Written by: Rita

Recently, the AI infrastructure sector has experienced a notable pullback, with Nvidia retracting about 12% from its peak, and stocks like Broadcom and Marvell also declining. The market's doubts about the sustainability of capital expenditure in AI continue to intensify. Morgan Stanley released a global thematic strategy report on July 27, proposing a contrary judgment that the gap between supply and demand for computing power will further widen, presenting a reallocation opportunity during this adjustment, and outlining five major investment directions.

Bottleneck in AI Infrastructure

Conditions for power access, construction capabilities, and land resources constitute the core barriers in the industry. The report defines companies with such resource endowments as "power shell providers," with representative targets including Talen Energy, Vistra, Bloom Energy, and Hut 8, Cipher Mining, TeraWulf, Riot Platforms, among others. The scarce land indicators and grid access permissions are key constraints upstream of data center construction and represent the highest difficulty in replicating the AI industry chain.

Computing Power Manufacturing Ecosystem

The track covers semiconductor manufacturing, equipment, materials, and packaging links. Nvidia remains the preferred choice for institutional semiconductor portfolios, while Broadcom and Micron are also included in the recommendation list. Morgan Stanley points out that Nvidia's valuation in the semiconductor sector is significantly underestimated.

China's AI Solutions

Institutions judge that domestic AI companies are developing faster than the market generally expects. ByteDance is likely to continue increasing its investment, with capital expenditure expected to rise to $80 billion, while Alibaba and Tencent are also ramping up AI business investments. The self-sufficiency rate of domestic AI chips is expected to increase from 42% in 2025 to 70% in 2030, with the scale of domestic AI models landing in overseas markets steadily rising.

Energy Security

Energy storage, grid equipment, and natural gas power generation are key focus areas. Due to the shortage of transformer supplies in the United States, South Korean power equipment manufacturers are facing development opportunities, with HD Hyundai Electric and LS Electric as core targets. Institutions predict that South Korean companies' share in the U.S. transformer import market will continue to rise.

Large-scale Manufacturers

Meta, Google, Microsoft, and Amazon rank high on the list. The market generally underestimates the return on investment ratio of the capital expenditure of leading cloud providers in AI, where the existing business moats combined with the commercialization potential of AI lead to fundamentals exceeding expectations. Meta possesses 3.5 billion daily active users, and there is still room for optimization in AI-enabled advertising and content recommendation systems.

Supply and Demand Conflicts and Core Market Discrepancies

The underlying logic of the five major themes is based on the judgment of an imbalance in supply and demand for computing power. Google executives have publicly stated that the company's computing power scale is planned to double every six months and expand 1000 times within 4 to 5 years. In contrast, Nvidia's annual compound growth rate in computing power shipments is about 140%, creating a significant gap between the demand expansion and supply growth, which is difficult to repair in the short term.

In response to the two major concerns most concerning the market, the report provides clear reasoning.

On one hand, cases of companies controlling AI spending are increasing; Uber exhausted its AI budget for 2026 already in April, and Meta has also set monthly usage limits. However, institutions define this phenomenon as short-term adjustment friction. Data shows that the average monthly token consumption for enterprise users is about $11, and a single AI transformation project can save costs of $55, with a return on investment ratio exceeding 20 times. Budget constraints are a phase adjustment and do not indicate the arrival of a demand turning point.

On the other hand, after the release of Kimi K3, the market is once again discussing the "DeepSeek moment," worrying that the narrowing of the technology gap between China and the U.S. suppresses computing power demand. Morgan Stanley holds an opposing view: efficiency improvements in models will reduce per-unit token costs, driving overall usage scale upward, ultimately generating greater demand for computing power, aligning with the Jevons Paradox in economics.

Supply-side Bottlenecks and Layout Ideas

The growth certainty on the demand side is relatively strong, while constraints on the supply side are becoming increasingly prominent. Institutions estimate that from 2026 to 2028, the gap in power supply and demand for data centers in the U.S. will reach 38GW, while the scale of projects that have started construction and can be smoothly interconnected is only 15GW each. Even with additional transitional power supply solutions such as gas turbines and fuel cells, there remains a persistent gap of 1GW to 11GW. The shortage of skilled labor is also difficult to alleviate quickly, with insufficient supply of skilled workers. At the policy level, 14 states across the U.S. are brewing data center control bills, with New York State implementing a one-year suspension policy.

The report also shares a set of backtesting conclusions: increasing positions to 150% when the AI thematic index pulls back 5%, and increasing positions to 200% when it pulls back 15%, with a holding period set at 3 months, showing long-term returns outperforming static holdings. The volatility in the AI infrastructure sector is mostly driven by emotions and does not represent a reversal in fundamental trends.

In summary, the core logic of this research report can be summarized as follows: The demand for AI computing power is increasing exponentially, while supply is limited to linear expansion due to various physical conditions, continuously widening the supply-demand gap. Short-term stock price fluctuations are more about emotional noise, while the imbalance of supply and demand for computing power is the main theme in the medium to long term.

Disclaimer

This article is a整理与解读 (整理与解读 means organization and interpretation) of a third-party broker research report (Morgan Stanley, July 27, 2026) by Chao Xiang Research. The ratings, target prices, earnings forecasts, and related judgments quoted in the text are the opinions of the analysts of that brokerage, representing only the position of their respective institutions, not the views of Chao Xiang Research, and do not constitute any investment advice.

The market has risks, and decisions must be independent. This article should not be used as a basis for buying or selling any securities.