In the first week of August, major U.S.-listed AI supply-chain stocks broadly gained more than 10%. Nvidia logged five consecutive sessions of gains, while Marvell rose by nearly 20%. After a steep setback in July, AI-linked names recovered a sizable portion of lost ground.
PANews said the rebound came after a forceful late-July leverage washout rather than a clear change in the underlying story. Citing Morgan Stanley China chief economist Xing Ziqiang, the report said recent volatility in the AI theme did not reflect worsening fundamentals. Instead, it described the move as a temporary “halftime break” caused by a mix of crowded positioning, liquidity being drained by fundraising from large technology companies, and higher oil prices feeding expectations for rate hikes. The article added that this reading is becoming a common view among major Wall Street banks.
Even with AI-related U.S. equities still showing strong performance, the earlier frenzy is fading, the report said. In the first phase of the trade, capital clustered around compute chips and other “picks-and-shovels” assets. Morgan Stanley’s research team, as cited by PANews, sees the next phase moving along two tracks: one toward applications, where investors look for measurable efficiency gains and cash-flow conversion, and another toward the physical economy, where energy, raw materials, and other hard assets that AI cannot avoid are being revalued.
Three pressures hit the AI “picks-and-shovels” trade
Crowded positioning and a fragile leverage structure
The article described the first-half AI rally as a textbook crowded trade. Leveraged money and momentum capital poured into upstream segments including compute chips, semiconductors, and memory, pushing position concentration to elevated levels. That structure made the market more fragile and increased the risk of disorderly deleveraging.
In late July, that one-way positioning began to crack. PANews cited Goldman Sachs data showing that assets under management in leveraged semiconductor ETFs fell from a June peak of about $163 billion to $100 billion, a drop of nearly 40% and the largest decline since April 2025. Over the same period, semiconductor ETFs accounted for about 63% of outflows from all leveraged ETFs in the United States.
The report said the deleveraging also removed some of the excess tied to purely thematic speculation and helped cool an overheated AI trade.
Liquidity drain from massive capital spending
The report said the AI infrastructure race is consuming large amounts of market liquidity. Hyperscale cloud providers are planning to spend hundreds of billions of dollars on AI infrastructure, but cash flow alone is not enough to cover the gap. As a result, major technology firms have repeatedly turned to equity issuance and large corporate bond sales to raise funding.
Citing the Financial Times, the article said cumulative AI capital investment by four Silicon Valley giants had reached $1.1 trillion as of the second quarter of this year. It also cited JPMorgan research saying AI-related debt now accounts for more than 15% of the U.S. investment-grade bond market, making it the largest single debt segment. If monetization downstream falls short, the report said, excessive borrowing could become a threat to corporate credit ratings.
As secondary markets are steadily drained of liquidity, valuation pressure follows. In the article’s framing, the more aggressive the compute buildout, the stronger the siphoning effect on capital.
Oil and rates add valuation pressure
Macro conditions also weighed on expensive assets. The report said rising geopolitical tensions in the Middle East pushed up international oil prices, reviving worries about sticky inflation and lifting expectations that the Federal Reserve could tighten again. Higher risk-free rates raise the discount rate applied to future cash flows.
That is especially painful for AI-linked companies still in the investment phase and not yet delivering those cash flows. Under those three pressures, the once-favored compute suppliers lost momentum.
The investment case is shifting in two directions
PANews said Goldman Sachs and Morgan Stanley both see model training giving way to large-scale inference deployment. In that setting, simply adding more compute and bigger parameter counts no longer supports the same valuation premium. The market’s anchor is moving toward business execution and physical resource constraints.
Applications move from storytelling to ROI
The article said the next phase of AI investing will be a test of financial delivery. In the first phase, a company could often secure a valuation premium simply by attaching itself to the AI theme. In the next phase, parameter size will matter less than return on investment.
Investors are expected to pay closer attention to whether companies can actually use AI to reduce costs, improve efficiency, and convert those gains into revenue and cash-flow growth. As inference costs keep falling, application companies with closed-loop ecosystems, exclusive data assets, and sticky customer relationships may stand out. PANews pointed to businesses that embed AI into game development, ad placement, or digital operational workflows as examples of companies that can lower unit operating costs and turn AI into an internal productivity engine and a source of pricing power.
The article said that as capital moves closer to ROI, AI vendors will be pushed away from “competing on parameters” and toward “competing on implementation,” accelerating the move from lab experiments to real industry use. It added that at the application layer, only a small number of leaders such as Palantir (PLTR) have already shown financial growth, while other names still need to prove themselves through later market data.
HALO assets gain traction
The second track runs through the physical world. PANews said Goldman Sachs argued last month that “the endgame of AI is energy and raw materials,” a line that is now gaining wider acceptance in the market. HALO refers to Heavy Assets, Low Obsolescence — tangible assets with high barriers to entry and limited risk of rapid replacement, including copper mines, power grids, infrastructure equipment, and nuclear resources.
These are the physical assets AI cannot easily move, dismantle, or recreate. In its report The HALO Effect, Goldman Sachs said global markets are going through a repricing of scarcity. During the past decade, investors favored asset-light software models with high scalability. But AI has lowered the barrier to information processing and, according to the article, compressed the valuation and margin ceiling for software and IT services companies. At the same time, inflation and supply-chain regionalization have pushed up the replacement cost of physical assets.
The report argued that large models now iterate on a weekly basis, flattening the moat around algorithms and software services. Some asset-light SaaS firms that rely on simple code or intermediary services face disruption from AI agents. Open-source models can overtake proprietary algorithms. AI agents can rewrite software. Power grids cannot be copied at will, copper mines cannot be created on demand, and nuclear plants cannot be built overnight.
Four sectors inside the HALO theme
Using Wall Street’s classification, the article grouped the HALO theme into four areas: power and nuclear energy, grids and infrastructure, key raw materials, and engineering and manufacturing. PANews said these are all physical bottlenecks for AI buildout and include names that firms such as BlackRock and Goldman Sachs see with relatively high conviction.
Power and nuclear
In power and nuclear energy, the report highlighted Constellation Energy (CEG), Vistra Corp (VST), and NextEra Energy (NEE). With expansion of public grids constrained, these companies benefit from licensing advantages and behind-the-meter power delivery models that place data-center capacity next to generating units.
The article said that setup is turning nuclear plants into major energy suppliers for data centers. Long-term power purchase agreements, or PPAs, signed with technology giants and backed by floor pricing are also converting what used to be cyclical utility activity into cash-flow assets with stronger visibility.
Grids and infrastructure
On grids and infrastructure, PANews said the spread of high-power GPUs is pushing traditional air cooling to its physical limits, making a shift toward liquid cooling in data centers increasingly unavoidable. Vertiv (VRT), with its position in precision cooling and thermal management, was cited as a likely beneficiary of that upgrade cycle.
Eaton (ETN) and Quanta Services (PWR) were highlighted for their capabilities in power-distribution equipment, transformers, and high-voltage grid construction, all of which influence how fast grid expansion can actually happen. The long cycle of physical grid upgrades also creates meaningful barriers to entry.
Key raw materials
In raw materials, the article pointed to Freeport-McMoRan (FCX), saying the company controls high-quality large-scale copper resources and mining rights. Copper remains an irreplaceable conductor in power transmission, transformer windings, and internal data-center wiring.
The report added that long mine-development cycles and declining ore grades reduce the elasticity of new copper supply. A widening long-term supply-demand gap, in that view, should strengthen pricing power for copper resources.
Engineering and manufacturing
In engineering and manufacturing, PANews named Caterpillar (CAT) and Deere & Co (DE). The article said both companies have large physical factories, proprietary engineering know-how, and global supply-chain networks, creating hard barriers that code or algorithms cannot easily replace and allowing them to keep winning orders during infrastructure expansion.
Old-economy assets, new strategic role
The report said assets once treated as part of the old economy — Eaton’s transformers, Caterpillar’s heavy excavators, and Newmont’s copper mines — are now being assigned new strategic value in the next stage of AI investing. The repricing of HALO assets, it said, could provide the physical base for the next wave of AI infrastructure demand.
That does not remove the risks. HALO assets take time to build and require large amounts of capital. If commercialization at the AI application layer comes through more slowly than expected, early investment in energy and compute infrastructure could still result in excess capacity and stranded assets.
PANews concluded that this “halftime break” is a necessary step in a more rational market split. Future excess returns, the article said, may stretch in two directions at once: toward real commercial use cases and toward hard assets with durable physical moats. Companies that combine monetization ability with those physical defenses are the ones most likely to stay ahead after the reset.