Meet TurboQuant: The Korean Professor's Tech Behind Semiconductor Stock Plunge

Google's recently unveiled artificial intelligence memory optimization technology "TurboQuant" is shaking the semiconductor market. In simple terms, TurboQuant dramatically reduces the memory required when large language models (LLMs) like ChatGPT process longer and more complex queries. After Google announced TurboQuant, the memory semiconductor market trembled. On the 30th (local time), the semiconductor index on the New York Stock Exchange plunged 4.23%, while shares of Micron (9.88%) and SanDisk (7.04%) also fell sharply. Concerns spread that demand for memory semiconductors could decline. The new technology that shook the global semiconductor market was developed based on two papers written by Han In-su, a professor in the School of Electrical Engineering at KAIST, South Korea's top science and technology university. "The more complex and longer the problem to solve, the more memory is needed, and costs and response times increase accordingly," Han said during an online briefing held at KAIST on the 30th. "TurboQuant can be applied directly to a user's existing model to reduce memory while minimizing performance degradation." Below is a Q&A based on Professor Han's explanations. **Q. What is TurboQuant?** A. When AI generates a response, it doesn't just compute a single line in that moment. It must continuously remember and retrieve previous conversations, lengthy document contents, and conditions set by the user. The temporary storage space needed for this is called the "KV cache." The problem is that the longer the AI reads and thinks, the larger this storage space grows. The longer an AI model thinks, the more memory it consumes. TurboQuant is a technology that compresses this memory into a much smaller size for storage. Think of it as transferring a thick book into a smaller notebook while retaining as much of the key content as possible, rather than carrying the entire book around. **Q. Why is memory important in AI?** A. In the AI industry, "memory" is now as big an issue as "computation." Previously, greater computing power was considered the key to making models smarter, but recently, as long-context processing and agentic AI have grown, the memory burden has increased sharply. The longer the context, the more information needs to be stored and referenced, which can lead to higher costs, response delays, and deployment constraints. **Q. How effective has it been?** A. TurboQuant achieved a minimum 6x memory reduction while maintaining performance with no accuracy loss on major long-context benchmarks. Additionally, when 4-bit quantization was applied on Nvidia H100 GPUs, attention logit computation speed could improve up to 8x compared to 32-bit. This means it doesn't just reduce storage space — depending on the situation, it can also improve speed. **Q. Why did investors see this as bad news for semiconductors?** A. It's because if AI uses less memory, companies may not need to buy as many expensive memory semiconductors as before. After Google's announcement, memory-related stocks such as Micron and SanDisk were indeed rattled, and reports followed that Korean and Japanese memory stocks were also affected. In particular, once it became known that TurboQuant significantly reduces KV cache memory used in AI inference, the market immediately interpreted it as "even if AI demand grows, the pace of memory sales growth could slow." There are also concerns that high-performance hardware chips may be needed for TurboQuant. However, TurboQuant's computations can be performed on existing GPUs. I hope TurboQuant serves as an opportunity for the AI market, which has so far focused solely on hardware, to also pay attention to software. For AI to spread widely, software development such as algorithms and theoretical research are also important. **Q. Where can TurboQuant be applied?** A. The scope of application is broader than one might think. The first area mentioned is on-device AI. Currently, many AI functions operate by computing on servers and sending results back, but going forward, there will be more cases where AI runs directly inside smartphones, cars, robots, and wearable devices. However, since devices cannot use as much memory as servers, technology that delivers the same performance with less memory becomes crucial. TurboQuant can help spread on-device AI by reducing the memory burden needed for AI to remember previous context in such environments. Another important area is search and recommendation algorithms. Recent AI services don't just stop at generating sentences — they are evolving toward quickly finding and connecting what users want from vast collections of documents, products, videos, and posts. This process requires storing and comparing massive amounts of data in vector form, which ultimately makes memory efficiency critical. The presentation materials also explain that TurboQuant can be broadly extended beyond KV cache compression to large-scale vector search systems. There is also significant potential in retrieval-augmented generation (RAG). RAG is a method where AI retrieves necessary information from external documents or databases before generating an answer, and as documents increase, the efficiency of the search process determines performance and cost. Since TurboQuant is a technology that secures both high search precision and efficiency with fewer bits, it can be particularly useful in systems that combine search and generation, such as RAG. **Q. So is the memory semiconductor era really ending?** A. It's difficult to make that conclusion. In fact, there are strong arguments for the opposite view. When efficiency improves, more AI services can be run for the same cost, and longer contexts and more users can be handled. This means AI itself could spread more widely, potentially increasing overall semiconductor demand again. Recent market analyses have suggested that the stock price decline was an overreaction, and some analyses have also argued that AI data center investment and memory supply-demand dynamics remain strong in the long term. Put simply, it's similar to how better fuel efficiency in cars doesn't immediately kill the oil industry. Efficiency improvements can actually drive increased usage. **Q. How should readers interpret this technology?** A. Rather than seeing TurboQuant as a story about "AI consuming less," it's more accurate to view it as "a technology that reveals where AI is currently hitting bottlenecks." It means that the race to run models cheaply and quickly has become just as important as the race to build bigger models. The presentation materials also explain that TurboQuant can extend beyond KV cache compression to vector-based systems such as search, recommendations, and RAG. Ultimately, the significance of this technology lies not in simple compression, but in opening a path for AI services to spread more cheaply and broadly. **About Professor Han In-su** Han In-su is an assistant professor in the School of Electrical Engineering at KAIST, who has also served as a visiting researcher at Google Research since 2025. He has continuously participated in research spanning QJL, PolarQuant, and TurboQuant. QJL was released in 2024 and presented at AAAI 2025, while PolarQuant was released in 2025 and is scheduled for presentation at AISTATS 2026. He published the TurboQuant paper in 2025, which is scheduled for presentation at ICLR 2026. *HBM Demand Down 6x? Three Reasons the Market Is Completely Misunderstanding | Google TurboQuant*

Meet TurboQuant: The Korean Professor's Tech Behind Semiconductor Stock Plunge

Mine-Hunting Dolphins Return to Spotlight

Defense AI Will Determine Military Power

SK hynix Is Not the Benchmark for Bonus Standards

Gwanwoldang Return Marks Beginning, Not End

Parents Who Raise Top-Tier Children

Hyundai Motor Heads to Silicon Valley to Recruit Future Tech Talent

Smart Glasses Sales Hit 8.7 Million in 2024, Korea's Share Near Zero

SK Broadband to Train All Employees in AI, Target 400 "AI Performers"

Samsung Group ETF Surges 57% This Year on Chip, Battery Rally

Google, Marvell Launch Two AI Inference Chips; Pilot Production Next Year