A single artificial intelligence paper from Google has sent shockwaves through the global semiconductor market. The study, interpreted as suggesting AI could use less memory, triggered sharp declines in shares of Samsung Electronics (005930.KS) and SK hynix (000660.KS).
Google's 'TurboQuant' Sends Korean Chip Stocks Reeling
On May 25, Google Research published a paper on an AI memory efficiency technology called "TurboQuant."
Markets reacted immediately. On May 26, Samsung Electronics fell more than 4% on the Korea Exchange, while SK hynix plunged around 6%. On May 27, Samsung Electronics slipped 400 won (-0.22%) to close at 179,700 won, and SK hynix dropped 11,000 won (-1.18%) to finish at 922,000 won.
Chae Min-suk, a researcher at Korea Investment & Securities, said in a report that day that the sell-off stemmed from "an interpretation error caused by confusing the roles of memory capacity and memory bandwidth."
The bottleneck in AI inference is determined not by insufficient memory capacity but by memory access speed and data transfer efficiency, Chae explained.
"TurboQuant should be understood as a technology that partially alleviates this bottleneck, improving GPU efficiency so that more tokens can be processed with the same GPU resources," Chae said. The claim that TurboQuant compresses the KV cache (Key-Value Cache, temporary storage for intermediate computation values) by up to six times does not mean reducing the required memory capacity itself, but rather significantly lowering the data size occupied by the KV cache and the resulting memory access burden, the researcher explained.
The Name Alone Is Daunting — What Exactly Is TurboQuant?
The core of TurboQuant is not simply a technology that reduces memory. In short, it is a data compression technology.
The main culprit behind bottlenecks that limit AI performance is not memory capacity but memory bandwidth and access speed. AI stores data in a space called the "KV cache (Key-Value Cache)" to remember conversations and context. The problem is that as this data accumulates, memory usage increases explosively.
To put it simply, data can be thought of as cars and memory bandwidth as a tunnel. Previously, as cars (data) increased, the tunnel (bandwidth) had to be continuously widened, which drove explosive growth in demand for HBM (high bandwidth memory).
However, the real issue was not the size of the tunnel but "how quickly the cars pass through" — the speed at which memory is read. GPUs compute extremely fast, but when the speed of fetching data from memory cannot keep up, computing units enter an idle state. According to industry research, a significant portion of AI computation is consumed by this "wait time," creating a bottleneck where data processing backs up.
TurboQuant addresses this bottleneck in a different way. It compresses the KV cache data used by AI by up to six times. Rather than reducing memory capacity itself, it reduces the amount of data that must be read from memory.
By simplifying data structures and minimizing information loss, the compression reduces the time GPUs spend waiting for data, allowing more computations to be processed with the same GPU resources.
Foreign Media vs. Securities Firms — Crisis or Opportunity?
Foreign media were quick to assess the technology's impact.
U.S. IT outlet VentureBeat called it "a technology that could change data center cost structures through software optimization alone." TechCrunch also noted it "could reduce AI operating costs by cutting the KV cache by up to six times." Both outlets pointed out, however, that "it is still an early-stage technology."
The investment community also acknowledged the technology's significance while noting limits in its scope of application and level of verification.
Chae Min-suk of Korea Investment & Securities said, "It is difficult to conclude that memory demand will decline just because an efficiency improvement technology has emerged." She added that "actual GPU utilization improves and throughput — the number of tokens that can be processed per unit of time — increases, lowering cost per token."
Chae also noted that "the experiments were conducted primarily on relatively small models with around 8 billion parameters, and whether the same effects can be replicated on large models with tens of billions of parameters used in actual industry has not yet been verified."
Global investment bank Morgan Stanley also weighed in, saying "this technology does not reduce total memory demand but could serve as a catalyst that expands the AI market itself," adding that "it could rather be seen as a buying opportunity."
