Google’s new AI tech reduces memory use, raising industry concerns while experts predict AI growth will sustain memory demand despite fears
Google’s New York headquarters in Manhattan, New York. (Reuters)
As Google announced its ‘TurboQuant’ technology, which dramatically reduces memory usage to enhance artificial intelligence (AI) performance, speculation is rising that the super-boom in memory semiconductors may be coming to an end. Consequently, there have been repercussions, including a sharp drop in the stock prices of memory semiconductor producers Samsung Electronics, SK hynix, and the U.S.'s Micron.
However, experts assess such concerns as an overreaction. They explain that even if TurboQuant technology reduces memory demand, overall memory demand will not decrease because AI demand is growing at a faster rate.
In particular, some predict that if TurboQuant technology speeds up AI, it could fuel the acceleration of AI adoption, causing an explosive rise in memory demand. This would be a boon for Samsung Electronics and SK hynix.
According to industry sources on March 26, Google Research recently announced a new AI compression algorithm, TurboQuant.
TurboQuant is a technology that compresses the ‘KV cache’, a temporary memory for Large Language Models (LLMs), to a 3-bit level, reducing memory usage by about six times. Simply put, if TurboQuant technology is commercialized, it can be interpreted that memory demand will fall to one-sixth of its current level.
Google explained that the technology can compress data with fewer errors than existing algorithms and increases AI processing speed by eight times.
The core of TurboQuant technology consists of three main parts. First, it reduces memory usage by making the data itself smaller. For example, while a conventional AI might use 0.123456789 for a calculation, TurboQuant would use only 0.12.
Additionally, it does not store intermediate calculation values but recalculates them when needed. This means that while memory usage decreases, the amount of computation increases.
Finally, it transforms problems into a more compressed form. This is a method of solving the same problem with a smaller formula.
An IT industry official explained, “While existing AI used a lot of memory for fast processing, TurboQuant uses less memory and performs more calculations,” adding, “It’s about making data smaller, storing less, and calculating only what is necessary.”
Experts assess that while TurboQuant technology can reduce memory usage, concerns about a slowdown in memory demand are somewhat excessive.
The prevailing view is that although memory usage may decrease as Google announced, this will lead to advancements and the spread of AI technology, causing overall memory demand to actually increase.
The trend of AI competition shifting from simple repetitive tasks to agent-based AI that performs real-world tasks also dispels these concerns. The accelerating expansion of the AI agent market is expected to further fuel memory demand.
Kim Il-hyuk, a researcher at KB Securities, predicted, “As less KV cache memory is needed for the same inference, more information can be processed, which will expand the context window (the maximum range of text an AI can remember) of AI models.”
He continued, “If less KV cache memory is needed, not only can the context window be expanded, but multiple agents can also be run simultaneously.”
There is also the opinion that the concerns over TurboQuant technology are similar to the concerns about DeepSeek last year.
Researcher Kim pointed out, “This is a similar situation to when the market fell into a panic early last year on the news that DeepSeek had trained an AI model at one-twentieth of the cost.” He added, “Recalling that the AI market’s growth accelerated afterward, this can be seen as ‘Jevons’ paradox’ (a phenomenon where technological progress that increases the efficiency with which a resource is used tends to increase, rather than decrease, the rate of consumption of that resource).”
It is also suggested that Samsung Electronics and SK hynix will benefit from the commercialization of TurboQuant technology. This is because the increase in memory demand could act as a positive factor.
Lee Young-jin, a researcher at Samsung Securities, said, “If the TurboQuant algorithm is utilized, inference costs can be reduced, but demand will explode,” adding, “It will be possible to use long context windows and large-scale batching without sacrificing speed or quality.”
In Samsung Electronics’ case, an increase in orders for its semiconductor contract manufacturing (foundry) business is also expected. Samsung Electronics is manufacturing Nvidia’s ‘Groq3’ Language Processing Unit (LPU) chip.
The Groq3 LPU chip, unveiled by Nvidia CEO Jensen Huang at ‘GTC 2026’ on the 16th (local time), is a chip that divides roles with the Graphics Processing Unit (GPU) to enhance inference performance and efficiency, and Samsung Electronics is in charge of its manufacturing.
The LPU directly incorporates large-scale, high-speed memory (SRAM) on the chip, improving the bottleneck issue that occurred in existing GPUs using HBM based on the DRAM process. GPUs, which are strong in parallel processing that handles multiple operations simultaneously using thousands of cores, experience bottlenecks when exchanging data with external memory.
As AI technology advances with a focus on inference, memory demand is expected to expand overall, not just for High Bandwidth Memory (HBM) or large-capacity memory (DRAM). In other words, as demand for inference-focused AI chips based on SRAM increases, Samsung Electronics, which produces them, could see a rise in orders.
Researcher Lee Young-jin analyzed, “If inference costs decrease due to TurboQuant, it will enable long-context processing and the use of large-scale batching,” adding, “Overall inference demand will actually increase.”
Kim Eun-jin
댓글
0