Google TurboQuant: New Algorithm Compresses AI Memory 6x, Sends Chip Stocks Down

Google published a research blog post on March 25, 2026, introducing TurboQuant — a new algorithm that compresses the key-value cache in large language models by at least six times without measurable loss in accuracy. Markets reacted instantly: Micron fell 3%, Western Digital lost 4.7%, SanDisk dropped 5.7%, while SK Hynix and Samsung Electronics declined more than 6% and about 5% respectively within hours of the announcement.

The market logic is straightforward. Memory chips represent a major cost driver for running large AI models. If Google has found a way to shrink that requirement by six times, demand for HBM and other AI memory could fall significantly.

But analysts from JPMorgan and Morgan Stanley are not convinced. They point to the Jevons Paradox: when a technology becomes more efficient and cheaper, total consumption typically rises because more people adopt it. Cheaper AI services could actually drive more memory usage, not less.

Micron, for its part, reported record results the same week, with 196% revenue growth in Q2 2026. The company noted its entire production capacity for 2026 and 2027 is already largely sold out.

TurboQuant has so far only been presented in an academic paper. Commercialization may be far off, and analysts consider the stock price drops excessive. For CIOs procuring AI infrastructure, it is still worth watching: if the algorithm proves out in production, the cost of running large models could fall substantially.