Google TurboQuant raises new questions about demand for memory chips

In short: Google has described a method called TurboQuant that could reduce how much memory AI systems need, and experts disagree on whether that will hurt or boost demand for memory chips.

What's going on

Google Research outlined TurboQuant in a recent blog post. The idea is to make AI systems use less memory, which is a key part inside the computers that run popular AI services.

After the post, shares of South Korean chipmakers Samsung Electronics and SK Hynix fell sharply. Both companies make high-bandwidth memory, which is a type of very fast memory used in AI servers.

TurboQuant focuses on something called a “key value cache,” which is like the AI’s short-term notepad that helps it remember what was said earlier in a conversation. Google’s approach compresses that notepad, then rebuilds it when needed, with little reported loss in accuracy. Researchers say this could cut memory use by as much as six times and lower the cost of running some AI models.

Samsung’s latest profit guidance, which points to a very strong quarter, eased some investor worries. It suggests memory chips are still in tight supply for AI work.

Some analysts argue that if TurboQuant makes AI cheaper to run, more people and businesses will use AI more often. That could increase total demand for computing power and memory, even if each individual AI task needs less. This is similar to how a more fuel-efficient car can lead people to drive more because each trip costs less.

What to watch

TurboQuant is still mostly a concept described in a blog post. Its real impact will be clearer after it is presented at an AI research conference in late April, and other people can test whether it works well outside Google. Investors will also watch whether big tech companies sign longer supply contracts with chipmakers, which could make chip demand steadier over time.

Source: Financial Times