South Korea’s KAIST, the nation’s main science and expertise college, stated the compression expertise behind Google’s TurboQuant will strengthen reminiscence semiconductor demand moderately than weaken it, pushing again in opposition to the investor fears that drove chip shares decrease this week.
The college’s evaluation carries uncommon weight as KAIST electrical engineering professor Han In-su co-developed two of TurboQuant’s three underlying algorithms and stays a visiting researcher at Google.
Google launched TurboQuant on Tuesday, describing a compression methodology that may scale back the working reminiscence AI fashions want throughout inference, when a skilled mannequin processes queries and generates output, by as much as sixfold with out significant accuracy loss.
Memory chip shares fell sharply within the days that adopted, with Samsung Electronics and SK hynix, the world’s two largest reminiscence chipmakers, each declining as buyers apprehensive the breakthrough would dampen demand for DRAM and high-bandwidth reminiscence.
In a press launch on Friday, KAIST stated the expertise marks a shift from high-capacity to high-efficiency computing that may make AI cheaper and extra accessible, “driving both qualitative advancement and quantitative expansion of memory demand at the same time.” While decreased reminiscence necessities per mannequin could seem to gradual demand within the quick time period, the college stated a decrease threshold would dramatically widen the vary of AI purposes, from on-device AI in smartphones and home equipment to large-scale knowledge facilities, finally creating new demand at a bigger scale.
“The rapid growth of memory consumption as models become more powerful has long been cited as the biggest constraint,” professor Han stated. “This research presents a new direction for effectively reducing that bottleneck while maintaining accuracy.”
Han co-developed QJL, a way that compresses knowledge to a single bit per knowledge level whereas preserving the mathematical relationships AI fashions rely on, and PolarQuant, a compression methodology to be offered on the AISTATS 2026 convention in May. Both function core constructing blocks of TurboQuant.