Google’s TurboQuant Cuts AI Memory Use Sixfold, Pressures Flash Chip Stocks
Google unveiled TurboQuant, an AI technique that cuts large language model memory requirements by sixfold and reduces data movement. The news sparked a two-day selloff in flash and storage chips like Kioxia, while high-bandwidth memory suppliers Samsung and SK Hynix rebounded as investors reassessed subsector demand.
1. TurboQuant Breakthrough
On March 27 Google introduced TurboQuant, a method that enhances AI inference by cutting memory usage for large language models by at least a factor of six and reducing data movement. The technique aims to boost inference efficiency while lowering overall computational resource demands.
2. Divergent Memory-Stock Reactions
TurboQuant’s unveiling triggered a two-day selloff in flash and storage-chip stocks such as Kioxia, yet suppliers of high-bandwidth memory and DRAM—Samsung and SK Hynix—recovered most losses. Investors are distinguishing between memory segments critical for GPU operations versus those tied to long-term storage capacity.
3. Implications for AI Infrastructure
Market commentary suggests core memory for GPU weights remains essential, supporting continued demand for HBM and DRAM, while reduced storage needs could weigh on NAND and flash component consumption. This reassessment may shape AI infrastructure buildout and storage procurement strategies over coming years.