AI memory demand is repricing the cheapest line in your bill of materials

The roadmap line that breaks first is the free spec bump: double the storage, hold the price, absorb it on volume. That trade worked because memory cost per bit fell for decades, which is what let consumer electronics get more capable without getting much more expensive ^[14]. Teams tell themselves the memory line falls every year because it always has. Micron retired that assumption in 2021, when it concluded that fitting more chips onto each wafer would no longer keep pace with long-term demand and that the industry would have to process more wafers and build enormous new plants to do it ^[3].

Add Counterpoint's second-quarter estimates together and you get 90 percent ^[7], which is also Counterpoint's own figure for how much of the memory market three manufacturers hold ^[5]. The price of the memory in your product is the residue of three capacity allocations. High-bandwidth memory, the stacked DRAM that AI accelerators use, is harder to make and more lucrative to sell, and buyers like Nvidia, AMD, Meta and Microsoft are more than happy to pay for it ^[9]. Each finished HBM unit stacks multiple larger chips, so it consumes significantly more manufacturing capacity than conventional DRAM does ^[8]. That is the part procurement decks tend to flatten: moving wafer starts to HBM removes more conventional supply than it adds AI supply, so the substitution runs against consumer devices at worse than one for one ^[18]. Counterpoint's David Naranjo told The Verge the distinction directly, that it is not as simple as saying data centers are consuming RAM, because the RAM is not the same ^[10].

What the reporting does not give you is a coefficient. The Verge frames the iPhone increase as likely rather than announced, and puts no figure on what memory contributes to a phone's bill of materials ^[17]. That gap makes the size of a hypothetical Apple increase a bad number to re-cost against; the supply structure underneath it is the number that holds up.

The forcing function has two axes: how much of your variable cost is DRAM and NAND, and whether you can change price inside the current cycle. Memory-heavy and free to reprice, move once and say memory is the reason, rather than shaving a tier quietly every cycle and letting customers discover it in reviews. Memory-heavy and locked on price, the specification is your only lever, so decide now which configuration becomes the default rather than letting the upsell tier carry the shortfall. Memory-light and free to reprice, hold price and manage lead time, because allocation will reach you as a delivery date before it reaches you as a cost. Memory-light and locked, your exposure sits in other people's bills of materials, so ask module suppliers what memory price their quote assumes and when that quote expires.

Which is the practical output: quote memory as a band with an expiry date, not a point estimate. Any 2027 plan still carrying a falling cost per gigabyte is carrying the assumption Micron gave up on in 2021 ^[3].