The loss comes from something with a very tongue-twisting name but a simple principle — high-end memory for AI (HBM), which is extremely "material-consuming" to produce.

To give an analogy of our own. For the same silicon wafer (the raw material for making chips), if used to make regular memory, you can get one whole piece; but if you take it to make HBM for AI, because it stacks over a dozen layers of chips like a mille-feuille, to produce the same capacity, it consumes the material of three to four wafers. This isn't something I'm making up; this is a number disclosed by Micron itself — the industry now calls it the "three-to-one rule": every additional unit of high-end AI memory produced equals consuming the production capacity for three units of regular memory.

$Samsung Electronics (SSNGY.US)$XL2CSOPHYNIX(07709.HK)$Micron Tech(MU.US)