Every time memory stocks go on a tear, somebody on Wall Street says the same thing: this time is different.
Memory used to boom and bust because demand was fickle and supply was rigid, and the argument is that AI changes that. The pros who have been doing this for decades disagree.
Memory chips have been on a tear since ChatGPT launched in late 2022, which kicked off a sustained demand wave for high-bandwidth memory, or HBM - the specialized chips needed to train large AI models.
Samsung and SK Hynix are the dominant suppliers, and the stock prices reflect it. Samsung is up 114% this year, SK Hynix is up 186%, U.S.-listed Micron has climbed 141%, and SanDisk is up 156%.
Those are full-year gains, and we are not even at the halfway mark.
The bull thesis goes like this: AI demand is structural, not cyclical, memory factories take years to build, so supply cannot catch up fast, companies have learned discipline, prices stay high for years, and margins stay fat.
Most of South Korea is along for the ride, with Samsung and SK Hynix together now over half of the Kospi - two stocks effectively running an entire country's stock market.
If you want a sharper read on the AI trade every weekday morning, Market Briefs covers it in five minutes - and includes a free investing masterclass when you sign up.
BlueBox Asset Management's William de Gale was on CNBC's Europe Early Edition last week, and his take was blunt.
"In the long run it's a pretty dreadful industry," he said. "I suspect that's still the case every time people make an argument that the memory cycle is gone, and it's now a long-term value-creating industry - just before it all goes horribly wrong."
That is the kind of line that ages well after a crash.
There is also a technology risk. Back on March 24, Alphabet's Google introduced something called TurboQuant, a compression technique that the company claims can shrink the memory footprint of large AI models by roughly six times.
Less memory needed means less demand for memory chips, and memory stocks dropped when the announcement landed.
Deutsche Bank told clients to "continue to brace themselves for continuous AI-related disruption," and other AI shifts are also forcing policymakers and investors to keep recalculating their assumptions.
Whether TurboQuant proves out is a separate question, but anything that makes AI models more efficient eats into the memory thesis.
Andrew Lapping, chief investment officer at Ranmore Fund Management, summed up the worry in a line: "A leopard does not often change its spots."
Standard Chartered's Steve Brice was in Korea recently and told CNBC's Squawk Box Asia his team was advising clients to take some profits and diversify away from Korean stocks.
Jon Cunliffe at JM Finn said today's share prices assume "prices stay high for a long time, companies stay very disciplined about not over-investing, and profit margins remain much better than in the past." In plain English: a lot has to go right.
Nomura is still bullish, and the firm sees Samsung climbing another 20% and SK Hynix doubling again over the next year. But cycle-aware investors know what those calls usually look like near a top.
The chips might still be hot. The history is too.
If you want to see how stories like this play out for the broader market, join Market Briefs every morning and get the 45-minute investing course tossed in for free.
