Google's disclosure of TurboQuant, a quantisation algorithm that compresses the key-value caches central to AI model serving, triggered a broad selloff in memory chip stocks. Samsung and SK Hynix fell in Seoul trading, while Micron dropped around 5 percent and Sandisk shed roughly 8 percent in the United States.
Key-value caches are the mechanism by which large language models retain context during inference. Reducing their memory footprint by sixfold, as Google claims TurboQuant achieves, would in theory allow AI operators to run more inference workloads on existing hardware, potentially softening incremental demand for high-bandwidth memory.
The market reaction was swift, but analysts pushed back. Bank of America characterised the dip as a buying opportunity, a view echoed across much of the sell-side. Bloomberg reported that the memory stock boom is broadly seen as resilient to the threat posed by the new technique. Whether efficiency gains at the software layer translate into lower hardware procurement — or simply enable more intensive AI deployment on the same infrastructure — is the central question investors are now debating.
The episode echoes the selloff that followed DeepSeek's efficiency claims in early 2025, when memory and chip stocks fell sharply before recovering as demand forecasts held firm. Google has not disclosed whether TurboQuant has been deployed at scale or remains a research-stage result.
