From late June to the end of July, the global AI sector underwent a sharp correction. The adjustment spread from South Korea to North America and then to A-shares—this was not a break in the industry logic, but a concentrated purge of high-valuation assets under the pressure of leverage contraction.
A research report released by a leading Chinese securities firm on August 9 pointed out that this pullback was triggered by three forces together: the loosening of crowded trades, the failure of AI application growth to consistently exceed expectations, and the shift in the South Korean memory chain's trading logic from "earnings upgrades" to "peak earnings and expansion pressure." The combination of these three factors amplified the volatility of the global AI hardware chain.
Since the end of July, the earnings reports of North American cloud giants have become a key anchor for stabilization. Three new narratives are reshaping AI pricing. First, the market is no longer simply punishing cloud giants for high CapEx and deteriorating free cash flow, but is instead rewarding the delivery of revenue, orders, and profits. Second, upward revisions to cloud giants' CapEx are once again becoming a positive for the industry chain, alleviating market concerns about a slowdown in orders for chips, communication connections, and storage. Third, although the CapEx growth rate of traditional cloud giants may decline in 2027-2028, the inference demand from model makers and new cloud providers, as well as the new training demand for Recursive Self-Improvement (RSI), are expected to become an important source of接力 for computing power investment in the next phase.
The securities firm believes that the iterative demand from agents to the breakthrough of RSI could become the main logic for the next phase of AI investment. The core of the RSI logic is that improvements in AI capabilities may simultaneously bring about two main lines: "expansion of training demand" and "opening of application demand."
At the same time, the firm pointed out that stabilization does not mean a return to a one-sided upward trend; the subsequent period is more likely to enter a phase of high volatility and high rotation. The AI sector may form a first阶段性 top, but this is not the end of the industry cycle. The subsequent main line is likely to continue rotating along the path of "hardware → cloud giants → model makers → applications."
Adjustments Caused by Leverage Clearing
The starting point of the adjustment was June 22-23. At that time, the KOSPI was around 9,114 points, and the Philadelphia Semiconductor Index was near 14,634 points, with market valuations lacking the safety margin for further increases. South Korean regulators subsequently stated that the decision to allow the listing of single-stock leveraged ETFs "may have been too hasty" and would review investor protection and product risks. This statement directly triggered margin calls and reductions in leveraged products.
On June 25, Apple raised prices on some Mac and iPad models, citing rising storage and memory costs. The next day, the Philadelphia Semiconductor Index fell 5.29%, and the Nikkei 225 fell 4.15%. The industry narrative also shifted: memory price increases changed from "a positive for upstream profitability" to "downstream cost pressure." This was a typical collective risk contraction after high valuations and high leverage encountered a consensus sell signal, not a collapse of industry fundamentals.
Three New Narratives Reshaping AI Pricing After Cloud Giants' Earnings
In late July, the four major North American cloud giants集中 disclosed their earnings, and the AI sector began to stabilize. Behind this are three new narratives reshaping pricing logic.
Narrative One: The market is no longer simply punishing high CapEx but is instead rewarding revenue delivery. Free cash flow generally deteriorated for the four major cloud giants, but their stock prices saw a clear rebound from July 29 to August 7. Amazon and Microsoft rebounded 21.1% and 28.0% from their lows, respectively, because they achieved more significant revenue and operating profit beats while raising their CapEx guidance. In contrast, Meta's Q2 operating profit missed expectations, and its free cash flow fell 91% year-on-year, causing its stock price to drop 7.95% the next day, with the subsequent stabilization lagging behind the other three cloud giants. The shift in pricing logic is clear: whether CapEx can be converted into orders, revenue, and profit is now the standard by which the market scores.
Narrative Two: Upward CapEx revisions are once again a positive for the hardware chain. After the Q2 earnings reports, except for Microsoft, which lowered its 2026E guidance from $190 billion to $175 billion due to an accounting reclassification (without actually cutting any equipment purchases), the CapEx expectations for the other cloud giants all moved higher. The total 2026E CapEx for the Big Four increased by $25 billion compared to expectations at the beginning of the year. This directly alleviated market concerns about a slowdown in the growth of orders for upstream chips, communication connections, and storage.
Narrative Three: Can model makers and new cloud providers fill the CapEx growth gap in 2027-2028? This narrative is the most complex and is currently the area of greatest market divergence. The absolute CapEx of the Big Four is still growing, but the growth rate is expected to decline sharply: from 105% in 2026E, to 45% in 2027E, and then to 9% in 2028E. To maintain the 67% growth rate of 2025, an additional $505.5 billion in capital investment would be needed in 2028; to maintain the 105% growth rate of 2026, the gap in 2028 would be as high as $1.9375 trillion.
The market's new narrative is RSI—Recursive Self-Improvement. The logic is that, unlike the past where humans trained models, the RSI model evolves into models training models, allowing for a massive number of attempts in a short time and continuous iteration without rest. The training speed far exceeds the old paradigm, thus significantly increasing the required computing power. Especially after the SSI company received a strategic investment of approximately $5 billion from Nvidia, the logic of "model makers taking over computing power demand" has gained more market attention. Systematic calculations were made for the 2027-2028 CapEx of model makers and new cloud providers: In 2027, OpenAI (Stargate project, 6GW GPU deployment agreement with AMD), Anthropic (signed agreement for up to 2GW computing system deployment with AMD, with the first 1GW planned to start in the first half of 2027), xAI (AI business CapEx will reach $73.9 billion in 2027), and new cloud providers like CoreWeave are expected to add approximately 10GW of computing power demand, corresponding to a median CapEx of about $460.9 billion. In 2028, the CapEx increment for these players is further expected to grow to a median of about $525.2 billion.
The conclusion is that if referencing the 67% growth rate of 2025, the CapEx increment from currently known projects of model makers and new cloud providers can fill the gap. However, to maintain the 105% growth rate of 2026, the $1.9376 trillion required in 2028 far exceeds the approximately $525.2 billion expected from model makers and new cloud providers, leaving a significant gap. The direction of the third narrative is correct, but it needs more project agreements to be landed for quantitative validation.
RSI Not Only Creates Training Demand, But May Also Open a Second Wave of Applications
Most of the currently widely used agents are still at the L1 level—completing subtasks like coding and debugging according to human instructions, essentially "efficient execution tools." The real change lies in the L2-L3 stages. At the L2 stage, AI no longer just follows given instructions but can autonomously modify constraints or even algorithms based on execution trajectories. For example, in financial forecasting, the model might discover the need to add new factors like "replacement cycle" and feed such improvements back into the training data, achieving high-quality data self-generation and driving the recursive evolution of its own CoT and workflow—this involves reinforcement learning and significant computing power support. At the L3 stage, the current generation of models substantially participates in the training and deployment of the next generation, with the evolution speed being so fast that humans cannot intervene or keep up.
Beyond computing power demand, RSI may also open up a new round of application demand. The market is currently concerned that the penetration rates of coding agents and office assistants are approaching a阶段性 ceiling. However, AI investment has the characteristic of "supply creating demand"—breakthroughs in model capabilities often open up new demand boundaries, similar to the wave of application penetration that followed the upgrade from 2G to 4G in communications. If AI at the L2 stage moves from passive execution to active participation in "pattern discovery" and "rule creation," the commercialization form could expand from coding agents to enterprise-level work agents, and even to new agent models whose forms are unimaginable at this stage.
Recovery Divergence: Overseas Leading, A-shares Still Awaiting Confirmation
From late July to early August, the pace and magnitude of the recovery showed clear divergence globally. In the US, stabilization began with the most severely hit AI hardware semiconductors and then spread to cloud giants. Microsoft and Amazon, which significantly beat revenue and profit expectations, rebounded the strongest, while Meta only slightly recovered its losses. The South Korean KOSPI fell approximately 39% from June 22 to July 30, then rebounded 17.9% in a single day on July 31—a典型的 oversold bounce. As of August 7, the KOSPI had stabilized 18.9% from its lowest point. Leveraged ETFs had de-levered 55.21% from their year-to-date peak, but credit financing had only de-levered 21.24%, and on-exchange derivatives had de-levered 10.50%, indicating that the clearing process is still ongoing. Notably, while the KOSPI rose 3.76% on August 5, retail investors still showed a net redemption of 2.56% in individual stock leveraged ETFs, suggesting that the rebound did not re-attract leveraged funds to chase gains.
The recovery in A-shares was relatively limited. The ChiNext index only slightly rose, lacking synchronous earnings verification and incremental funds—overseas cloud giants' capital expenditure validated that HBM orders remain strong, but this transmission chain has a relatively indirect impact on A-share technology stocks.
The AI Sector May Have Hit a Phase Top, But the Main Line Rotation Continues
A three-dimensional bubble framework shows that the expansion of real investment is only equivalent to the mid-stage of 1994-1995, far from overheating; investor sentiment is roughly equivalent to the level of 1999-2000, optimistic but not frenzied; valuation levels are similarly comparable to around 1999, requiring more fundamental delivery to support further upside. Since the end of July, with the verification of North American cloud giants' earnings, continued upward CapEx guidance, and oversold bounces in semiconductors and the South Korean memory chain, the AI sector has shown initial signs of stabilization. However, stabilization does not mean a return to a one-sided upward trend; the subsequent period is more likely to enter a phase of high volatility and high rotation. The AI sector has already formed a first阶段性 top, but this is not the end of the industry cycle.
The order of the subsequent recovery is likely to unfold along the chain of "hardware → cloud giants → model makers → applications." The hardware chain cannot maintain its position as the sole star for long; the market will gradually shift to the delivery of capital expenditure by cloud giants, the commercialization capabilities of model companies, and ultimately, the application layer. Key indicators to closely track include: whether the ARR of model makers continues to grow (OpenAI's and Anthropic's current ARR have exceeded $47 billion and $30 billion, respectively); whether cloud revenue growth can consistently cover the rapidly expanding capital expenditure (AWS's Q2 revenue grew 37% year-on-year to $42.2 billion, the highest growth rate in nearly 18 quarters); and whether the growth in usage volume can proportionally translate into revenue against the backdrop of continuously declining token prices. Only when a truly phenomenal application with user scale, retention, and willingness to pay emerges will the main line of the AI market shift from "shovel sellers" to companies that "use AI to generate cash flow." This step has not yet occurred, but it is the endpoint of the entire narrative chain.
Comments