AI Investment 'Halftime': Computing Stocks Lose Appeal, Focus on Two Key Areas in the Second Half

Author: Jae, PANews

In the first week of August, mainstream stocks in the AI industry chain on the US stock market generally achieved a remarkable increase of over 10%. Nvidia's stock price rose for five consecutive days, while Marvell gained nearly 20%. After experiencing a "Waterloo" in July, AI concept stocks finally regained their footing in the new month, recovering a significant amount of lost ground.

However, this rapid rise is actually inseparable from the intense leverage cleansing that took place in late July. At that time, upstream targets in the semiconductor and computing sectors faced significant valuation pullbacks, and the market was filled with panic over a "bubble burst." Morgan Stanley's chief economist for China, Xie Zhiqiang, pointed out in his latest analysis that the recent fluctuations in the AI sector are not due to a deterioration in fundamentals, but rather a phase of "halftime adjustment" caused by crowded trading, large firms withdrawing financing, and rising oil prices amplifying interest rate hike expectations. This view is becoming a consensus among major investment banks on Wall Street.

Currently, although AI industry chain stocks in the US still show strong performance, the fervor in the AI sector is undoubtedly gradually cooling down, and the hottest engine in the global capital market is quietly shifting gears. In the first half of AI investment, capital flocked to computing chips, betting on the certainty of "selling shovels"; Morgan Stanley's research team believes that the investment focus in the second half may shift along two paths: one sinking towards the application end, chasing real cost reduction and cash flow realization; the other extending into the physical world, reassessing hard assets such as energy and raw materials that AI cannot bypass.

"Three Mountains" Crushing AI's "Selling Shovels"

The combination of three forces has collectively pressed the "pause button" on the AI sector.

Crowded Trading: A Fragile Balance Built on High Leverage

The AI market in the first half of the year is a typical example of "crowded trading."

A large amount of leveraged funds and momentum investors rushed into upstream sectors such as computing chips, semiconductors, and storage, causing the concentration of chips to rapidly rise to historical highs. The crowded chip structure amplified the market's fragility, leading to a stampede-like deleveraging sell-off.

In late July, overly consistent trading expectations faced a concentrated loosening of chips. According to Goldman Sachs, the assets under management (AUM) of leveraged semiconductor ETFs fell from a peak of approximately $163 billion in June to $100 billion, a decline of nearly 40%, marking the largest drop since April 2025. During the same period, semiconductor ETFs accounted for about 63% of the outflows from all leveraged ETFs in the US.

However, deleveraging has also effectively cleared the bubble of pure concept speculation, cooling the overheated sentiment in the AI sector.

Capital Withdrawal: Liquidity Backlash from Hundreds of Billions in Capital Expenditures

The other side of the AI arms race is the continuous consumption of liquidity in the secondary market.

Global hyperscale cloud service providers plan to invest hundreds of billions of dollars in AI infrastructure to seize the computing power high ground, but their cash flows cannot fully cover such a massive capital expenditure gap. As a result, tech giants frequently resort to large-scale financing through stock issuances and issuing large corporate bonds. According to the Financial Times, the cumulative capital investment in AI by the four major Silicon Valley giants has reached $1.1 trillion as of the second quarter of this year. Morgan Stanley's research team also pointed out that AI-related debt now accounts for over 15% of the US investment-grade bond market, becoming the largest single debt sector. If downstream monetization does not meet expectations, excessive debt will pose a potential threat to corporate credit ratings.

As the secondary market continues to be "drained," the supply of funds tightens, and valuations naturally come under pressure. Simply put, the more aggressively computing power expands, the stronger the siphoning effect on liquidity.

Interest Rate Clouds: Valuation Squeeze from Inflation Rebound

Macroeconomic variables have become the last straw that broke the camel's back for high valuations.

The escalation of geopolitical conflicts in the Middle East has pushed up international oil prices, and concerns about sticky inflation rebounding have resurfaced, forcing expectations for Federal Reserve interest rate hikes to rise. The upward movement of risk-free interest rates has raised the discount rate for future cash flows. For AI targets that are still in the investment phase and have not yet realized cash flows, the increase in the discount rate further suppresses their valuations.

Under the triple pressure, the business of "selling shovels" has suddenly become difficult.

AI Investment Logic Shift: One Hand on ROI, One Hand on Hard Assets

Goldman Sachs and Morgan Stanley point out that foundational model training is transitioning towards large-scale inference deployment, and the story of simply stacking computing power and competing on parameters is beginning to lose its marginal effectiveness. The valuation anchor in the capital market is also shifting towards the ability to realize business models and the resource bottlenecks of the physical world.

AI Application Side: From Storytelling to Calculating ROI

The second half of AI investment is an elimination race based on financial realization capabilities.

In the first half, any target associated with the AI concept could enjoy valuation premiums; in the second half, parameter scale is no longer the main indicator, and the return on investment (ROI) becomes the basis for obtaining high valuations. The market will pay more attention to whether companies can use AI to achieve cost reduction and efficiency improvement, translating into revenue and cash flow growth.

As inference costs continue to decline, application-oriented companies with strong closed-loop ecosystems, exclusive data assets, and high customer stickiness will stand out. For example, companies that embed AI in game development, advertising placement, or digital business processes can significantly reduce unit operating costs, turning AI technology into an endogenous efficiency engine and product pricing power.

Capital's shift towards ROI will force AI vendors to transition from "competing on parameters" to "competing on implementation," accelerating AI's move from the laboratory to real industries. Currently, in the AI application sector, apart from a few leading players like Palantir (PLTR) that have achieved financial growth, other targets still need subsequent market data to speak for themselves.

HALO Assets: The End of AI is the Physical World

"The end of AI is energy and raw materials," a judgment from Goldman Sachs last month is becoming market consensus.

HALO (Heavy Assets, Low Obsolescence) assets refer to tangible assets with high barriers to entry that are difficult to be quickly replaced by technology, such as copper mines, power grids, infrastructure equipment, and nuclear energy resources. These physical hard assets that AI "cannot move, dismantle, or create" may attract global capital.

Goldman Sachs pointed out in its report "The HALO Effect" that the global market is undergoing a "repricing of scarcity." Over the past decade, the market has favored "light asset, high expansion" software models, but AI has lowered the threshold for information processing, significantly compressing the valuation and profit margin ceilings of software and IT service companies. In contrast, the reset costs of physical assets have risen significantly due to inflation and the re-regionalization of supply chains.

In simple terms, large models are iterating on a weekly basis, the barriers for algorithms and software services have been greatly flattened, and light asset SaaS companies that rely on simple code or intermediary services face the disruptive risk of being replaced by AI Agents. Algorithms can be surpassed by open-source models, software can be rewritten by AI Agents, but power grids cannot be easily replicated, copper mines cannot be created out of thin air, and nuclear power plants cannot be built overnight.

According to Wall Street's classification standards, HALO themes cover four main sectors: power and nuclear energy, power grids and infrastructure, critical raw materials, and engineering manufacturing, each of which is a physical checkpoint that AI computing power expansion cannot bypass, and there are also potential targets with high consensus among institutions like BlackRock and Goldman Sachs.

In the power and nuclear energy sector, independent nuclear power giants represented by Constellation Energy (CEG), Vistra Corp (VST), and NextEra Energy (NEE) are becoming the focus of capital attention. Against the backdrop of limited public grid expansion, they leverage licensing advantages and the "behind-the-meter" direct supply model, making nuclear power plants the main energy suppliers for data centers, attracting tech giants to sign long-term power purchase agreements (PPAs) with guaranteed minimum prices, transforming originally cyclical public utilities into cash flow assets with high certainty.

In the power grid and infrastructure sector, the widespread use of high-power GPUs has pushed traditional air cooling to its physical limits, making the transition of data centers to liquid cooling technology an inevitable trend. Vertiv (VRT), relying on its leading position in precision cooling and thermal management technology, will benefit from the upgrade of data center cooling. Eaton (ETN) and Quanta Services (PWR) control the construction capabilities of distribution equipment, transformers, and high-voltage power grids, determining the actual speed of grid expansion. Additionally, the long-cycle projects of physical grid upgrades also create high competitive barriers for them.

In the critical raw materials sector, Freeport-McMoRan (FCX) holds high-quality copper mine resources and mining rights. Whether for power transmission, transformer windings, or internal wiring of data centers, copper is an irreplaceable physical conductive medium. The long development cycle of mines and declining ore quality significantly reduce the supply elasticity of new copper mines, and the long-term expansion of the supply-demand gap will continue to push up the pricing power of copper mine resources.

In the engineering manufacturing sector, Caterpillar (CAT) and Deere & Co (DE) possess vast physical factories, proprietary engineering technologies, and global supply chain networks, creating physical barriers that are difficult to be replaced by code or algorithms, continuously securing orders amid the infrastructure boom.

In the second half of AI investment, Eaton's transformers, Caterpillar's giant excavators, and Newmont's copper mines, once regarded as "old economy" assets, are suddenly endowed with new strategic value. The repricing of HALO assets will lay a solid physical foundation for the next stage of even larger AI infrastructure demands.

However, the construction cycle of HALO assets is long, and the capital investment is substantial. If the commercialization of downstream AI applications does not meet expectations, the pre-invested energy and computing infrastructure may also trigger risks of overcapacity and asset obsolescence.

"Halftime adjustment" is a necessary stage for the capital market to move towards rational differentiation, and future excess returns will stretch towards real business scenarios while digging into solid hard assets. Only players with both commercialization capabilities and physical moats can maintain their lead in the long race after the "halftime adjustment."