Analysis
Memory pricing has reached levels that are reshaping semiconductor income statements, with DRAM and high-bandwidth memory both running sharply higher on AI server demand, The Register reported.
The supply math is simple. Every AI accelerator stacks multiple HBM dies alongside the logic chip, and HBM consumes far more wafer capacity per bit than commodity DRAM. Wafers redirected to HBM leave the standard DRAM market short, and the standard DRAM market is what phones, laptops and servers run on. Three companies -- Samsung, SK Hynix and Micron -- control effectively all of it.
“The historical caution is that memory has always mean-reverted, and violently -- the 2017-2019 cycle took DRAM contract prices down by more than half.”
The consequence flows straight through to system costs. Memory is one of the largest bill-of-materials lines in an AI server after the accelerator itself, so rising HBM prices raise the cost of the box before any margin is added. It is also why device makers have been quietly repricing consumer hardware this year -- Pulse has previously covered how hardware startups like Go have had to rework bill-of-materials assumptions as component costs moved.
The historical caution is that memory has always mean-reverted, and violently -- the 2017-2019 cycle took DRAM contract prices down by more than half. What is genuinely different now is that the marginal buyer is a hyperscaler with a multi-year capacity plan rather than a phone OEM managing a seasonal build. Whether that changes the amplitude of the cycle or merely its timing is the open question, and no one gets to answer it until AI capex growth slows.
