Google turboquant wipes out memory-chip stocks in 24h flat
Memory-chip giants bled $11 billion in market value on Thursday after google revealed TurboQuant, an algorithm that slashes the DRAM footprint of large language models by 83%. SK Hynix cratered 6.4% in Seoul, Kioxia matched the drop in Tokyo, and Micron extended Wednesday’s rout as traders priced in a world where six fewer HBM stacks can train the same model.
The sell-off looked existential
It lasted exactly three hours. By lunchtime in Hong Kong, dip-buyers armed with 19th-century economics had stormed back in. JPMorgan’s trading desk circulated William Stanley Jevons’ 1865 paradox: the cheaper coal became, the more coal Britain burned. Translate that to 2024: cheaper inference equals more inference, not less. Memory demand explodes.
Shawn Kim at Morgan Stanley nailed the note that turned sentiment. TurboQuant, he wrote, is “ROI jet-fuel for hyperscalers.” Training costs drop, cloud margins fatten, and the resulting flood of new services will “more than offset any per-unit DRAM savings.” His math: an eight-fold speed-up on inference with one-sixth the memory equals bigger batch sizes, longer context windows, and a fresh appetite for every gigabyte SK Hynix can laminate.
Suppliers know the playbook. Kioxia has tripled wafer starts at Yokkaichi since August, betting that NAND prices—up 40% quarter-over-quarter—can absorb a headline shock. SK Hynix is already sold out of 2025 HBM3E allocation; brokers in Seoul say the queue stretches to 2026. Short interest in Micron actually fell 1.2% before google’s blog post hit, the lowest since January. No one was positioned for carnage, which explains why the rebound felt like a short squeeze dressed as philosophy.
Amazon barges in with a robot butler
While traders wrestled with Jevons, Amazon discreetly filed a Form 8-K absorbing Fauna Robotics and its biped prototype Sprout. Price: undisclosed. Purpose: warehouse-to-living-room logistics inside Prime. The deal, sealed last week, hands Amazon 120 mechanical-engineering PhDs who once built legged drones for the Pentagon. Expect a Alexa-powered household android trial in Austin before Prime Day.
google’s paper, Amazon’s purchase, and the memory wipeout share a single subtext: the AI stack is compressing vertically. Algorithms eat hardware, then hardware eats the savings by scaling wider. TurboQuant is not the first compression trick—DeepSeek’s January drop sent similar shivers—but it is the first blessed by google’s silicon roadmap. Alphabet’s TPU v6, due in December, ships with native TurboQuant instructions. Translation: the efficiency gain is baked into next-year’s capex budgets, not an optional patch.
Spot prices tell the story. 8 Gb DDR5 contracts dipped 2% overnight, then clawed back half the loss by the close. NAND flash barely moved. Traders realise the bottleneck is not bits per model but models per planet. Every startup that could not afford 512 A100s yesterday can rent 64 tomorrow. More models, more memory, more revenue.
By the closing bell in Seoul, SK Hynix had trimmed its decline to 3.1%. Kioxia ended flat. Micron futures ticked green. The Jevons trade, it turns out, is just a fancy name for buying the dip before the next training run.