AI bull market second half: downstream commercialization may be more important than Capex

After three years of an industrial bull market, the U.S. AI industry is facing a critical structural turning point—a widening rift between the high prosperity of upstream computing hardware and the lag in downstream application commercialization. The direction in which this contradiction is resolved will determine whether the current AI cycle enters a long-term expansion or comes to an end with the peak in capital expenditure.

In a report dated July 2, the strategy team at Zheshang Securities pointed out: "The biggest structural contradiction in the current U.S. AI industry chain is that the high prosperity of upstream equipment and materials is highly dependent on the aggressive capital expenditure expansion of midstream cloud vendors, while the effective uptake of downstream application demand has not yet been fully realized." The stock prices of several representative AI application companies have fallen below or are approaching the critical technical support level of the "888-day" moving average, and market concerns about the midstream and downstream continue to intensify.

From a data perspective, the absence of a commercialized closed loop in the downstream application layer is creating tangible pressure. Salesforce's remaining performance obligations (RPO) growth rate has fallen from 21% in 2022 to 12% in Q3 2026 and is still declining; C3.AI's RPO has been continuously declining since the AI boom; the marginal return on investment (ROI) of North American cloud vendors has been declining since 2024, with Amazon breaking below the breakeven point.

Zheshang Securities judges that the subsequent tracking priority should be: downstream commercialization data > cloud vendor CapEx guidance > U.S. Treasury yield fluctuations.

Review of the First Year of AI: Profit Resilience, Not Liquidity Easing, Drove the Uptrend

To understand the current predicament, one must go back to the beginning of this AI bull market. In 2022, the Federal Reserve's aggressive rate hikes pushed the 10-year U.S. Treasury yield from 1.3% at the start of the year to a rapid climb to 4.24%, the highest level since 2008. The NASDAQ index fell 34.02% for the year, with a maximum drawdown of 36.36%. Microsoft fell 29.32% for the year, and Netflix saw a maximum drawdown of over 70%. At that time, the tech sector experienced a systemic "valuation compression."

However, the situation reversed starting in 2023. Even with U.S. Treasury yields remaining above 4% for an extended period, upstream AI hardware, driven by the rigid demand for large model training, achieved both volume and price increases, forming a numerator-side support against high interest rates. NVIDIA's data center revenue soared from $2.05B in Q1 2022 to $193.74B in Q4 2026; Google Cloud's 2025 revenue was $58.7 billion, up 35.8% year-over-year. This tech rally was not driven by liquidity easing but by the resilience of industrial earnings overcoming liquidity tightening.

In its report, Zheshang Securities uses the "888-day" moving average as a tool to verify the industrial cycle. This indicator corresponds to approximately 42.67 months (888 trading days), encompassing the evolution of a complete inventory cycle. History shows that when an industry index historically crosses this moving average, it often signals the substantial completion of underlying capacity clearing or a major reassessment of fundamental logic. Zhongji Innolight and Foxconn Industrial Internet both broke through the 888-day moving average in early 2023, confirming the start of a new prosperity cycle for upstream computing hardware.

Upstream Divergence: Cracks Appear in the Boom Driven by Midstream CapEx

The current high prosperity of upstream computing hardware is essentially a mirror image of the aggressive capital expenditure by midstream cloud vendors, rather than a transmission of genuine downstream demand. This logic is being tested.

From a financial data perspective, the revenue and gross margins of upstream companies in both China and the U.S. remain high. NVIDIA's gross margin has been above 70% since Q4 2024, and most leading Chinese companies maintain gross margins above 50%. However, divergence has already appeared at the stock price level—the 20-day moving averages of Lumentum Holdings, NVIDIA, and Broadcom have recently turned downward, while those of Applied Materials, Micron Technology, and SanDisk continue to rise, showing a clear "shrinking circle" in trading patterns.

Financial data from midstream cloud vendors reveals deeper underlying risks. According to Zheshang Securities, Microsoft's capital expenditure as a percentage of free cash flow surged to 637.54% in Q4 2025, and Amazon reached a peak of 3,587.91% in Q1 2026—meaning that for every $1 of free cash flow earned by cloud vendors, they face several times or even dozens of times more capital expenditures. Capital investment has become highly disconnected from operating profits, creating a strong reliance on external financing. Once the financing environment tightens, the pace of midstream expansion may face a pullback, which would directly transmit to upstream hardware demand.

Meanwhile, the marginal ROI of North American cloud vendors has been declining since 2024. Taking Amazon as an example, the marginal ROI has fallen from a historical high of 44.93% to below the breakeven point, indicating that new investments cannot guarantee profitability. There is a clear time mismatch between the depreciation of computing hardware and the release of demand.

Downstream Stall: Weakness in RPO and Warning from the 888-Day Moving Average Appear Simultaneously

The downstream application layer is the terminal export of the AI industry and the weakest link currently. Both domestically and internationally, the C-end remains dominated by free tools, and the B-end is mainly enterprise pilot projects. Replicable, scalable paid scenarios have not yet matured.

RPO (Remaining Performance Obligations) data is a forward-looking indicator of downstream willingness to pay. Salesforce's RPO growth rate has continuously declined from 21% in 2022 to 12% in Q3 2026, and it is still trending downward; Adobe's RPO percentage growth has declined in a volatile manner since 2022, with new contract signings slower than RPO consumption; C3.AI's RPO has been declining since the AI boom, with many projects stuck in the pilot stage. Zheshang Securities points out that this means enterprise customers' attitudes towards long-term AI subscription services are shifting from committed investment to a wait-and-see approach.

Technical signals further confirm these concerns. Microsoft's stock price has effectively broken below the support of the 888-day moving average; Netflix broke below this moving average on June 23; Oracle is also challenging the effective support of this moving average. Zheshang Securities believes this is a chart-based verification that the commercial closed loop for the AI application layer has not yet formed, and the transmission path of "terminal demand → computing power procurement → hardware profitability" is still in the磨合 (running-in) stage.

Risk Signal System: Three-Layer Observation of Macro, Industry, and Demand

Zheshang Securities has constructed a "macro-industry-demand" three-layer observation system to quantify the current risk position.

At the macro level, the 10-year U.S. Treasury yield is consistently operating within the 4.65% to 4.85% range, approaching the upper edge of the box range since 2023; U.S. core CPI has exceeded expectations for three consecutive months; the probability of a Fed rate cut has fallen from 60% to less than 30%, and the dot plot projections have shifted towards further rate hikes. This combination not only suppresses valuation expansion momentum but also prompts global long-term capital to reassess the allocation weight of the AI track.

At the industry level, the combined year-over-year growth rate of capital expenditure by North American cloud vendors has significantly declined, with a deceleration trend emerging; the increment in AI server shipments is slowing; the second derivative of the Philadelphia Semiconductor Index's year-over-year growth has shown an inflection point, indicating a structural shift from rapid recovery to moderate growth.

At the demand level, the predicted ARR growth rate of global AI companies continues to slow, and the proportion of C-end paid users still needs improvement. However, Zheshang Securities also points out that China's daily Token call volume has increased from 0.1 trillion in early 2024 to 180 trillion in June 2026, and the absolute ARR value of major AI large model companies is still growing, indicating that the medium-term trend of industry prosperity is still continuing. The current state is one where "there is a risk of a double whammy (valuation and earnings compression), but the market movement has not yet completely ended."

Two Scenarios: Application Implementation Determines the Final Outcome of the Market

Zheshang Securities presents two opposing forward-looking scenarios.

Optimistic scenario: If enterprise-level AI agents, vertical SaaS, and industry applications enter a phase of large-scale implementation, with user payments and inference demand continuing to grow, the AI industry may usher in a second growth phase driven by demand. The driving force of the industry will shift from the supply side to the demand side, and market attention will shift from capital expenditure growth rates to application revenue and user payment rates. Upstream hardware prosperity will be sustained, and investment opportunities will expand from computing hardware to models and application layers, forming a self-reinforcing positive cycle sufficient to offset the valuation suppression caused by high interest rates.

Pessimistic scenario: If AI application commercialization progresses slowly over the next three to six months, with enterprise willingness to pay and AI budget growth falling short of expectations, and against the backdrop of high U.S. Treasury yields, the market will focus more on the return on capital expenditure of cloud vendors. Once the growth rate of capital expenditure slows, demand expectations for the upstream industry chain will be revised downward accordingly, and the hardware sector will face simultaneous compression of earnings forecasts and valuation levels. Upstream hardware will shift from growth expectation pricing back to cyclical manufacturing pricing logic. Funds will concentrate on leading companies with strong order certainty, while theme-based targets will face significant valuation pullback risk.

Zheshang Securities emphasizes that the current crowding in the AI sector is unlikely to end due to a single unconfirmed risk but is more likely to manifest as increased volatility. The priority of core tracking indicators is as follows: AI paid ARR growth rate and vertical application paid penetration rate, cloud vendor CapEx guidance and global semiconductor sales year-over-year, U.S. Treasury yields and expectations of Fed monetary policy shifts.

Risk Warning and Disclaimer

        The market carries risks, and investment requires caution. This article does not constitute personal investment advice and does not consider the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at your own risk.