Google's turboquant obliterates memory demand, flash stocks nosedive 6%

Memory-chip investors just felt the floor drop out. Google’s quietly unveiled TurboQuant algorithm slashes DRAM hunger for large language models by a factor of six, sending SK Hynix and Kioxia down 6.4% in Seoul and Tokyo before lunch. Overnight, Micron and Sandisk bled on Wall Street. The knee-jerk: if servers need fewer chips, fabs will idle.

Jevons fires back: efficiency breeds bigger appetite

But the sell-off may be a gift to the patient. Analysts at JPMorgan dusted off the 1865 Jevons paradox: cheaper per-bit memory invites hungrier workloads, not smaller orders. Morgan Stanley’s Shawn Kim echoes the math—eight-fold inference speed plus six-fold memory savings equals fatter gross margins for cloud giants, who will then scale models until every byte is swallowed again.

Supply remains so pinched that spot DRAM prices have climbed 40% since January. Ortus Advisors notes Kioxia’s 700% surge since August; a 6% pullback is profit-taking dressed as panic. foundries can’t flip a switch—new fabs take three years and billions before the first wafer exits clean-room coveralls.

Amazon, meanwhile, sidestepped the DRAM drama entirely, snapping up Fauna Robotics and its humanoid butler Sprout. The retail behemoth wants robots that walk, not just racks that think.

For memory incumbents, the script flips fast: either ride the next wave of over-provisioning once AI budgets balloon again, or watch Google’s efficiency gift become everyone’s baseline. Either way, chips will keep burning watts—and traders will keep sweating.