On March 25, $Alphabet(GOOGL)$ unveiled a new AI memory compression algorithm, TurboQuant, claiming it can reduce memory requirements during large language model inference by sixfold while increasing computational speed by eightfold, all without sacrificing accuracy. Specifically, as AI models grow more powerful, context windows continue to expand, and key-value (KV) cache storage grows geometrically, becoming a core bottleneck for both performance and cost. TurboQuant leverages PolarQuant and error correction (QJL) to maintain full model accuracy and a 100% retrieval recall rate.
The technology can be directly deployed on existing AI systems, raising market concerns that demand for memory chips could be weakened. As a result, shares of $Micron Technology(MU)$ , $SanDisk Corp.(SNDK)$ , $Western Digital(WDC)$ , and $Seagate Technology PLC(STX)$ declined sharply.
The selloff in memory stocks also spread to other semiconductor companies. $NVIDIA(NVDA)$ fell 4.2%, $Taiwan Semiconductor Manufacturing(TSM)$ dropped 6.2%, and $ASML Holding NV(ASML)$ declined 4.6%.
Semiconductor ETFs were also hit, with the largest, $VanEck Semiconductor ETF(SMH)$ , falling over 4.5%, $iShares Semiconductor ETF(SOXX)$ down more than 4.7%, $First Trust Nasdaq Semiconductor ETF(FTXL)$ dropping over 4.6%, and PSI plunging 5.9%.
This selloff recalls January 2025, when DeepSeek emerged. Its development cost was only a few million dollars, yet its performance rivaled OpenAI’s models. Most notably, DeepSeek did not rely on Nvidia’s most advanced GPUs.
At that time, concerns about potential impacts on Nvidia’s GPU demand triggered a wave of selling. On January 27, 2025, Nvidia plunged 17%, wiping out $589 billion in market value in a single day, the largest one-day loss in U.S. stock market history.
Compared with Nvidia’s case, the market reaction to Micron has been more measured. The day after Google announced TurboQuant, Micron fell only 3.4%. However, amid escalating tensions in the Middle East and a broader tech selloff, Meta dropped 8%, Microsoft hit a one-year low, and the Nasdaq fell 2.4%. Against this backdrop of rising risk aversion, Micron declined 7% yesterday—significant, but still far less severe than Nvidia’s earlier drop.
According to Wall Street analysts, TurboQuant is unlikely to disrupt the structural tightness in memory supply. Morgan Stanley analyst Joseph Moore noted that the claim of a sixfold reduction in memory usage may be overstated, as the optimization mainly targets KV cache rather than total memory demand. He added that the recent decline in memory stocks is partly driven by this overinterpretation.
The report further highlighted that Google’s latest models, Gemini 3 and 2.5 Pro, already support context windows of up to one million tokens, while Gemini 1.5 Pro has demonstrated up to ten million tokens in testing but was not released due to high inference costs. As technologies like TurboQuant reduce costs, future AI models may adopt larger context windows and more complex computations, ultimately increasing overall demand for compute and memory.
Based on this view, Morgan Stanley reiterated its “Overweight” rating on Micron (MU.US) and SanDisk (SNDK.US). Analysts emphasized that the key bottleneck in AI development remains insufficient memory supply, particularly in DRAM, where rapid data center demand growth has already crowded out memory availability for PCs and smartphones.
Moore stated, “Our supply chain checks show no indication that memory or storage demand is declining.”
Similarly, Lynx Equity Strategies noted in a client report that Google’s proposed solution is unlikely to reduce demand for memory and flash storage over the next three to five years due to extremely tight supply conditions.
Taken together, the recent pullback in memory stocks appears more like a normal correction following strong gains, compounded by negative headlines and geopolitical tensions, rather than a breakdown in the core supply-demand imbalance.
That said, it is worth noting that since the beginning of the year, major tech stocks led by Microsoft have remained under pressure. From their historical highs, $Microsoft(MSFT)$ has declined as much as 34%, $Meta Platforms, Inc.(META)$ over 31%, and $Amazon.com(AMZN)$ about 20%.
The primary concern behind this pullback is whether massive AI capital expenditures can generate sufficient returns. At the same time, as AI technology advances, some investors are beginning to question whether it could disrupt certain core businesses of major tech companies.
If weakness in big tech stocks persists, it could eventually affect future AI investment. Analysts estimate that tech companies’ AI-related capital expenditures will exceed $650 billion this year, growing over 40% year-on-year. However, by 2027, the growth rate is expected to slow significantly to around 12%.
In addition, from a pricing perspective, the rally in DDR5 memory chips has already stalled:
According to analysts’ estimates, the memory chip market is expected to reach supply-demand balance in the fourth quarter of 2027, with the shortage peaking in the third quarter of 2026:
Supply–Demand Analysis of Standard DRAM
From this perspective, the favorable conditions for memory chips may not be as sustainable as those for companies like Nvidia or TSMC. As supply and demand move toward balance, the upward momentum in memory prices is likely to ease. In other words, Micron’s current gross margin of around 80% may be difficult to sustain:
Although conditions are unlikely to remain this favorable indefinitely, Micron’s capacity for this year has already been fully sold out. Some major customers have also begun signing five-year long-term supply agreements with the company. In addition, SK Hynix has indicated that the memory shortage could persist through 2030. As a result, investors may not need to be overly concerned about Micron’s fundamentals, and price fluctuations are increasingly driven by valuation.
For example, escalating tensions in the Middle East have reduced market risk appetite, tightening overall liquidity. Technology stocks were already under pressure, and the negative news surrounding Google further added to the sentiment—making a decline in Micron’s share price unsurprising.
From a valuation perspective, Micron’s price-to-book ratio has fallen to around 5.5x, noticeably below TSMC’s 8.6x. If a deeper pullback occurs, the risk-reward profile would improve significantly:
If Google’s TurboQuant can be seen as a mild headwind for memory chips, then the declines in companies like TSMC and ASML appear somewhat excessive. After all, memory is only one component within GPUs, and a reduction in its usage does not materially affect TSMC, since memory chips are produced under the IDM model and are not outsourced to TSMC.
Semiconductor equipment stocks such as ASML may also have been caught in the selloff. Compared with memory, capacity expansion in logic chips remains far more aggressive. For example, TSMC’s capital expenditure is expected to reach around $54 billion this year, with analysts projecting it to rise further to about $58.6 billion by 2027.
Micron’s capital expenditure is also expected to remain elevated over the coming years:
Micron Capital Allocation Outlook
In addition, ASML’s current price-to-earnings ratio is only around 44x, still below the peak levels seen during the 2021 semiconductor bull market, suggesting there may be further upside ahead:
Therefore, the sharp pullback in semiconductor ETFs may not represent a risk, but rather a potential opportunity!
Comments