Google’s TurboQuant slashes AI inference memory use, but analysts say efficiency gains may boost long-term demand via a rebound effect
Google personnel are introducing ‘TurboQuant.’ (Photo captured from Google site)
After Google announced its ‘TurboQuant’ technology, which significantly reduces memory semiconductor usage, stock prices of semiconductor companies, including the top three memory makers, fluctuated wildly, plummeting before rebounding.
Over four trading days (March 26-31), Samsung Electronics fell by 11.53% and SK Hynix by 18.89%, but on April 1, they recovered by 13.44% and 10.66%, respectively. On the New York stock market, Micron Technology’s stock price also fell by 9.88% on March 30 (local time) alone, before closing up 4.98% the next day. Experts forecast that the top three memory makers still have ample room for stock price growth, as the emergence of TurboQuant could actually boost semiconductor demand in the long term.
On March 24, Google published a paper titled “‘TurboQuant: Redefining AI Efficiency with Extreme Compression’” on the Google Research Blog. TurboQuant is a technology that enhances memory efficiency used in Artificial Intelligence (AI) by resolving memory bottlenecks that occur during the inference process of Large Language Models (LLMs).
AI technologies like ChatGPT and Gemini are evolving from training-based to inference-based models. To infer and generate responses, contextual data, such as conversations with users and search results, are stored in a temporary repository called ‘KV cache’. As conversations accumulate, the memory usage occupied by the KV cache also increases proportionally. In such cases, bottlenecks occur when retrieving data from temporary storage. TurboQuant is a technology that reduces memory overload by reducing its size by a factor of six while maintaining data accuracy. Google is officially scheduled to present detailed research results related to TurboQuant at the international AI academic conference ‘ICLR 2026’, to be held in Brazil from April 23-27.
Professor Han In-soo of KAIST’s School of Electrical Engineering, who participated in the development of the core algorithm used in TurboQuant, stated at an online research presentation on March 30, “This is an example of software influencing the hardware market,” adding, “The biggest strength of TurboQuant technology is its applicability to any AI model.” As such, semiconductor stock prices plummeted on observations that demand for semiconductors, including High Bandwidth Memory (HBM) where KV cache is stored, would decrease with the commercialization of TurboQuant technology.
However, a prevailing view among the semiconductor industry and experts is that this sharp decline is a market misunderstanding. They argue that the so-called ‘Jevons’ paradox’ applies, where increased technological efficiency leads to lower resource costs and an explosive increase in overall usage. Distinguished Professor Park Jae-geun of Hanyang University’s Department of Electronic Engineering explained in a phone interview, “If memory efficiency for inference improves, more people will jump into developing inference AI, and semiconductor demand will also grow together.” Sean Kim, an analyst at Morgan Stanley, also forecasted to Bloomberg on March 26, “If TurboQuant can lower memory requirements without performance degradation, costs will significantly decrease, increasing the profitability of AI adoption,” adding, “Lower costs will also increase product adoption demand, benefiting memory manufacturers in the long term.”
The research results recently disclosed by Google via its blog re-highlighted a paper published by the same research team in April of last year. Lee Hyung-soo, CEO of HSL Partners and a semiconductor investment expert, said in a phone interview, “Since TurboQuant technology was already published as a paper a year ago, it can be interpreted that the market created a basis for profit-taking,” adding, “Even if 1/6 compression is possible in situations where KV cache will surge in the future, such as with Agent AI and Physical AI, additional semiconductor demand is guaranteed.” Kim Young-geon, a researcher at Hana Securities, pointed out in a report on March 30, “This stock price adjustment was an opportunity for some profit-taking needs to be met amidst macroeconomic uncertainty,” and “Considering that Nvidia also submitted a paper in a similar direction, this reaction is more a matter of timing than content.”
This market plunge triggered by TurboQuant is comparable to the ‘DeepSeek Shock’ in January of last year. At that time, when DeepSeek, built with low-spec chips, emerged in China, AI-related stocks suffered a significant impact, with Nvidia falling by approximately 17% in a single day. However, within two weeks, Nvidia’s stock price recovered to its pre-DeepSeek levels and went on to set new record highs. Kim Dong-won, a researcher at KB Securities, explained in a report on March 27, “Both TurboQuant and DeepSeek are significant as software optimization technologies for implementing low-cost, high-efficiency AI, but considering the anticipated pace of AI data center investment expansion over the next five years, there are practical limitations to meeting the surging AI demand with only such technologies.”
Although the stock prices of the top three memory makers plummeted, the securities industry is maintaining target stock prices, reflecting the global memory boom. KB Securities and Kiwoom Securities presented target prices for Samsung Electronics of 320,000 won (approximately $213) and 260,000 won, respectively, on March 30. KB Securities announced the next day that its target price for SK Hynix remained unchanged at 1,700,000 won. Researcher Kim Dong-won explained, “Similar to the case where coal demand actually increased after steam engine efficiency improved in the 1860s, and paper usage surged after the introduction of the internet in the 1900s, a ‘rebound effect’ where efficiency improvements actually expand demand is expected to occur.” He added, “The short-term stock price adjustment due to heightened Middle East tensions and the Google TurboQuant issue presents a meaningful buying opportunity.”
Kim Eun-jin
댓글
0