Google's TurboQuant unlikely to weaken memory demand: analysts

An introduction of Google's TurboQuant technology published on Google Research website / Captured from Google Research

Google’s announcement of TurboQuant is weighing on the share prices of memory companies, as the technology is expected to cut artificial intelligence (AI) models’ memory usage to about one-sixth of current levels.

For analysts, however, concerns on the memory chip demand may be overblown, as they noted that even if memory demand per model declines due to TurboQuant, overall demand for AI continues to grow at a faster pace, keeping the broader memory market on a solid growth trajectory.

Announced on Tuesday by Google Research, TurboQuant is a compression technology designed to maximize AI efficiency. The gist of the technology is compressing an AI model’s key value cache memory (KV cache) to just 3 bits, cutting its size by more than sixfold.

KV cache is an AI model’s short-term memory, where it stores keys and values already calculated so it can generate the next words faster. A sixfold reduction in KV cache size effectively lowers memory usage to about one-sixth of current levels, making similar performance possible even with only one-sixth of the required memory.

As AI services become more advanced, the AI industry has been seeking ways to address the growing burden of KV cache, while the demand of AI memory chips such as high-bandwidth memory (HBM) has remained strong. This has even led to discussions on using advanced NAND storage to support HBM.

As expectations grew that TurboQuant’s commercialization could reduce memory demand, shares of memory companies came under pressure. Samsung Electronics shed 4.71 percent Thursday, while SK hynix fell 6.23 percent. Micron also declined 6.97 percent on the same day. Samsung Electronics and SK hynix extended their losses Friday, falling 0.22 and 1.18 percent, respectively.

This picture taken on Oct. 22, 2025 shows a mockup of a chipset featuring Samsung Electronics' high-bandwidth memory technology on display during the 2025 Semiconductor Exhibition in Seoul. AFP-Yonhap

While TurboQuant could reduce memory use, analysts said concerns over a slowdown in overall memory demand are excessive.

“There have been efforts to improve AI models to optimize chip usage, but more efficient models tend to lower overall costs and, in turn, drive greater demand for AI computing,” Samsung Securities analyst Lee Jong-wook said. “Rather than reducing semiconductor demand, such optimized models are being used to deliver higher-performance AI services with the same chip resources.”

Lee cited the Jevons paradox, which occurs when increased efficiency will increase the use of a resource. In AI computing, the paradox means improvements in AI efficiency reduce the cost of computing, ultimately driving much higher demand.

“Factors that could lead to a decline in AI memory demand are likely to emerge when AI capabilities get into a stalemate, such as a slowdown in service improvements or weakening competition between AI model developers,” Lee said.

“As long as AI companies compete on performance rather than cost, optimization will not weigh on semiconductor demand.”

Korea Advanced Institute of Science and Technology professor Han In-su, who participated in the TurboQuant project, also said that it will be “used as a key technology that can improve AI model's operation efficiency,” rather than weakening memory demand.

Hana Securities analyst Kim Rok-ho also said TurboQuant’s commercialization will improve cost efficiency for data center operators, driving up the overall demand for AI memory chips.

“Compression technologies are not new, and it remains uncertain whether they will be widely adopted across the industry,” Kim said. “Even if such technologies become more widely used over the mid to long term, it will lower memory cost barriers, expanding overall AI use. There are limited chances of decline in demand for DRAM and storage.”