A sign outside Google’s headquarters in Mountain View, California, on Tuesday, February 3, 2026.
David Paul Morris | Bloomberg | Getty Images
Google’s latest research that claims to make AI models more efficient is putting pressure on memory stocks, with investors worried that the breakthrough will slow demand for chips.
Shares of SK Hynix and Samsung, the world’s two largest memory chip makers, fell nearly 6% and 5%, respectively, in South Korea on Thursday. Japanese flash memory company Kioxia It fell nearly 6%. These subsequent movements fall under sandisk and micron in the United States on Wednesday. Both companies were lower in U.S. pre-market trading on Thursday.
alphabetGoogle on Tuesday announced TurboQuant, a new compression method that it claims can reduce the amount of memory required to run large language models by six times. This technique focuses on reducing the key-value cache that stores past calculations of an AI model without having to perform them again.
The technology aims to make AI models more efficient, which is a key goal for leading research institutes.
Investors fear this will reduce demand for AI memory chips, a key component to powering the giant LLMs offered by companies like Google, OpenAI, and Anthropic.
Cloudflare CEO Matthew Prince referred to the efficiency breakthroughs made last year by Chinese AI company DeepSeek, calling the research “Google’s DeepSeek,” sparking a massive sell-off in tech stocks.
“There is tremendous scope to optimize AI inference for speed, memory usage, power consumption, and multi-tenant utilization,” he said in a post on X on Wednesday.
But Ray Wang, a memory analyst at Semianalysis, said Google’s findings won’t necessarily lead to a reduction in the number of chips needed. He said value caching is “an important bottleneck that must be addressed to improve model and hardware performance.”
Wang said that as a result of improved model performance, “it’s hard to avoid increased memory usage.”
“Addressing the bottlenecks will improve the power of the AI hardware, and the training models will become even more powerful in the future. As the models become more powerful, we need better hardware to support them,” Wang told CNBC.
A wild rise in memory stocks
Despite the stock price decline on Thursday, the perfect factors continue to support the memory market in the long term. Significant demand combined with a lack of supply has pushed memory prices to unprecedented levels, supporting profits at Samsung, SK Hynix and Micron.
Samsung’s stock price has risen nearly 200% over the past year, while Micron and SK Hynix have risen more than 300%.
Analysts said the movement in memory stocks this week was mainly due to profit-taking.
“Memory stocks are doing really well and this is a very cyclical sector, so investors were already looking for reasons to take profits,” Ben Ballinger, head of technology research at Quilter Cheviot, told CNBC.
“Google Turboquant’s innovation added to the pressure, but it is evolutionary, not revolutionary. It does not change the long-term demand landscape of the industry. In a market ready to avert risks, even incremental developments can be seen as a signal to defuse the situation.”
