On Apr. 1, the securities industry said concern spreading in the stock market over "TurboQuant" "looks excessive."
On the 24th, Google introduced "TurboQuant" on its blog. TurboQuant technology is a data compression algorithm method that reduces the memory usage of artificial intelligence (AI) models. Analysts say it could affect AI investment and demand for memory semiconductors. Right after the announcement about TurboQuant technology, share prices in the memory sector plunged.
Kim Young-geon, a researcher at Mirae Asset Securities, said, "We judge that this week's price adjustment served as an opportunity to resolve profit-taking demand amid an uncertain macro environment," adding, "Concerns about TurboQuant look excessive."
As grounds, he cited: ▲ the latest announcement is not entirely new ▲ some aspects of the technological advance appear to be exaggerated ▲ the purpose of efficiency technologies is to extract maximum performance from the same resources, not to reduce the total amount of resources投入.
First, what Google disclosed on its blog this time is not a new technology announcement. It is content from a paper disclosed in Apr. last year, and the company said it was resurfaced ahead of a presentation at the International Conference on Learning Representations (ICLR) 2026 in the AI field in Apr. this year.
Kim said, "There are also some exaggerated aspects of the technological advance, as the '8x speed improvement' emphasized in the blog is a result in a specific operation, not overall inference throughput."
He added that the models tested were limited, so it is currently unverified whether the same effect can be expected in commercial models with tens of billions of parameters.
Finally, he emphasized that the purpose of efficiency technologies is performance maximization, not resource saving. Kim said, "If compute resources and performance are proportional in AI infrastructure, the goal is to extract maximum performance from the same resources, and we judge that the objective is not to reduce the total amount of resource投入."
He went on to explain, "Google developed TurboQuant not to buy less memory, but to process more tokens with the same H100 and implement longer context to outperform competitor models."
At the same time, he said the target price for Samsung Electronics is maintained at 300,000 won and the target price for SK hynix at 1,540,000 won. He also judged that it is time to approach from a perspective of holding continuously and making new purchases if share prices fall.
Kim said, "With a tight supply-demand environment continuing, there is a high possibility that prices will remain at elevated levels due to buyer-side waiting demand," adding, "Even if the upward pace of memory prices slows, if prices stay around current levels in the mid to long term, we forecast a high-profit structure with return on equity (ROE) of more than 50% will continue."