Google has recently unveiled its new TurboQuant compression method, which is significantly affecting the stock prices of memory chip manufacturers in both Asia and the United States. This technology, which the company claims can reduce the memory needed to run large language AI models by six times, has raised concerns among investors about a potential decline in demand for memory chips, a crucial component of AI infrastructure.
On Thursday, shares of South Korea's SK Hynix and Samsung Electronics, the world's two largest memory producers, fell by 6% and nearly 5%, respectively. Japanese company Kioxia, known for its flash memory products, also saw its stock drop by almost 6%. In the U.S., shares of Sandisk and Micron Technology experienced declines, continuing in pre-market trading on Thursday.
TurboQuant works by reducing the amount of cache used for key-value pairs, which store the outcomes of previous computations within AI models. The goal is to enhance the efficiency of these models, a critical objective for leading research labs. Investors are worried that a decreased need for memory during AI training could slow the growth in demand for specialized chips.
Matthew Prince, CEO of Cloudflare, pointed out that there are still significant opportunities for optimization in AI inference regarding speed, memory usage, energy consumption, and resource efficiency in multi-user environments. Ray Wang, a memory market analyst at SemiAnalysis, did not connect Google's research with an imminent decrease in chip demand. He stated, "When you eliminate a bottleneck, you make the hardware platform for AI more efficient. As models become more powerful, better hardware is required to support them."
Despite the stock declines, the memory market is still bolstered by high demand and supply shortages. This combination has driven memory prices to unprecedented levels, supporting profitability for companies like Samsung, SK Hynix, and Micron. Over the past year, Samsung's stock has surged nearly 200%, while Micron and SK Hynix have risen over 300%. Analysts suggest that the stock movements this week were largely influenced by profit-taking.
Ben Barringer, a technology sector research leader at Quilter Cheviot, noted that companies in the memory sector had already shown robust growth and that the industry is known for its cyclical nature, prompting investors to seek opportunities to reduce their positions. He emphasized that Google's TurboQuant innovation has intensified pressure but represents an evolutionary rather than revolutionary shift, not altering the long-term demand outlook for the industry. In a market already tuned to risk reduction, even incremental advancements can be perceived as signals to trim investments.
This development highlights potential shifts in the memory chip market and poses new challenges for competitors who may need to adapt to changing demands fueled by advancements in AI technologies.
Informational material. 18+.
