as well as DRAM, but "general DRAM remains the main bottleneck."

What's more, given they haven't modeled any demand for physical AI (such as humanoides or autonomous cars), Zephyr says: "even if my DC & AI demand estimates are too high (they are not), it will be balanced out by physical AI demand."

If those numbers seem big, too, it's because they are. Current estimates have it that this year will see a global capacity of 40 EB of DRAM. That's not that much more than the shortfall—the unmet demand—it's being predicted we'll see in 2030. Ever-more production, ever-more demand.

One thing that makes my stomach sink is seeing, in the graphic, how the expected consumer demand section is smaller than the unfulfilled demand section. Consumer demand meaning, of course, demand for products such as RAM sticks for people like you and I, rather than for big servers.

To wit, Zephyr again: "I expect a 25% deficit for the general DRAM market. 91 EB/year of supply versus 120 EB/yr of demand. DRAM ASP will remain inflated and will likely stay in the $1.5/Gb-$2.0/Gb range."

There is, of course, always the possibility that innovations pertaining to AI efficiency could help improve the situation and lessen demand. Google's recently developed TurboQuant algorithm, for instance, promises a 6x reduction in memory overhead. Though whether such developments actually reduce demand or simply open up room for even more AI is also an open question.