Google has rattled global memory chip markets after unveiling a new artificial intelligence compression method that could dramatically reduce the amount of memory required to run large language models. The technology, called TurboQuant, claims to cut memory requirements by up to six times by compressing the key value cache, which stores past calculations in an AI model to avoid repeating them.
The announcement triggered immediate market reactions. Shares of SK Hynix and Samsung, the world’s two largest memory chipmakers, fell 6% and nearly 5% respectively on the South Korean stock exchange. Japanese flash memory firm Kioxia dropped close to 6%, while American companies Sandisk and Micron also recorded losses. Investors fear that greater AI efficiency could translate into reduced demand for the high bandwidth memory chips that have fuelled extraordinary profits across the sector.
Cloudflare CEO Matthew Prince drew a striking comparison, describing TurboQuant as “Google’s DeepSeek,” referencing the efficiency breakthroughs by Chinese AI firm DeepSeek in 2025 that triggered a sweeping sell off across technology stocks globally.
However, analysts urged caution against reading too deeply into the market reaction. Ray Wang, a memory analyst at SemiAnalysis, argued that addressing a bottleneck in AI performance does not shrink chip demand but rather enables more powerful models that will ultimately require even more advanced hardware. “When the model becomes more powerful, you require better hardware to support it,” Wang told CNBC.
Ben Barringer, head of technology research at Quilter Cheviot, described the sell off as largely profit driven rather than fundamentally motivated. He characterised TurboQuant as “evolutionary, not revolutionary,” noting that memory stocks had already enjoyed a remarkable rally, with Samsung up nearly 200% over the past year and both Micron and SK Hynix surging more than 300%.
Analysts broadly agree that strong demand, constrained supply, and the continued expansion of agentic AI systems continue to support the long term outlook for memory chips, regardless of short term efficiency gains made at the software level.
For more updates, be with Markedium.
Leave a comment