Zhongtai Securities: As the global AI capital expenditure paradox sparks heated debate, four key variables determine the adjusted interpretation path

According to the Zhitong Finance APP, Zhongtai Securities released a research report stating that after experiencing a deep adjustment, the global technology stocks' core focus on AI investment is shifting from "scale worship" to "value verification." The post-adjustment trajectory of technology stocks needs to focus on four key variables: First, whether the impact of the Federal Reserve's interest rate hikes has been fully priced in—after the dissenting votes at the July FOMC and the market's shift to a data-driven model following Waller's abandonment of forward guidance, the current pricing of interest rate hikes is likely excessive; Second, whether the capability gap between frontier and open-source models can be reopened—regulatory procedures have added extra time costs to closed-source models, objectively providing a catch-up window for the open-source camp; Third, whether the process of deleveraging is nearing its end—hedge fund holdings in technology have dropped to a near three-year low, and the concentrated liquidation of leveraged ETFs in South Korea has also been fully released; Fourth, whether the IPO financing window for model companies can be successfully opened—if the listings of OpenAI and Anthropic do not progress as expected, the primary market valuations will face systemic reassessment.

The main points of Zhongtai Securities are as follows:

Deleveraging and earnings season: Systematic changes in the "aesthetics" of AI investment: After experiencing this round of deep adjustments, the market's core aesthetics regarding AI investment is undergoing a systematic shift from "scale worship" to "value verification." In the second quarter earnings season of 2026, Alphabet and Microsoft recorded drastically different stock price reactions against the backdrop of increased capital expenditures, clearly indicating that funds are no longer simply paying for scale but are demanding that every capital expenditure must correspond to verifiable revenue growth and clear financial explanations. Meanwhile, hedge funds continue to reduce their holdings in U.S. technology stocks, with the technology sector's exposure shrinking to a near three-year low. Coupled with the forced liquidation of single-stock leveraged ETFs in South Korea, the impact on the funding structure has been concentrated and released, and the market is seeking a new equilibrium point from the deleveraging.

"AI Capital Expenditure Paradox": Why the market is beginning to question the authenticity of AI demand: Current market doubts about AI demand are concentrated in the "capital expenditure paradox"—the same investment in computing power is both a signal of industry prosperity and a source of erosion of free cash flow. The key to unraveling this paradox lies in distinguishing between two types of buyers: the cash flow-supported demand represented by Google and the high-leverage financing-driven demand represented by Oracle. The latter has remaining performance obligations of $638 billion by the end of fiscal year 2026, but its free cash flow has turned negative for fiscal year 2026. Whether it can convert orders into revenue as expected is a core variable in judging the sustainability of this round of capital expenditure. Historically, Oracle's high-leverage model bears similarities to the financing expansion of telecom operators in 2000, but the presence of mature tech giants like Google (GOOGL.US), Microsoft (MSFT.US), and Amazon (AMZN.US) in the current buyer structure, along with the national security attributes of AI and the reallocation capabilities of computing power, constitute a systematic rebuttal to the "simple analogy of bubble bursts." The Essential Differences Between the 2000 Internet Bubble and the Current AI Cycle: The underlying factor for the severity of the 2000 tech bubble burst was the break in the financing loop—telecom operators built network capacity far exceeding actual demand relying on IPOs, supplier financing, and high-yield bonds. Once the financing window closed, orders based on "expected demand" instantly lost support. Comparing today with 2000, there are indeed similarities—long-term contracts and prepayments functionally resemble the supplier financing of that time, as orders do not equate to cash. However, the differences constitute a systematic rebuttal: today's buyers are primarily giants like Google and Microsoft, supported by mature business cash flows, rather than loss-making startups; computing power has clear capacity limits and reconfiguration capabilities, contrasting sharply with the large amounts of idle fiber optics in 2000; and the national security attributes of AI further imply the existence of policy backing. High-leverage examples like Oracle are worth noting, but simply applying Cisco's outcome to the entire AI industry overlooks the essential differences in buyer structure and underestimates the real industrial support behind this round of demand.

The Arms Race of Betting on Inference: The moat and sustainability of the AI model layer: Revenue growth in the AI model layer shows significant differentiation—leaders have already forecasted their first profitable quarter, while followers remain mired in large losses. More crucial than the total amount is the shift in revenue structure towards enterprise-level entry, as AI transforms from a technological novelty into a business necessity. From a moat perspective, the leading window for advantages in model architecture and algorithms is compressing from "years" to "months," with true stickiness reflected in the depth of ecological integration—once AI Agents are deeply embedded in enterprise codebases and business processes, migration costs become decoupled from the extent of model capability leadership. This competitive landscape determines that a single winner is unlikely to emerge at the model layer, and as long as a few leading companies continue to chase each other, the demand for computing power driven by the arms race will not systematically stagnate due to any single company's commercialization falling short of expectations.

Four Key Variables Determine the Adjusted Derivation Path: Focus needs to be on four key variables: First, whether the impact of the Federal Reserve's interest rate hikes has been fully priced in—after the dissenting votes at the July FOMC and the market's shift to a data-driven model following Walsh's abandonment of forward guidance, current interest rate pricing is likely excessive; Second, whether the capability gap between frontier and open-source models can be re-established—regulatory procedures have added extra time costs for closed-source models, objectively providing a catch-up window for the open-source camp; Third, whether the process of deleveraging in funds is nearing its end—hedge fund tech holdings have fallen to a near three-year low, and the concentrated liquidation wave of leveraged ETFs in South Korea has also been fully released; Fourth, whether the IPO financing window for model companies can be smoothly opened—if the listings of OpenAI and Anthropic do not progress as expected, primary market valuations will face systematic reassessment.

Risk Warning:

AI commercialization and the penetration of knowledge work Agents may fall short of expectations; there may be significant discrepancies between the revenue disclosure standards of model companies and actual financial results; high-leverage computing cloud and data center projects may face refinancing obstacles; there may be overcapacity in devices such as GPUs and storage; depreciation and asset impairment may exceed expectations; rising energy prices, electricity access, and regulatory constraints; The release of cutting-edge models has been delayed beyond expectations due to regulatory scrutiny; the Federal Reserve's interest rate hike expectations have fluctuated; related technology asset prices may face significant volatility