Is the Era of Memory Semiconductors Over? Google’s “TurboQuant” Sends Shockwaves

According to The Financial News, after Google announced "TurboQuant," a technology that sharply reduces memory usage while improving artificial intelligence (AI) performance, some began to predict that the super-boom in memory semiconductors might be coming to an end. In response, the share prices of memory semiconductor producers such as Samsung Electronics, SK hynix, and U.S.-based Micron Technology plunged, sending ripples through the market.
However, experts argue that such concerns are an overreaction. They contend that even if TurboQuant reduces memory demand per model, overall memory demand will not fall because AI demand is growing at an even faster pace. Some analysts even forecast that by speeding up AI, TurboQuant could accelerate the spread of AI and drive an explosive increase in memory demand. This could ultimately turn into a positive factor for Samsung Electronics and SK hynix.■ What is Google’s "TurboQuant" for optimizing AI models?According to industry sources on the 26th, Google Research recently unveiled a new AI compression algorithm called TurboQuant.
TurboQuant compresses the "KV cache," a temporary memory store used by Large Language Models (LLM), down to 3 bits, cutting memory usage by about a factor of six. Taken at face value, if TurboQuant were fully commercialized, memory demand could be interpreted as dropping to roughly one-sixth of current levels. Google explained that the technology can compress data with fewer errors than existing algorithms and increase AI processing speed by up to eight times.
TurboQuant’s core ideas can be grouped into three main points. First, it shrinks the data itself to reduce memory usage. For example, if a conventional AI calculation used the value 0.123456789, TurboQuant would work with something like 0.12 instead. Second, it avoids storing intermediate results and instead recomputes them when needed. This reduces memory usage at the cost of increased computation. Finally, it transforms the problem into a more compressed representation, essentially solving the same problem with a smaller mathematical expression.
An IT industry insider explained, "Traditional AI systems used a lot of memory to process tasks quickly, whereas TurboQuant uses less memory and performs more computation," adding, "It makes the data smaller, stores less of it, and only computes what is necessary."
On the day of the announcement, Samsung Electronics’ share price closed at 180,100 won, down 4.71% (8,900 won) from the previous day, while SK hynix finished trading at 933,000 won, a drop of 6.23% (62,000 won).■ "It will actually increase memory demand... concerns are overblown"Experts acknowledge that TurboQuant can reduce memory usage, but they say worries about a slowdown in memory demand are somewhat excessive.
They note that while memory usage per model may decline as Google claims, this will in turn drive advances and wider adoption of AI, leading overall memory demand to rise rather than fall. The ongoing shift in AI competition—from simple repetitive tasks to AI agents that perform real-world work—also helps dispel these concerns. The rapid expansion of the AI agent market is expected to further fuel memory demand.
Some observers say the current anxiety over TurboQuant is reminiscent of the fears that surfaced when DeepSeek was unveiled last year.■ "Could be a boon for Samsung Electronics and SK hynix"There is also a view that Samsung Electronics and SK hynix could actually benefit from the commercialization of TurboQuant, as stronger memory demand would work in their favor. For Samsung Electronics in particular, an increase in orders for its semiconductor foundry business is also anticipated. Samsung Electronics is currently manufacturing Nvidia’s Groq 3 Language Processing Unit (LPU) chips.
The Groq 3 LPU chip that Nvidia CEO Jensen Huang unveiled on the 16th (local time) at NVIDIA GTC 2026 (NVIDIA GPU Technology Conference 2026) is designed to share roles with the Graphics Processing Unit (GPU) to enhance inference performance and efficiency, and Samsung Electronics has been tasked with its production.
The LPU integrates large amounts of ultra-high-speed Static random-access memory (SRAM) directly on the chip, alleviating the bottlenecks seen in conventional GPUs that rely on High Bandwidth Memory (HBM) built on Dynamic Random Access Memory (DRAM) processes. GPUs excel at parallel computation, using thousands of cores to handle many operations simultaneously, but they suffer from bottlenecks when exchanging data with external memory. As AI technology becomes more inference-centric and sophisticated, memory demand is expected to expand across the board—not only for HBM and large-capacity DRAM, but also for other types of memory. In other words, as demand grows for inference-focused AI chips based on SRAM, Samsung Electronics, which produces such chips, could see a rise in orders.

june@fnnews.com Lee Seok-woo Reporter