Storage Sell-off a Misstep? Morgan Stanley: Traditional Cycle Selling Logic Doesn't Apply, TurboQuant Panic is Cognitive Bias!

U.S. storage chip stocks have recently faced significant sell-offs due to demand concerns stemming from Google's TurboQuant memory optimization technology. However, Morgan Stanley believes the market is mispricing a structural shortage story using traditional cyclical stock logic.

In its latest report, the firm maintains its overweight rating on U.S. storage chip stocks, viewing the recent market sell-off as normal profit-taking rather than a signal of peaking cycles. Memory has transitioned from being a beneficiary of AI demand to a core bottleneck in AI infrastructure expansion, and tight supply and demand are expected to persist for years, with current valuations remaining attractive.

Morgan Stanley believes that market concerns over rising capital expenditures, demand destruction, and Google's memory optimization technology are being overinterpreted.

Regarding Google's TurboQuant memory optimization technology, Morgan Stanley analyst Joseph Moore believes it represents an incremental evolution, aiming to expand context windows and enhance model capabilities rather than cut storage demand, having virtually no material impact on the storage market.

Google's TurboQuant Technology Impact Overstated

The report specifically addresses the Google "TurboQuant" memory optimization technology that has recently triggered declines in storage stocks. Google introduced a data compression algorithm for Key-Value Cache (KV Cache) and recently announced its upcoming release. Widespread market rumors that "Google will reduce memory usage by six times" have put renewed pressure on storage stocks.

After communicating with industry insiders, Morgan Stanley believes this is an incremental technological evolution with virtually no material impact on storage demand. The report explains that KV Cache is typically stored in High Bandwidth Memory (HBM), whose capacity is fixed and cannot be changed; if additional cache offloading is required, it is usually transferred to 18TB LPDDR5 memory within the rack, which also has a fixed capacity. More efficient KV Cache usage might have some impact on Level 3 storage (SSD or HDD), but industry insiders generally reflect that related improvements are primarily used to expand context windows and enhance model capabilities rather than cut storage costs.

The report also cites Google's own case, noting that Gemini 1.5 Pro tested a 10-million-token context window with good results but was not released due to excessive inference costs. As similar optimization technologies reduce costs, they are expected to be used to support products with higher intelligence and computing power rather than reduce storage demand.

AI Consuming Capacity, Storage Becomes Bottleneck for Compute Expansion

Morgan Stanley's report states that this storage cycle differs fundamentally from historical patterns. The report points out that AI's consumption of DRAM has become so massive that it has caused supply shortages in other end markets—PC and smartphone production are both being constrained by insufficient memory supply.

For the past three years, the market widely believed there was slack in DRAM supply, but this buffer has disappeared. Demand for High Bandwidth Memory (HBM) from AI data centers continues to climb, and the complexity of HBM4 will further absorb capacity, while memory capacity demand related to the Rubin Ultra platform is expected to double after its launch next year. Meanwhile, demand for low-power DDR5 for racks and enterprise-grade storage is also experiencing explosive growth.

Morgan Stanley also notes that AI capital expenditure growth exceeds 50%, and as AI's share of overall spending continues to expand, this trend strengthens every year. While higher capital expenditure will indeed bring more supply, this is not a supply expansion driven by the traditional 3% to 5% growth in the smartphone and server markets; its magnitude is incomparable.

The report cites OpenAI's suspension of its Sora AI video generation app as evidence of insufficient computing power supply. Morgan Stanley believes this event clearly validates its investment theme—with token counts growing at double-digit rates weekly, computing power supply is severely insufficient, and demand far exceeds supply. Video generation is extremely storage- and HDD-intensive, and storage shortages may have been a factor in Sora's suspension.

The report points out that for commodities where demand is so strong that it cannot be fully met, it is difficult to maintain an excessively pessimistic stance at current valuation levels.

Micron Target Price $520, SanDisk Target $690

Morgan Stanley maintains its overweight rating on Micron Technology, with a target price of $520, corresponding to approximately 25 times through-cycle earnings per share (EPS) of $21, representing approximately 36% upside from the current stock price of $382.09. In a bull case scenario, the target price is $700 (28x through-cycle EPS of $25), and in a bear case, $240. The firm forecasts Micron's FY2026 GAAP revenue to reach $110.489 billion, with non-GAAP EPS of $59.36 and a non-GAAP gross margin of 77.8%.

For SanDisk, Morgan Stanley sets a target price of $690, corresponding to 23 times through-cycle EPS of $30, an upside of approximately 1.8% from the current stock price of $677.86. The bull case target price is $875, and the bear case is $350. The firm forecasts SanDisk's FY2026 GAAP revenue to reach $15.499 billion, with non-GAAP EPS of $41.09 and a non-GAAP gross margin of 60.2%.

The report notes that at current earnings levels, both companies' annualized free cash flow can reach 15% to 25% of their respective market capitalizations. Even considering cyclical factors, a period of high profitability sustained for more than two years is sufficient to support a substantial increase in stock prices.