Memory chip stocks were big winners in 2026 — until lately. What to do now
Memory chip stocks that dominated semiconductor gains through much of 2026 are suddenly losing momentum, as a breakthrough from Alphabet ‘s Google raises questions about the durability of artificial intelligence-driven demand. Stocks including former leaders Micron Technology and Sandisk started selling off at the end of last week, but losses accelerated this week after Google unveiled a new compression technique , dubbed TurboQuant, that could cut memory requirements for AI large language models sixfold. The slide reflects investor concern that one of the industry’s most powerful tailwinds — surging demand for high-bandwidth memory tied to AI — may prove weaker than previously thought. The development “highlights emerging efficiencies that could reduce memory intensity in AI workloads,” Evercore ISI said. “The market may be beginning to price in a more favorable cost/supply backdrop.” That could ease cost pressures for hardware buyers such as server makers, with Evercore ISI pointing to potential upside for companies including Dell Technologies and Hewlett Packard Enterprise if lower memory requirements improve system economics. Before this latest sell-off, memory stocks were some of the best performers in the market this year, with Sandisk almost tripling, up 170%, and Micron rising more than 40%. DeepSeek moment Morgan Stanley described the development as “another DeepSeek moment ,” arguing the read-through is broadly positive for hyperscalers and model platforms given the improved economics of AI deployment. Cheaper per-query costs, particularly in long-context and retrieval-heavy applications, could drive stronger returns on investment and wider adoption. At the same time, Morgan Stanley said the implications for computing and memory are “neutral” in the near term. While compression reduces memory traffic and GPU-hours required per workload, lower costs per token could spur greater usage, potentially offsetting some of the demand impact. UBS urged calm, saying that TurboQuant isn’t necessarily a turning point but is instead a reflection of continued work across the industry. The bank maintained its base case that DRAM pricing will peak around mid-2027, with equity markets likely to discount that turning point roughly a year earlier — meaning in a few months. “While the situation is still fluid, in our view, TurboQuant does not represent a sudden technological inflection but rather an extension of work that has been public since April 2025,” UBS said in a note to clients. Deeper shift Others on Wall Street doubt the Micron and Sandisk selloff marks a deeper shift. Analysts at Mizuho, for example, urged investors not to overreact, characterizing the move as typical profit-taking after a strong rally. “While painful and annoying, I am here to say do not panic … this is normal profit-taking and consolidating after a strong rally across the sector. GOOG compression tech white paper is noise. It will blow over sooner vs later,” the Mizuho analyst wrote. The firm added that skepticism and positioning resets are healthy after crowded bullish trades, adding it expects memory names to trade higher over the next three to six months. — CNBC’s Michael Bloom contributed reporting.