CITIC Securities: Technology stock adjustments are not due to US bond interest rates. AI forward pricing stores three major “narrative variables”. The rate of commercialization and “anti-distillation” are the keys
The Zhitong Finance App learned that CITIC Securities released a research report saying that the recent adjustments in technology stocks cannot simply be attributed to high interest rates on US long-term bonds. There are three key narrative variables: 1) whether the rate and space of commercialization can keep up with market expectations; 2) whether the computing power advantage brings share and pricing power advantages; 3) whether the current computing power gap will clearly widen the AI model gap in the long term. Among these, the current consensus concern is the rate and space for commercialization. The biggest potential variable is whether “anti-distillation” will widen the model gap again in the future. As for macroeconomic factors, the impact of the US Treasury's announcement on the repurchase of long-term bonds is very limited. The factors that cause long-term interest rates in the US to continue to rise have not fundamentally changed, and there may be continued disturbances for some time to come. Under the influence of these external disturbances, the short-term capital structure of A-shares determines that the complexity of the market game is still increasing. During this volatile market phase, it is necessary to control psychological expectations and avoid too many grand narratives.
CITIC Securities's main views are as follows:
Recent adjustments in technology stocks cannot simply be attributed to high interest rates on US long-term bonds
Higher interest rates on US long-term bonds reflect more of the crowding out effect of AI investment on social capital, as a result of AI investment still thriving. The debt financing scale of the seven major US technology companies last year was US$87.49 billion. As of 2026, the scale has expanded to US$219.22 billion, an increase of 150.6% over the full year of last year. In the past, with huge cash reserves and operating cash flow, these companies were important marginal purchases of highly liquid assets such as treasury bonds; with the rapid expansion of AI computing power, data centers, and energy infrastructure investments, their role is shifting from the supply side of debt capital to the demand side. This means that corporate cash and social capital, which may have originally been allocated to treasury bonds, are being attracted more to the AI investment sector. High long-term interest rates and the expansion of bonds issued by technology companies are essentially two sides of the same round of AI capital expenditure boom, reflecting the repricing of limited capital between treasury bonds and AI investments. The rise in long-term interest rates is more reflected in the rise in real interest rates. It is unreasonable to use the rise in interest rates on long-term bonds to explain the decline in technology stocks.
Behind the adjustments is a forward pricing issue for AI-related stocks. There are three key narrative variables
1) Whether the rate and space for commercialization can keep up with market expectations. Investors have repeatedly disputed Anthropic's ARR since the end of May. The growth rate did slow in July, with an average monthly increase of only about 18%; however, if OpenAI is also taken into account, the slowdown is much smaller. According to a CNBC August 19 report, OpenAI CFO Sarah Friar said at an August internal meeting that ARR has increased by about 35% since the quarter. In other words, considering the two cutting-edge model manufacturers at the same time, the overall ARR growth did not slow significantly from month to month. Recently, the market has also formed a new narrative, that is, a significant proportion of cutting-edge model token consumption is not directly included in the model manufacturer's ARR, but is implemented through CSP's TaaS (Token as a Service) channel, that is, enterprises call the model via Bedrock/Vertex/Foundry within the framework of an existing AWS/Google/Azure contract, and pay per usage, and CSP is divided into model manufacturers. Looking at model manufacturers' ARR alone, they will systematically underestimate the growth in the scale of terminal payments; in fact, the growth rate of TaaS channels is faster than the growth rate of model factories' direct revenue. This narrative currently supports the optimism of many computing power investors and supports the stock prices of the core hardware manufacturers in the North American chain, but as long as agents do not have more potential commercial payment scenarios in the non-coding process, the TaaS narrative still doesn't seem to be able to attract more new capital into the market.
2) Does the computing power advantage bring share and pricing power advantages. According to the AI Index of enterprise expenditure management company Ramp, OpenAI's share of API spending covered by more than 70,000 US companies rose from 28.5% in May and 28.1% in June to 36.0% in July, while Anthropic fell from 71.2% to 63.4%. At the model level, this change was almost entirely contributed by GPT-5.6 Sol. The model's share of expenditure in May was still zero, and it alone accounted for 14.9% in July. In terms of the number of paying companies rather than the amount of expenses, Anthropic is still leading with about 44% to 40%, but OpenAI's growth momentum has clearly recovered. One possible explanation is that OpenAI currently has more computing power, so it is more relaxed in terms of the pace of release of next-generation models and the supply of inference capacity, so “grabbing more computing power to obtain a higher application market share” seems to constitute a reason for model manufacturers to continue to bet on computing power. However, there is a key problem: static shares are not exactly equivalent to pricing power. If the capabilities and agent functions of cutting-edge models tend to be homogenized, and user switching costs are not high and stickiness is very low, then the significance of static share is limited, and overall market space is far more important than share allocation. According to OpenRouter data, the weekly token usage of the Anthropic model declined significantly after mid-July. At the same time, the usage of tokens for models such as OpenAI, Deepseek, and Minimax all increased markedly. In a situation where capabilities converge, computing power advantages bring phased shares rather than sustainable excess profit margins. Ultimately, we still have to go back to the first problem, which is the overall commercialization space.
3) Whether the current computing power gap will clearly widen the AI model gap in the long term. The answer to this question has the greatest impact on the forward pricing of computing power “sellers” because it directly affects the market's expectations about the intensity and continuity of the computing power competition. Currently, the key influencing the answer to this question may be whether cutting-edge model manufacturers can use some means to “prevent distillation” in the future to transform the computing power advantage of the training side into the technical cost difference of next-generation models, thereby obtaining pricing power. In mid-August, researchers from institutions such as MATS Research and the ELLIS Tübingen Institute published a paper called “Stealing Traces from Inference LLM APIs”, which analyzed in detail the extractability of cutting-edge model inference chains under current mainstream API architectures. The paper did not make a conclusion on model distillation, but there is plenty of evidence that at least indicates that the inference capacity barriers formed by large manufacturers spending huge amounts of computing power do indeed face the risk of being overtaken by low costs. If this problem persists, then the pricing of computing power “sellers” will eventually return to the traditional public infrastructure chain pricing model, exchanging time for space, and explosively decline. However, the market currently also has another expectation that cutting-edge model manufacturers may solve the anti-distillation problem and simultaneously release a new generation of cutting-edge models with significantly stronger capabilities, then the scaling advantages on the training side will be transformed into long-term competitive barriers and pricing rights. The intensity of the computing power competition will continue to be upgraded, and AI hardware will also be treated as scarce resource pricing rather than a public infrastructure chain.
The impact of the US Treasury buyback is limited, but weakening interest rate hike expectations are conducive to the convergence of K-type differentiation in the global market
The US Treasury Department announced on August 19 that it will expand the scale of repurchases of long-term treasury bonds to provide greater liquidity support to the bond market. According to the statement, the scale of long-term treasury bond repurchases will be raised from 2 billion US dollars to 4 billion US dollars, and the repurchases will cover 10-year to 30-year bonds. Compared to the US debt stock held by the public (held by the public) US debt stock that surpassed $32 trillion, the impact of the $4 billion repurchase scale on the bond market is limited. The more practical significance of this operation is to reinforce the expectation that “the Ministry of Finance will adopt 'YCC-like' regulation when long-term interest rates rise in a disorderly manner”, guide the market to form an expected upper limit on long-term interest rates and suppress the tail risk of term premiums. The negative effects of this approach are also obvious. It may further deepen the market's distrust of fiscal discipline and increase the sell-off of US debt. The direct impact of these macro-narrative changes is to weaken expectations of the Fed's interest rate hike during the year. The indirect impact is to promote the convergence of K-type differentiation in the global market, because the non-AI sector is more sensitive to interest rate costs than the booming AI sector.
The factors responsible for the continued rise in long-term interest rates in the US have not changed
First, the high return on AI hardware investment in the short term will still crowd out demand in the bond market and push up actual interest rates. Currently, in an environment where computing power is scarce, the return on static investment in data centers is still impressive, and the cloud business EBITDA rate of major CSP vendors is still rising. As long as the computing power competition continues, AI will continue to squeeze out demand in the treasury bond market and drive up long-term interest rates. Second, the price increase of energy and chemical products is more sticky than at the beginning of the US-Iran war. The impact of inventory reduction since the second quarter on supply and demand buffers in the crude oil market has decreased. As China accelerates broad fiscal spending in the second half of the year, the effect of suppressing demand is also decreasing, and the probability that the Strait of Hormuz issue will “end” is still increasing. The impact of these factors on inflation expectations in Europe and the US may heat up again.
Under the influence of external disturbances, the short-term capital structure of A-shares determines that the complexity of the market game is increasing
According to CITIC Securities channel research data, active private equity increased their positions sharply during the rebound in the first week of August, from 71.7% at the end of July to 79.0%, increasing their positions by 7.3 pcts in a single week, which is the second largest weekly increase since 2017 (after 7.8 pcts for the week of October 12, 2018). In addition, according to data from the Private Equity Ranking Network, as of August 14, 2026, the stock position index for large-scale private placement (with a management scale of more than 5 billion yuan) had reached 88.56%, setting a new high during the year; at the same time, large-scale private equity accounts for 77.11% of full positions, which also hit a new high during the year. It can be seen from this that since August, capital with the most aggressive risk appetite in the market has increased its positions and driven this wave of rebound. For a market that has gone long unilaterally, such as A-shares, optimistic expectations have largely been priced into it. At the same time, the active public equity index is still highly correlated with the trend of telecommunications and semiconductor ETFs, and there has been no significant adjustment in the position structure. Unlike the typical “collapse” market in history, “avoiding institutional stocks” is currently not an effective strategy. After controlling market capitalization factors, the bank found that the market rebound since August, whether broad base, theme, industry, or style, there was no clear correlation between the rise and fall of individual stocks and the size of institutional holdings. In many industries, individual stocks with a higher percentage of institutional holdings rebounded even more (after controlling the market capitalization factor). The bank believes that not so much that the market has been avoiding institutional tickets since August, it is more like shunning large-cap stocks. This may be related to changes in the strategic exposure of quantitative capital in the market.
During the turbulent market phase, it is necessary to control psychological expectations and avoid too many grand narratives
Currently, many sectors have performance and prosperity, but there is no room for valuation improvement in the short term. For example, North American AI, domestic computing power, non-color, energy storage, innovative drugs, etc., all have similar characteristics. Investors should be cautious when optimistic narratives prevail, and when the risk of killing valuations explodes, it instead forms a layout point. The market has just experienced a few months of high sector volatility. Controlling psychological expectations come first, and it is necessary to avoid frequently falling into grand narratives. In terms of allocation strategies, within the technology sector, it is recommended to use AI to rebound in price increases, adjust positions to core assets (such as gas turbines, wafer manufacturing platforms, semiconductor equipment, etc.) in a timely manner, pay more attention to “quantitative certainty”, and treat “price explosiveness” carefully. For the non-technology sector, it is recommended to focus on adding energy, non-color, innovative drugs, and leading brokerage firms with the potential to go overseas.
Risk factors: Frictions in the fields of technology, trade, and finance between China and the US have intensified; China's policy strength, implementation effects, or economic recovery have fallen short of expectations; macro-liquidity at home and abroad has tightened beyond expectations; conflicts between Russia, Ukraine, the Middle East and other regions have further escalated; and China's real estate inventories have fallen short of expectations.