The AI infrastructure sector is undergoing a market-level stress test. Over the past two weeks, the performance of hard tech sectors in China and the United States has shown a clear divergence. Market disagreements have intensified around issues such as whether high capital expenditure can be sustained and the ongoing correction in the Token expenditure index. The strategy team at Changjiang Securities believes this divergence is more akin to a stress test for AI infrastructure rather than the end of the current market trend.
The slowdown in capital expenditure growth and the correction in the Token expenditure index are the two main themes of the current market divergence. In a report dated July 18th, Changjiang Securities pointed out that a slowdown in CapEx does not necessarily lead to the end of the hardware market trend. Meanwhile, the correction in the Token expenditure index reflects a market shift from "technology feasibility trading" to "economic feasibility trading"—investors are beginning to demand that AI demonstrate its ability to generate sustainable revenue, rather than merely consuming computing power.
This shift in logic has a direct impact on asset allocation. The team judges that the main theme of the AI market trend is shifting from "upstream hardware beta" to "genuinely scarce segments" and "application revenue realization," i.e., moving from CapEx trading to ROI trading. Mapping this to the A-share market, scarce hardware directions such as HBM, advanced packaging, and high-speed optical interconnects are more likely to regain a premium after adjustments, whereas ordinary servers, ordinary storage, and high-valuation crowded trades face greater pressure.
Sino-US Hard Tech Divergence: Domestic Computing Power Chain Outperforms Against the Trend
From June 29th to July 9th, 2026, against a backdrop of historically high crowding in the technology sector, the hard tech sectors in China and the US experienced varying degrees of correction, with significant internal divergence.
Domestically, overseas computing power directions such as optical modules saw substantial corrections. However, with Changxin Technology's IPO approaching and supported by expectations of production expansion certainty, domestic computing power chain directions like semiconductors performed relatively well. Internationally, storage giants like Samsung and SK Hynix also saw significant corrections, while companies like Nvidia and Apple performed relatively better.
Changjiang Securities believes this divergence pattern stems from a significant increase in market discussion regarding the sustainability of high CapEx spending. With overall crowding in the tech sector at historically high levels, this has triggered a systemic stress test on AI infrastructure.
First Stress Test: Slowing Capital Expenditure Growth
The slowdown in CapEx growth is one of the core triggers for the current market divergence. Judging from the capital expenditure data and forward guidance of major North American cloud providers, overall capital expenditure growth for 2027 may see a marginal decline.
Citing a historical review of the 2010s mobile internet market trend, Changjiang Securities points out that the peak inflection point in capital expenditure growth typically leads the market peak by about one quarter. In the early stages of a market trend, the order in which capital expenditure growth peaks across TMT sub-sectors is: communications > electronics > computer = media. Towards the end of a trend, capital expenditure growth in the software end declines before that in the hardware end, with indices beginning to fall about one quarter after growth in all sectors has peaked and declined.
However, the team emphasizes that whether a CapEx slowdown evolves into an inflection point for the hardware market trend depends on three conditions: whether revenue growth slows simultaneously, whether supply is fully released, and whether the financing chain breaks. The AI chain has now entered the "first stage of CapEx anxiety." In the short term, expectations for broad-based gains in the mid-to-upstream sectors should be moderated. However, as long as Token demand, cloud AI revenue, and inference scenarios have not been disproven, the AI cycle may still continue.
Historical experience from the new energy vehicle industry chain also provides a reference. A review shows that during an industry upcycle, even if the penetration rate growth slows, high gross margins in the upstream can often be maintained for a period. The elasticity ranking of sub-sectors typically follows the pattern of upstream > midstream > downstream. This pattern only reverses after a downward trend in upstream gross margins is established.
Applying this to the current AI wave, Changjiang Securities believes the key variable lies in whether 2027 capital expenditure can be revised upward. This depends on two factors: whether AI model iteration can break through linear expectations and open up new paradigms for computing power demand; and whether the application end can accelerate ROI realization, driving capital expenditure to shift from "stockpiling GPU investments" to "genuine consumption-driven expansion."
Second Stress Test: Correction in the Token Expenditure Index
Since June 2026, the LLM Token Index (price per million tokens) has continued to correct. However, during the same period, the absolute value of global Token usage has maintained slight growth. This divergence signal is the second main theme of the current market disagreement.
Changjiang Securities interprets this phenomenon as a structural shift in AI trading logic: the market is not adjusting due to concerns about AI computing power demand, but rather starting to demand that AI prove it can not only consume Tokens but also create sustainable revenue at a lower cost. In other words, AI is transitioning from technology feasibility trading to economic feasibility trading.
From a model pricing perspective, the prices of closed-source large models (taking the Claude Opus series as an example) remain relatively high, while the prices of open-source large models (taking the DeepSeek V series as an example) have seen significant reductions and are notably lower than closed-source models, indicating a widening divergence. Meanwhile, the revenue growth of major North American cloud providers (Microsoft, Google, Oracle, Amazon, Meta) is gradually recovering, and the Annual Recurring Revenue (ARR) of leading overseas large model providers is also showing an accelerating upward trend, indicating that AI commercialization is still progressing.
The team points out that the market is shifting from the first stage of "willingness to pay a high price as long as AI can be used" to the second stage of "only willing to pay continuously for AI that can create ROI." This shift will reduce the indiscriminate premium for cutting-edge models and compress overly optimistic hardware valuations in some areas, but it will also drive the expansion of low-cost models, model routing, inference optimization, and enterprise-level AI applications.
Theme Shift: Scarce Hardware and Application Revenue Become the New Anchors
Considering both stress tests, Changjiang Securities believes the current divergence marks a shift in the main theme of AI investment, with market focus gradually turning towards application-end revenue certainty and high-end hardware scarcity.
Mapping this to the A-share market, the team clearly distinguishes between the benefiting and pressured directions: ordinary servers, ordinary storage, and high-valuation crowded trades are most vulnerable to impact; whereas genuinely scarce directions like HBM, advanced packaging, and high-speed optical interconnects are more likely to regain a premium after adjustments.
Taking a longer-term view, Changjiang Securities believes one cannot simply equate a CapEx slowdown with the end of the AI cycle. The current AI application trend has accumulated a solid hardware foundation over the past 10 to 20 years—spanning multiple industry cycles such as communications infrastructure (2010-2013), semiconductor equipment and materials (2020), cloud computing (2020), and new energy vehicles (2021). Combined with the continuous iteration of large AI models since 2024, the next wave of AI application market activity may already be "poised for takeoff."
The team also highlights three main risks: the limited guidance historical experience provides for the future; potential intensified market volatility if the capital expenditure guidance from North American tech giants during the earnings season falls short of expectations; and potential synchronized volatility in the tech sector if the ARR data from major AI model providers like Anthropic falls short of expectations.
Comments