BlockBeats News, July 3rd, the U.S. stock market closed early ahead of the Independence Day holiday, but AI trading did not calm down. Over the past week, investors have been chasing the imaginative space of Meta commercializing AI computing power while rapidly selling off previously soaring storage and semiconductor stocks. U.S. storage stocks such as Micron and SanDisk experienced a sharp single-day decline, and the South Korean KOSPI also plummeted under the leadership of SK Hynix and Samsung Electronics, indicating that the market is reevaluating the pace, returns, and supply bottlenecks of AI infrastructure expansion.

The trigger of this round of volatility was Meta's alleged consideration of launching cloud services and selling excess AI computing power to external customers. The news temporarily drove Meta's stock price up by 8.8% as investors saw its massive capital expenditure potentially converting into revenue more quickly. However, the same narrative also raised concerns in the opposite direction: if large tech companies already have computing power available for external sale, does the AI infrastructure suffer from overbuilding? This question quickly spread to storage, chip, and AI cloud service providers, becoming the core background of this week's semiconductor sector correction.

Morgan Stanley provided a different answer in its global NAND industry report released on July 2nd. The report believes that AI will continue to create a NAND shortage until 2027, with AI-related NAND demand expected to increase from 205EB in 2025 to 609EB in 2027, with its share of total demand rising from 18% to 41%. In its model, there is still about a 9% supply-demand gap in global NAND in 2027.

This implies that the current market is not trading a simple "AI demand disappearance" but repricing based on growth, valuation, and supply expansion pace. Morgan Stanley remains bullish on storage suppliers such as Micron, SK Hynix, and SanDisk, but also emphasizes that industry differentiation is intensifying: demand for servers, enterprise SSDs, and HBM remains strong, while the consumer end like PCs and smartphones has shown signs of rising inventory and order cuts by customers.

This is also the most contradictory aspect of this week's market. Micron's stock price dropped by 10.6% on July 1st and continued to weaken the next day; Korean chip stocks also faced concentrated selling pressure, with SK Hynix and Samsung Electronics leading the decline in the KOSPI. However, at the fundamental level, storage prices are still rising, and the demand for AI servers for DRAM, HBM, and enterprise SSDs has not significantly reversed. In other words, what investors are selling is the "one-way uptrend" trading crowding, not a complete denial of the storage cycle.

The Morgan Stanley report also suggests that risks will become more complex after 2028. If AI SSD demand continues to grow rapidly, and companies like Yangtze Memory Technologies maintain restrained capacity expansion, NAND supply may still be relatively tight; but if Chinese manufacturers or other players accelerate greenfield expansion, the industry may face oversupply pressure again. Therefore, the core variable of storage stocks has shifted from "is there demand" to "can demand growth outpace new supply."

Another Bernstein report further shifts the AI compute narrative from GPU to CPU. The firm believes that Agentic AI will bring about a so-called "CPU Renaissance": intelligent agent applications require task orchestration, tool invocation, context management, and multi-agent collaboration, all of which rely more on general-purpose computing power than traditional large-scale model training. Bernstein projects that the global server CPU market size will increase from approximately $39 billion in 2025 to $223 billion in 2030.

In the Chinese market, Bernstein sees Hygon as a key beneficiary. The report states that the Chinese x86 server CPU market size is expected to grow from $7 billion in 2025 to $27 billion in 2030, with Hygon's revenue share in the Chinese x86 server CPU market likely to increase from 19% to 36%. The rationale behind this is not just the IC policy, but post-2027, cloud players may accelerate the adoption of domestic alternatives due to unstable overseas CPU supply, rising procurement costs, and the expansion of domestic AI infrastructure.

These two reports collectively point to a shift: AI trades are moving from "buy all compute-related stocks" to a more granular bottleneck identification phase. Previously, the market revolved around GPU, HBM, and cloud capex pricing; now, enterprise SSDs, NAND, server CPUs, domestic alternatives, and long-term supply agreements are all starting to become part of valuation discussions.

For investors, this week's sharp pullback does not signal the end of the AI hardware theme. It is more of a stress test: as the market begins to question whether there is an excess of compute power, segments with true pricing power, supply constraints, and long-term agreement protections are more likely to differentiate from crowded trades. In the short term, AI hardware stocks may continue to endure profit-taking; in the medium term, storage and CPUs are becoming harder-to-bypass infrastructure variables in the next wave of compute cycles.