The Korea Advanced Institute of Science & Technology (KAIST), South Korea’s premier science and technology institution, has issued a statement asserting that Google’s TurboQuant compression technology will, in fact, bolster demand for memory semiconductors. This comes as a direct response to investor concerns that led to a downturn in chip stocks earlier in the week.
KAIST’s perspective is particularly significant as Professor Han In-su, an electrical engineering expert at KAIST, played a key role in developing two of the three fundamental algorithms underpinning TurboQuant. He also maintains a position as a visiting researcher at Google.
Google unveiled TurboQuant on Tuesday, presenting it as a compression method capable of reducing the working memory required by AI models during inference – the process of a trained model responding to queries and generating outputs – by up to six times without any considerable loss of accuracy.
Following the announcement, memory chip stocks experienced a sharp decline, with both Samsung Electronics and SK hynix, the world’s leading memory chip manufacturers, witnessing drops. This was fueled by investor apprehension that the technological advancement would diminish demand for DRAM and high-bandwidth memory.
In a press release issued on Friday, KAIST clarified that the technology represents a shift from focusing solely on high-capacity to prioritizing high-efficiency computing. The university argues this will make AI more affordable and widely accessible, consequently “driving both qualitative advancement and quantitative expansion of memory demand at the same time.” While reduced memory requirements per model might initially appear to slow demand, KAIST contends that a lower barrier to entry will significantly broaden the spectrum of AI applications, ranging from on-device AI in smartphones and appliances to expansive data centers, ultimately creating new demand on a larger scale.
“The rapid growth of memory consumption as models become more powerful has long been cited as the biggest constraint,” professor Han said. “This research presents a new direction for effectively reducing that bottleneck while maintaining accuracy.”
Han co-developed QJL, a technique that compresses data to a single bit per data point while preserving the mathematical relationships AI models depend on, and PolarQuant, a compression method to be presented at the AISTATS 2026 conference in May. Both serve as core building blocks of TurboQuant.
mjh
