(Bloomberg) — A two-day selloff in memory-chip stocks is revealing a split in the artificial intelligence trade, as Google touts a breakthrough that analysts say may curb demand for certain types of storage while leaving others largely unscathed.
On Friday, losses extended for the makers of flash memory — longer-term storage used to run AI — including Kioxia Holdings Corp., which have soared in past months. Meanwhile, leaders in high-bandwidth memory used in Nvidia Corp. AI accelerators stabilized, with Samsung Electronics Co. shares erasing all their losses and SK Hynix Inc. close to doing the same.
The market is beginning to realize that Google’s “TurboQuant” technique, which makes operating AI more efficient, poses more of a threat to the first group, analysts say.
“By shrinking memory usage and data movement, TurboQuant significantly improves inference efficiency,” Morgan Stanley analysts including Tiffany Yeh wrote in a note. “Yet it does not reduce demand for core memory like HBM.”
Investors in past months piled into the makers of flash and storage products on the belief that demand would skyrocket as artificial intelligence goes mainstream.
Sandisk shares were up more than 1,000% since the end of August, while Japan’s Kioxia had gained over 600%. They soared past Samsung, SK Hynix and Micron Technology Inc., the traditional leaders of the memory sector that became early darlings of the AI trade. Their lucrative high-bandwidth memory chips were a hot investment topic when the focus was on the training of large language models like ChatGPT.
Now, the flash memory firms are taking the brunt of an industry-wide selloff that began this week when investors cottoned on to Google’s breakthrough.
The US company said the algorithm can cut the amount of memory required to run a specific aspect of large language models by at least a factor of six, helping reduce the overall cost of artificial intelligence. Investors feared this could reduce the need for memory from hyperscalers like Meta Platforms Inc. — the largest data center operators. That would in turn push down the prices for components that are also used in smartphones and consumer electronics.
“Given the need to store model weights in graphics processing unit memory, high-bandwidth memory demand and DRAM produced by Micron will likely be unaffected,” Bloomberg Intelligence analyst Jake Silverman wrote in a note. “NAND demand will face deeper long-term effects.”
To be sure, tech stocks have broadly come under the microscope after big runups on the AI boom of the post-ChatGPT years. With inflation worries driven by the Iran war spurring caution over highly valued stocks, investors are taking profits on daily headlines.
“The hardware powering developments in the delivery and use of artificial intelligence technologies is a story that is secular and will play out over years and decades not over days and weeks,” said Ed Gomes, chief investment officer at SGMC Capital Pte. The selloff on TurboQuant is “short-term noise leading to very good buying opportunities.”
More stories like this are available on bloomberg.com
