CICC: Is AI shifting from a "friend" of traditional demand to an "enemy"?

Since July, the global AI chain has continued to adjust. Behind this, besides rising costs, more importantly, the market worries that the returns AI can generate cannot sustain large-scale investment. For example, the market appears to be worried about CoreWeave's credit risk, but behind that may be worries about a slowdown in Anthropic's ARR, so the market needs to calculate the account of ROIC (return) versus WACC (cost), which is exactly the same as the "script" of the market's worries about private credit in the fourth quarter of last year. Therefore, the key to determining the subsequent trajectory of AI lies in the iteration of large models and the realization of new business models (such as Anthropic's better-than-expected performance in the first quarter of this year), which is precisely the key to raising AI returns.

Chart: The surge in CoreWeave's CDS (default concerns) triggers a round of selling

Source: Bloomberg, CICC Research Department

Chart: As of the second quarter, the 19.9% ROIC of US cloud providers can still cover the 9.6% WACC

Source: FactSet, Bloomberg, CICC Research Department

In addition to having to "work hard" to prove that their own returns are sufficient to cover ever-expanding capital expenditures, AI in both China and the US has also reached another critical juncture: that is, their own capabilities (the money they earn themselves, operating cash flow) can no longer support their own ambitions (the investment demand of capital expenditures), and they need to obtain support through external financing (free cash flow turning negative), and can no longer invest only with the money they earn themselves as before. Of course, borrowing money is not the problem; the problem is borrowing money yet failing to earn matching returns.

But in any case, as AI in both China and the US begins to draw resources from the whole of society, another increasingly important question has gradually surfaced: will AI gradually move from being a "friend" that drives other economic sectors to gradually moving toward its opposite, becoming an "enemy" that crowds out traditional demand? In other words, can the upper half of the K-shaped divergence continue to offset the lower half, so that overall growth still trends upward?

Both China and the US actually face this problem: the recent rise in US Treasury yields, the exposure of CoreWeave risks, and the weakness of China's domestic demand are all some kind of refraction of this problem. It is just that the transmission and crowding-out paths, tolerance levels, and critical thresholds differ between China and the US. Clarifying this "critical threshold" also clarifies the boundary that the economy and policy can withstand, which is of great benefit to us in judging the policy paths of China and the US and choosing investment directions.

In the second quarter of this year, US AI reached a critical juncture: that is, it cannot use the money it earns itself (operating cash flow) to support increasingly inflated capital expenditures, and therefore needs external financing (equity or debt) to fill the funding gap. The landmark event is that in the second quarter of this year, the free cash flow of major cloud providers (Hyperscalers) other than Microsoft and Meta all turned negative (operating cash flow minus capital expenditures).

Chart: US cloud providers' capital expenditures approach operating cash flow

Source: FactSet, CICC Research Department

Of course, this is not to say that borrowing money is a bad thing; on the contrary, borrowing money can leverage more funds and can also raise ROE (ROE = profit margin * turnover * leverage). Only borrowing money yet failing to earn back the corresponding money is a bad thing. For cloud providers, as of the second quarter, a 19.9% ROIC can fully cover a 9.6% WACC, not to mention corporate bond financing of less than 5-7%.

This is so in the US, and similar in China. Moreover, because China's AI investment ROIC is lower (as of the second quarter, the weighted ROIC of China's leading cloud providers such as Tencent, Alibaba, and Baidu was about 11.9%, lower than the 19.9% of US cloud providers), and free cash flow in the second quarter has all turned negative, more external funding support is needed. In an environment where factor resource allocation is not entirely determined by market pricing, the crowding out of traditional demand may be faster.

Chart: As of the second quarter, the weighted ROIC of China's leading cloud providers was about 11.9%, lower than that of US cloud providers

The economic impact of a technological revolution cannot be observed only in the technology-producing sector. The impact of a technological revolution on the economy may come from investment and production in the new technology itself, may come from productivity improvements and new supply creation brought about by the application of the new technology, and may also affect sectors that have not yet effectively adopted the new technology through factor prices. It is not difficult to understand that the impact of new technology on different sectors differs markedly. Referring to the approach of Stiroh (2002), we divide the economy into four sectors to better reflect the impact of the technological revolution. Specifically:

1) The AI production sector, mainly involving the production links of AI large models, such as computers, software, data centers/cloud services, and IT services;

3) The AI application sector, including finance, professional services, R&D, pharmaceuticals, film and television media, etc.;

4) Other sectors, sectors not directly related to AI investment and consumption, such as food and beverages, real estate, transportation and warehousing, culture and entertainment, etc.

For the US, as of the first half of 2026, compared with 2022, the share of the AI production sector rose by 2.4ppt to 10.3%; the AI application and infrastructure sectors were basically flat, with a combined share of about 40%. Other sectors still exceeded 50%, but were down 1.4ppt compared with the end of 2022. In short, AI-related and non-related sectors are roughly split fifty-fifty.

For China, AI-related and non-related sectors are also roughly split fifty-fifty, but the structure of AI-related sectors differs markedly from that of the US. As of the end of 2025, the AI production sector accounted for less than 5%, only half that of the US; the share of the AI application sector was almost the same as that of the US; however, the share of the AI infrastructure sector was close to 14%, about twice that of the US; other sectors accounted for about 50%, close to the US.

Source: BEA, CICC Research Department

Chart: China is also roughly split fifty-fifty, but its structure differs considerably from that of the US, with a higher share of infrastructure

Source: Wind, CICC Research Department

► Directly driving growth through investment and exports. This has been particularly evident over the past few quarters, in both the US and China. In Q2, US real GDP grew at a seasonally adjusted annualized rate of 2.2% quarter-on-quarter; excluding the import factor, broad AI demand (AI production investment and infrastructure) contributed 28% of US growth, adding about 0.6ppt. Over the same period, China's real GDP grew at a seasonally adjusted annualized rate of 4.77% quarter-on-quarter, with AI contributing 47%. For China, the boost was mainly reflected in exports: AI-related products contributed more than 60% of the year-on-year increase in net exports, and net exports contributed 21% of GDP growth, adding 0.6ppt to GDP growth. The same is true of earnings: AI-related sectors contributed 53% of US equity earnings growth and 19% and 2.4% of A-share and Hong Kong equity earnings growth, respectively. In other words, strip out AI from the US, and half of earnings are gone; strip out AI from China, and half of growth is gone.

Chart: In Q2, US real GDP grew at a seasonally adjusted annualized rate of 2.2% quarter-on-quarter, with AI-related demand contributing 28% of growth

Source: Haver, CICC Research Department

Chart: In Q2, China's real GDP grew at a seasonally adjusted annualized rate of 4.77% quarter-on-quarter, with AI-related demand contributing 47% of growth

Source: Wind, CICC Research Department

► The wealth effect from market gains. Stock market gains during the AI rally generate a pronounced wealth effect, which in turn lifts households' consumption expectations. Using empirical tests on total wages in local non-tradable sectors (such as retail, food services and other service industries directly driven by local demand), Chodorow-Reich et al. (2020) show that for every $1 increase in stock wealth, households increase real consumption spending by about 3.2 cents per year. This is more evident in the US, thanks to stronger US equity performance and households' greater equity market exposure (401K plans, etc.). According to Federal Reserve data, in 2Q26, individual and other non-institutional investors held more than 42% of total US stock market capitalization, up 1.4ppt from the start of the year; equity-type assets accounted for 47% of households' balance sheets, up 2ppt from the start of the year; and dividends and other such income accounted for more than 9% of disposable income so far this year. By contrast, China's wealth effect still has room to improve.

Chart: After a stock market wealth shock, total wages in non-tradable sectors rise

Source: Chodorow-Reich et al., (2020), CICC Research Department

Chart: In 2Q26, equity-type assets accounted for nearly 47% of US household assets

Source: Bloomberg, CICC Research Department

Chart: So far this year, the median return of the 5,500 A-share companies is -17%

Source: Wind, CICC Research Department

Chart: As of 2025, the share of household assets allocated to equities remains very low

Source: CHFS, CICC Research Department

► Positive externalities from higher labor productivity and knowledge spillovers. The first principle behind AI's boost to the economy is its improvement in labor productivity by completing tasks at relatively low cost and high efficiency. Anthropic estimates that AI assistance can shorten the completion time of some tasks by an average of 80%, at a lower relative labor cost. As large-model application scenarios become increasingly rich, overall labor productivity in the US nonfarm sector has improved. Among industries with heavier AI investment, such as information technology and finance and insurance, labor productivity improvements have been particularly pronounced. Better production efficiency, on the one hand, raises total factor productivity and, on the other hand, drives continued rapid growth in token usage, sustained increases in cloud providers' computing investment, and rapid expansion of demand in AI computing power, infrastructure and related segments.

Knowledge spillovers generate positive externalities: the spread of models, software, data and organizational experience to more firms can lower subsequent costs, gradually turning other sectors that previously had lower AI usage intensity into application sectors. Knowledge diffusion may also generate positive externalities that are not fully priced, making AI's social returns higher than its private returns. For 2025, Brynjolfsson et al. (2026) estimate that US users derived about $116.2 billion in annual consumer surplus from generative AI chat tools. Over the same period, Gartner forecasts that global end buyers' spending on generative AI models was only $14.2 billion, so the former is about 8.2 times the latter. By comparison with other industries, every $1 in base fare paid by UberX users generates about $1.57 in consumer surplus, mobile communications generates 1 times consumer surplus, and generative AI reaches as high as 8.2 times. The results show that AI's social returns are significantly higher than its private returns, so measuring its economic contribution solely by model market transaction value would systematically omit the largest piece of social welfare value.

However, once the returns AI generates for society (its own ROIC plus positive externalities) fall below the resources AI crowds out from other sectors, AI may turn from a "friend" of traditional demand into an "enemy." AI beginning to require external financing and to push up raw material prices is a typical example of factor resources crowding out traditional demand, reflected in three aspects:

► Crowding out physical resources. AI consumes key resources such as electricity and pushes up the marginal cost and capacity utilization of key factors in other sectors. AI capital expenditure is not merely a financial outlay; it also requires physical inputs such as chips, servers, electricity, grid connections, and land. When the supply of the relevant inputs lacks elasticity in the short term, new AI projects may crowd out the physical resources of other sectors by pushing up electricity prices, equipment delivery prices, construction costs, and high-skilled wages. Since the end of 2022, U.S. private-sector spending on data center construction has grown nearly fivefold. Globally, data centers' share of electricity consumption in electricity generation rose from 1.0% in 2020 to 1.5% in 2025; in the United States in particular, this share rose rapidly from 2.6% in 2020 to 4.9% in 2025. In addition, over the past few years, with the development of AI, capacity utilization in U.S. industries such as oil and gas, electrical equipment, mining, machinery, and electronics has risen markedly. Since the start of this year, memory prices have risen 110%, pushing the year-on-year growth rate of the electronic components and accessories sub-index in U.S. PPI from 4% at the end of last year to as high as 29% in June, making it one of the main contributors to the year-on-year rise in U.S. PPI.

Chart: Data centers' share of electricity consumption in electricity generation has been rising steadily worldwide, especially in the United States

Source: IEA, Ember (2026), CICC Research Department

► Crowding out financial resources. After AI crosses the threshold at which it begins to require external financing, its crowding out of other sectors is reflected in two dimensions: price and quantity. 1) In terms of price, it pushes up overall financing costs. Strong AI demand, combined with external supply disturbances such as oil prices, has jointly pushed up costs. For example, as U.S. Treasury yields have risen, the 30-year mortgage rate has climbed to a high of 7.5%, and the short-term loan rate for small businesses has also risen markedly to 8.2%. Looking at credit spreads, the credit spread on U.S. investment-grade corporate bonds has not risen significantly relative to the technology sector, but high-yield bond spreads have widened markedly.

Chart: The credit spread on overall U.S. investment-grade corporate bonds has not risen significantly relative to the technology sector

2) In terms of quantity, massive AI bond issuance will also crowd out financing demand, including for U.S. Treasuries and other borrowing, but this is not yet significant. On the one hand, additional long-term investment-grade bond issuance by AI giants will crowd out long-only funds' allocation demand for U.S. Treasuries, and combined with the "mini-peak" in Treasury supply in the third quarter, this jointly led to the "Treasury storm" in August. On the other hand, AI bond issuance will also crowd out financing by non-AI companies; in particular, the market expects the issuance of credit bonds by U.S. technology leaders to increase by another $200–300 billion in 2027, though the crowding-out effect is not yet significant. The technology sector (information technology plus communications) has risen to 40–50% of U.S. non-financial corporate bond supply, but non-financial, non-technology issuance has not declined significantly. Even after deducting maturities, net investment-grade supply from non-financial, non-technology sectors has continued to grow.

Chart: The technology sector (information technology plus communications) has risen to 40–50% of U.S. non-financial corporate bond supply

Source: Bloomberg, CICC Research Department

Chart: Net investment-grade corporate bond supply from U.S. non-financial, non-technology sectors has continued to grow

Source: Bloomberg, CICC Research Department

► Constraining policy space. Strong AI demand also constrains the Federal Reserve's policy space. The U.S. economy is also undergoing K-shaped divergence: traditional demand, represented by real estate, is being suppressed by high interest rates, and existing home sales weakened again after May. By rights, the Federal Reserve should cut rates to boost traditional demand, but now it has to raise rates to cope with stronger investment and higher inflation. According to Federal Reserve estimates, the U.S. natural rate of interest has risen from 0.99% at the end of 2022 to 1.44%. Year to date, most of the 79bp rise in the 10-year Treasury yield has been contributed by the real interest rate (71bp) rather than inflation expectations (9bp).

Chart: Year to date, most of the rise in the 10-year Treasury yield has been contributed by the real interest rate (71bp) rather than inflation expectations (9bp)

Source: Bloomberg, CICC Research Department

Compared with the United States, China's crowding out is mainly reflected in fiscal and credit resources. On the one hand, fiscal resources are tilted more toward new quality productive forces such as the construction of the "six networks"; on the other hand, the allocation of credit resources also reflects a clear policy orientation. China's new loans are concentrated in technology and other fields, while the share of loans related to real estate and consumption has fallen sharply. Since the start of this year, consumption has remained weak and household balance sheets have contracted.

► Replacing some jobs. While AI improves production efficiency, it also substitutes for employment, especially in industries with high AI exposure. Since ChatGPT was launched at the end of 2022, U.S. nonfarm payrolls have increased by 4.7 million, but the gap across industries is striking: employment rose in education and health (+2.78 million), government (+610,000), and leisure and hospitality (+440,000), while industries more clearly affected by AI, such as business services (-370,000), information (-300,000), and financial activities (-90,000), saw larger declines. In addition, according to research by the Dallas Fed[2], wage growth is higher in industries with high AI exposure, but the wage gap is larger between younger groups and experienced groups, which may pose greater challenges for young job seekers.

Source: Haver, CICC Research Department

Chart: U.S. technology and financial employment has fallen markedly since the end of 2022

Source: Haver, CICC Research Department

V. When will AI turn from a "friend" into an "enemy"? The paths and tipping points differ between China and the United States; U.S. rates at 6%, China has a high tolerance for policy

Now that AI has begun drawing resources from society, the next question is: when will AI turn from a "friend" into an "enemy"? This is essentially a matter of estimating the costs that the economy and policy can bear—for example, when resources are all directed toward technology, how long can the upper end of the K-shape offset the lower end? When will the Fed stop raising rates, or even pivot to cutting? When will Chinese policy pay more attention to weaker domestic demand?

In the specific estimation process, it should be noted that because the sources of AI funding and policy objectives differ between China and the United States, the crowding-out paths and tipping points also differ: 1) The vast majority of U.S. AI financing comes from the private sector, so crowding out occurs mainly through interest rates and Fed policy; 2) In China, at least 30% comes from policy-related funds, and bank credit and other channels are also policy-guided, so crowding out occurs more through the tilt of fiscal and credit resources. In addition, China focuses more on long-term strategy, so its tolerance for short-term crowding out is also higher.

► How high an interest rate can the United States bear? About 6% overall, above 7% for AI, and about 4% for traditional sectors. The difference between investment returns and financing costs is significant for U.S. AI and traditional demand sectors. The ROIC of the five major cloud vendors (20%) remains far above their WACC (9%), while for traditional demand such as rate-sensitive real estate, the mortgage rate (7.5%) has long been above the rental yield (6.86%). Within U.S. equities, the gap between AI and non-AI is likewise large. We take the information technology and communication services sectors in the S&P 500 plus Amazon in the consumer discretionary sector as the representative of U.S. AI equities, and the remaining 406 constituents as the representative of non-AI traditional demand. As of the third quarter of 2026, the gap between ROIC (about 35%) and WACC (10–15%) for the AI segment was as high as 21%, while for non-AI it was only 3%.

Chart: As of Q3 2026, the ROIC-WACC gap for the U.S. AI segment was as high as 21%, while for non-AI it was only 3%

Source: Bloomberg, FactSet, CICC Research Department

AI can bear an interest rate of about 7–7.5%. We assume ROIC remains unchanged; if the gap between ROIC and WACC for the five major cloud vendors narrows to the average of 8ppt before the AI trend emerged in 2015-2023, the corresponding WACC would be about 11.9%. Taking into account the impact of rising U.S. Treasury yields on bond costs and equity costs, and assuming β remains unchanged, we back out a risk-free rate of about 7–7.5%.

The traditional sector can bear an interest rate of about 4%. 1) Taking real estate as an example, assuming the 30-year mortgage rate converges with the rental yield in recent years (6.2%), and then using the averages since 2015 of the MBS spread and the primary-secondary spread among the components of the mortgage rate (split into the 10-year U.S. Treasury yield, the MBS spread, and the primary-secondary spread), we derive a corresponding 10-year U.S. Treasury yield of about 4–4.2%. 2) For the "non-AI" segment of the S&P 500 (excluding information technology, communication services, and Amazon), assuming the gap between ROIC and WACC returns to the average of 4.5% since 2015, the corresponding WACC is 9.7%, and regression yields a corresponding 10-year U.S. Treasury yield of 4%.

The overall U.S. economy can bear an interest rate of about 5.5–6%, considering: 1) AI contributes about 30–50% to U.S. GDP; 2) the market-cap weights of AI and non-AI in the S&P are roughly "fifty-fifty," giving a weighted interest rate that the overall U.S. economy can bear of about 5.5–6%.

Chart: Assuming ROIC remains unchanged, if the gap between cloud vendors' ROIC and WACC returns to the average, the corresponding U.S. Treasury yield is 7–7.5%

Source: Bloomberg, FactSet, CICC Research Department

Chart: If the mortgage rate converges with the rental yield in recent years (6.2%), the corresponding 10-year U.S. Treasury yield is around 4%

Source: Bloomberg, Haver, CICC Research Department

► How high a cost increase can the United States bear? About 50% overall, 4–5% for AI application sectors, and 8–10% for other sectors. AI production investment on the one hand boosts demand in the upstream supply chain, but also drives up prices of raw materials and related goods, squeezing profits in other sectors. Our estimates show that assuming cloud vendors' capital expenditure reaches $900 billion, after excluding the import leakage effect, if AI-related costs rise 50%, then the $161.7 billion in total gross operating surplus (GOS) generated by capital expenditure would be offset by cost increases and import leakage; when costs rise 3.6%, the gains that application sectors obtain from capital expenditure would be crowded out by cost increases; if costs rise 8.6%, the gains of other sectors would also be wiped out. The AI production and infrastructure sectors as a whole suffer no obvious losses during cost increases, and some even benefit.

Source: BEA, CICC Research Department

► How much crowding out can Chinese policy bear? Because policy focuses more on long-term strategic objectives and cannot be simply quantified, the policy tipping point needs to be "estimated" from three dimensions: 1) Growth. The "Tutorial 100 Questions" points out that to achieve the long-term growth goal, per capita GDP should reach the level of moderately developed countries by 2035; 2) Employment, which concerns social stability and has not yet returned to the low point before the "924" policy push; 3) Leverage and balance sheets. Any low-return investment can ultimately only turn into debt and overcapacity. Whether it is bank or local government debt, the credit behind it is ultimately that of the central government. But in a narrow sense, fiscal space is relatively limited (the broad deficit ratio is already as high as above 8%, and the overall leverage ratio of the non-financial sector is 308%, even higher than 251% in the United States).

The "15th Five-Year Plan" devotes a dedicated chapter to "accelerating high-level technological self-reliance and self-strengthening, and leading the development of new quality productive forces," with independent controllability and technological innovation as the two key medium- and long-term levers. With clear strategic goals, controllable short-term pressure, and not-so-large fiscal space, we expect that a major shift in fiscal intensity and policy direction may still require some time or greater pressure.

Source: FactSet, CICC Research Department

Chart: China's capacity utilization rates for computers, communications and electronics, and general equipment manufacturing are at relatively high percentiles

Source: Wind, CICC Research Department

This article is sourced from: CICC Insight. Risk disclosure and disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account 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 particular circumstances. Investment based on this is at the user's own risk.