New AI algorithm targets inference memory bottleneck; Analysts say lower costs could ultimately drive chip demand
Google’s new AI memory compression algorithm, TurboQuant, sent memory chip stocks sliding this week, raising concerns over potential demand disruption. However, analysts suggest the long-term outlook remains positive, as the algorithm could unlock new application possibilities and ultimately boost demand for memory chips.
The tech giant recently unveiled TurboQuant, an algorithm designed to minimize the working memory needed for artificial intelligence inference. It achieves this by compressing the key-value cache, a crucial memory component that enables large language models to maintain context without repetitive computations.
Following the announcement on Tuesday, shares of Samsung Electronics and SK hynix experienced declines on Thursday, falling 4.71 percent and 6.23 percent, respectively. US-listed memory companies also saw significant drops, with the Philadelphia Semiconductor Index decreasing by 4.8 percent on Thursday. Nvidia shares fell by 4.2 percent, while Micron Technology dropped 6.97 percent and Sandisk plunged 11.02 percent.
Google states that TurboQuant can reduce AI memory usage by up to sixfold without sacrificing accuracy. Moreover, it claims to deliver processing speeds up to eight times faster than Nvidia’s H100 graphics processing unit under certain conditions.
According to Google, TurboQuant restructures data into a simpler form, allowing for efficient compression while preserving most of the core information. An additional step is then applied to correct minor errors resulting from the initial compression, ensuring accuracy in tasks like language processing and search.
The company anticipates that TurboQuant could eventually be integrated into its Gemini AI model and search devices.
Cloudfare CEO Matthew Prince hailed the breakthrough as Google’s “DeepSeek moment,” drawing a parallel to how China’s DeepSeek demonstrated the feasibility of building high-performance AI models at significantly reduced costs.
The advancement has sparked concerns that improved memory efficiency may lessen the demand for DRAM and high bandwidth memory chips, which are essential components for AI workloads.
However, analysts believe the market’s reaction might be an overreaction, citing uncertainties surrounding commercialization and broader demand trends.
“At this point, it remains a research paper that needs further validation, so its impact on actual memory demand seems minimal,” stated Yoo Hoi-jun, a professor at the KAIST Graduate School of AI Semiconductor.
“Efforts to reduce KV (key-value) cache have been ongoing, and cutting KV cache doesn’t necessarily mean lower HBM usage.”
Kim Rok-ho, an analyst at Hana Securities, shared a similar perspective, emphasizing that it is unclear how Google’s simulation results will translate into real-world implementation and the speed at which the technology can be scaled.
“Memory price forecasts for the second and third quarters have been revised upward following the first quarter. With supply shortages anticipated to continue throughout the year, the possibility of stronger-than-expected DRAM price increases remains,” Kim said.
Others highlighted the potential for a long-term recovery, noting that reduced memory requirements could decrease the costs associated with running AI systems and expedite adoption – ultimately driving up chip demand.
“Low-cost AI technologies like TurboQuant are likely to lower the barriers to adoption and significantly increase overall demand,” said Kim Dong-won, head of research at KB Securities.
“This would lead to increased computing workloads and higher memory content, ultimately positioning memory-chip manufacturers as the primary beneficiaries of the expanding AI ecosystem.”
Google is scheduled to present more details on TurboQuant at the International Conference on Learning Representation 2026 in Rio de Janeiro, commencing on April 23. The research was a collaborative effort between researchers at Google and Google Deepmind, along with Han In-su, an assistant professor of electrical engineering at the Korea Advanced Institute of Science and Technology.
