CITIC SEC: The adjustment in technology stocks is not due to U.S. Treasury yields. There are three major "narrative variables" in AI forward pricing, and the speed of commercialization and "distillation prevention" become key factors.

CITIC SEC has released a research report stating that the recent adjustments in technology stocks cannot be simply attributed to the high long-term bond yields in the U.S. The underlying issue behind the adjustments is related to the forward pricing of AI-related stocks, which involves three key narrative variables: 1) whether the pace and scope of commercialization keep pace with market expectations; 2) whether computational power advantages bring share and pricing power advantages; 3) whether the current gap in computational power will clearly widen the long-term AI model gap. Currently, the consensus concern revolves around the pace and scope of commercialization, with the biggest potential variable being whether "anti-distillation" will lead to a re-widening of the model gap in the future. As for macro factors, the impact of the U.S. Treasury's announcement to repurchase long-term bonds is very limited, and the factors causing the persistent rise in long-term U.S. yields have not fundamentally changed, suggesting that there may be ongoing disturbances in the near future. Given these external disturbances, the short-term capital structure of A-shares is determining that the complexity of market dynamics is on the rise, and during this stage of market volatility, it is important to control psychological expectations and avoid excessive grand narratives. CITIC SEC's main viewpoints are as follows: The recent adjustments in technology stocks cannot be simply attributed to high U.S. long-term bond yields. The high long-term bond yields in the U.S. reflect more of the crowding-out effect of AI investments on social capital, resulting from still robust AI investment. The seven major U.S. tech giants raised $87.49 billion in debt financing last year, which has expanded to $219.22 billion as of 2026, a growth of 150.6% compared to last year. In the past, these companies were significant marginal buyers of high-liquidity assets like government bonds due to their large cash reserves and operational cash flows; however, with the rapid expansion of investments in AI computing power, data centers, and energy infrastructure, their role is shifting from a supply side of debt capital to a demand side. This means that cash and social capital that would have been allocated to government bonds are increasingly attracted to AI investment. The high long-term yields and the debt expansion of tech companies are essentially two sides of the same boom in AI capital expenditures, reflecting a re-pricing of limited capital between government bonds and AI investments. The rise in long-term yields primarily reflects an increase in real interest rates; attributing the decline in tech stocks to rising long-term yields is not reasonable. The adjustments are rooted in the forward pricing issues of AI-related stocks, which involve three key narrative variables. 1) Whether the pace and scope of commercialization keep pace with market expectations. Since the end of May, there has been repeated debate among investors regarding Anthropic's ARR, and while the growth rate did slow down in July, averaging only about 18% growth from May to July, if we also consider OpenAI, the slowdown is much less pronounced. According to a report by CNBC on August 19, OpenAI's CFO Sarah Friar stated in an internal meeting in August that the ARR growth has been about 35% for the quarter to date. This indicates that when considering both cutting-edge model companies, the overall month-on-month ARR growth has not significantly slowed. Recently, the market has formed a new narrative suggesting that a considerable proportion of token consumption from cutting-edge models is not directly accounted for in the models' ARR but is realized through the CSP's TaaS (Token as a Service) channels, meaning enterprises invoke models and pay per usage within the existing AWS/Google/Azure contract frameworks through Bedrock/Vertex/Foundry, and CSP shares the earnings with model providers. Relying solely on the ARR of model providers would systemically underestimate the growth of end-user payment scale; the growth rate of the TaaS channel is actually faster than the direct revenue growth rate of model providers. This narrative currently supports the optimism of many investors in computing power and provides support for the stock prices of key hardware manufacturers in North America, but as long as more promising commercial payment scenarios for agents have not emerged in non-coding segments, the TaaS narrative seems unlikely to attract more new funds into the market. 2) Whether the computational power advantage brings advantages in market share and pricing power. According to AI Index from Ramp, which covers API expenditures of over 70,000 U.S. companies, OpenAI's share increased from 28.5% in May and 28.1% in June to 36.0% in July, while Anthropic dropped from 71.2% to 63.4%. Breaking it down to the model level, this change is almost entirely contributed by the GPT-5.6 Sol model, which had a spending share of zero in May but accounted for 14.9% in July. If we consider the number of paying enterprises rather than the expenditure amount, Anthropic still leads by about 44% to 40%, yet OpenAI's growth momentum has clearly recovered. One possible explanation is that OpenAI currently has more computational power, thus being more relaxed in releasing new generation models and providing reasoning capacity, which suggests that "capturing more computational power leads to obtaining higher application market shares." However, a critical issue remains: static shares do not entirely equate to pricing power. If the capabilities of cutting-edge models and agent functionalities tend to homogenize, and the switching costs for users are low and stickiness is weak, then the significance of static share becomes limited, as the overall market potential is far more crucial than share distribution. According to OpenRouter data, weekly token usage of Anthropic models notably declined after mid-July, while token usage for OpenAI, Deepseek, Minimax, and others has significantly increased. In conditions of capability convergence, the computational power advantage yields only a temporary share rather than sustainable excess profit margins, ultimately bringing us back to the first question regarding the overall commercialization scope. 3) Whether the current gaps in computational power will significantly widen the long-term AI model gaps. The answer to this question has the greatest impact on the forward pricing of "computational power sellers," as it directly affects market expectations regarding the intensity and sustainability of the computational power competition. Currently, a key factor influencing this question might be whether frontier model companies can implement certain strategies to "prevent distillation," thereby converting their computational power advantages during training into technological disparities in the next generation of models, thus gaining pricing power. In mid-August, researchers from MATS Research, ELLIS Tbingen Institute, and other institutions published a paper titled "Stealing Reasoning Traces from Proprietary LLM APIs," which detailed the extractability of reasoning chains of cutting-edge models under current mainstream API architectures. Although the paper did not reach a definitive conclusion on model distillation, several pieces of evidence at least suggest that the embedding barriers created by the significant computational power of large companies indeed face the risk of being pursued at lower costs. If this problem persists, then the pricing of "computational power sellers" is likely to revert to a traditional public infrastructure pricing model, trading time for space, and resultant drops. However, the market currently also harbors another expectation that frontier model companies may solve the anti-distillation problem by the end of the year while simultaneously releasing noticeably more powerful next-generation frontier models. In this case, the scaling advantages in the training phase will translate to long-term competitive barriers and pricing power, indicating that the intensity of the computational power competition will continue to escalate, and AI hardware will be priced as a scarce resource rather than as part of a public infrastructure chain. The impact of the U.S. Treasury's repurchase is limited, but weakening rate hike expectations are favorable for a convergence of global market K-type differentiation. On August 19, the U.S. Treasury announced an expansion of its long-term government bond repurchase scale to provide greater liquidity support for the bond market. According to the statement, the repurchase scale will increase from $2 billion to $4 billion, covering 10 to 30-year bonds. Relative to the $32 trillion public-held U.S. debt stock, the $4 billion buyback scale has a limited impact on the bond market; practically, this operation may further reinforce the expectation that "the Treasury will adopt 'quasi-YCC' control measures during disorderly upward movements of long-term rates," guiding the market towards a cap on long-term yield expectations and suppressing tail risks of term premiums. The negative effects of this approach are also evident, potentially deepening the market's distrust in fiscal discipline, exacerbating the sell-off of U.S. debt. The direct impact of these macro narrative changes is to weaken the expectation of Fed rate hikes this year, with indirect effects promoting the convergence of global market K-type differentiation, as non-AI sectors are more sensitive to interest rate costs compared to the more prosperous AI sectors. The factors causing the persistent rise in U.S. long-term yields have not changed. Firstly, the short-term high returns on AI hardware investments will continue to crowd out demand in the bond market, raising real interest rates. Currently, the static return rates on data center investments remain attractive under a computational power shortage environment, while the EBITDA rates of major CSPs' cloud businesses continue to rise; as long as the computational power competition persists, AI will continue to crowd out government bonds' market demand, pushing long-term yields higher. Secondly, price increases in energy commodities are stickier now than at the onset of the U.S.-Iran war, and the inventory reduction measures since Q2 have lessened their cushioning effect on oil market supply and demand. As China accelerates its broad fiscal spending in the second half of the year, the demand suppression effect is decreasing, and the probability of the situation in the Strait of Hormuz becoming unresolved is increasing; these factors may reignite impacts on inflation expectations in Europe and the U.S. Under the influence of external disturbances, the short-term capital structure of A-shares is determining that the complexity of market dynamics is on the rise. Data from CITIC SEC's channel research indicates that active private equity significantly increased positions during the rebound in the first week of August, rising from 71.7% at the end of July to 79.0%, marking a weekly increase of 7.3 percentage points, the second-largest weekly increase since 2017 (only behind the 7.8 points in the week of October 12, 2018). Additionally, data from Private Placement Network shows that as of August 14, 2026, the stock position index of large private equity firms (managing over 5 billion yuan) has reached 88.56%, setting a new high for the year; at the same time, the proportion of fully invested large private equity firms stands at 77.11%, also a year-to-date high. This indicates that since August, the most aggressively risk-seeking funds in the market have increased positions and propelled this rebound. For a market like A-shares that tends to make one-sided bullish bets, optimistic expectations are now largely priced in. Meanwhile, the correlation between active public equity fund indices and those for ETFs in communications and semiconductors remains high, with no significant adjustments in holding structures. Unlike previous typical scenarios where "collapses of overcrowded stocks" occurred, currently avoiding institutional stocks does not seem to be an effective strategy. After controlling for market capitalization factors, this institution found that the market rebound since August, whether broad-based, thematic, industry-based, or style-based, has shown no clear correlation between stock price movements and institutional holdings, with many stocks with higher institutional holding ratios actually rebounding more (after controlling for market capitalization factors). This institution believes that instead of saying the market has avoided institutional stocks since August, it might be more accurate to say the market is avoiding large-cap stocks, possibly related to changes in quantitative fund strategies. In this phase of a volatile market, it is crucial to control psychological expectations and avoid excessive grand narratives. Currently, many sectors show performance and favorable conditions but lack visible short-term valuation improvement potential, such as North American AI, domestic computing power, non-ferrous metals, energy storage, and innovative pharmaceuticals, which share similar characteristics. Investors should be cautious during periods when optimistic narratives dominate, while the eruption of valuation risks may provide opportunities for positioning. The market has just experienced a few months of high volatility among sectors, so controlling psychological expectations should take precedence, avoiding frequent entrapment in grand narratives. In terms of allocation strategies, within the technology sector, it is recommended to timely adjust towards core assets (such as gas turbines, wafer fabrication platforms, semiconductor equipment, etc.) in response to rebounds in AI pricing varieties, placing greater emphasis on "certainty in volume" and cautiously approaching "explosive pricing." For non-tech sectors, it is advisable to focus on increasing allocations in energy, non-ferrous metals, innovative pharmaceuticals, and leading brokers with overseas potential. Risk factors: intensifying friction in technology, trade, and finance between China and the U.S.; the policy, implementation effects, or economic recovery in China may not meet expectations; unexpected tightening of domestic and foreign macro liquidity; further escalation of conflicts in regions such as Russia-Ukraine and the Middle East; slow digestion of real estate inventory in China compared to expectations.