Reflections on the AI Investment Bubble Debate

Since the launch of ChatGPT in late 2022, the stock prices of leading US AI companies have significantly outperformed the broader market. Following the emergence of DeepSeek in early 2025, leading Chinese AI companies (primarily listed in Hong Kong) have also substantially outperformed their respective markets. In the context of the US stock market, despite rapid earnings growth from related companies, risk premiums are at extremely low levels, reflecting investors' optimistic expectations. Elevated stock valuations have recently spurred increased discussion about a potential bubble in AI-related asset prices. This essay is not a technical analysis on how to define or measure a bubble, but rather offers some reflections based on asset prices, innovation, and the relationship with the macroeconomy.

I. Cause and Effect

One possibility for digesting high stock valuations is falling interest rates, leading some investors to pin their optimistic expectations on Fed rate cuts. Conventional thinking suggests that interest rates and risk asset prices have a seesaw relationship; a decline in the risk-free rate prompts a shift in allocation towards risk assets, benefiting stock valuations. To assess whether this relationship still holds today, we must first explain why stock prices rose sharply in an environment of rising US dollar interest rates over the past few years.

The relationship between interest rates and the stock market has three potential dynamics. First is the traditional view where interest rates are the cause and stock prices are the effect. The second reverses this: the stock market is the cause, and interest rates are the effect. The AI-led stock market rally has been a key support for US aggregate demand, with the resulting inflationary pressures prompting the Fed to maintain interest rates at high levels. Since the beginning of this year, AI-related capital expenditure has contributed one-third of US GDP growth. Simultaneously, the wealthiest 10% of the population owns 85% of US stocks and accounts for half of total consumer spending, the highest share since records began. Under this logic, the wealth effect from the stock market boosts consumption, reduces savings, while AI-related capex increases investment. Together, these forces lead to a rise in the natural rate of interest required to balance supply and demand.

The third possibility is that both interest rates and stock prices are effects, driven by a third force. The US AI-led stock market has attracted capital from around the world. According to US Treasury TIC data on foreign investor holdings, as of September 2025, foreign investors held $21.2 trillion of US stocks, representing 31.3% of the total market capitalization. Historically, this ~30% market share is the highest since the end of World War II. Investors worldwide are betting on US tech giants' stocks, fueling the US market. The related wealth effect stimulates aggregate demand, thereby supporting the level of interest rates.

Looking ahead, regarding the interest rate-stock market relationship, the key is to distinguish the cause of any potential rate decline. If investor optimism about AI wanes or is disproven, stock prices and AI-related capex fall, leading to weaker aggregate demand, a lower natural rate, and prompting Fed rate cuts. In this case, stock prices are the cause and rates are the effect; lower rates would not support stock prices. Particular attention should be paid to the herding effect among investors in the era of digital economy and AI.

Another possibility for digesting high valuations is earnings growth from the relevant listed companies. An intuitive concern here is that price increases have so far been concentrated in a few stocks. From an optimistic perspective, AI is a General-Purpose Technology (GPT), similar to steam and electricity in the past, which will gradually permeate all industries. This takes time, with AI technology R&D and upstream players benefiting first.

II. Cost and Benefit

A characteristic of this round of AI technological development is the coexistence of low application-layer technological maturity and high expected profits, necessitating support from capital markets, particularly equity investment. Investors must ultimately consider the cost-benefit ratio of AI. One constraining factor is the R&D investment required for innovation. Some costs are visible, such as the R&D and application costs for large models. R&D costs primarily include computing power, personnel, electricity, and other items, while inference energy consumption is a significant component of application costs.

Notable new features in AI-related investment models are emerging. The traditional capital-light software distribution model familiar to venture capital firms is shifting towards a capital-intensive hardware production model (advanced chips and data infrastructure). Leading this transition are dominant tech giants, who are not only investing huge sums but also becoming the primary backers of major AI startups like OpenAI and Anthropic, taking over the historical role of VCs.

Despite high hopes and expectations, how much potential does AI truly have in industrial applications? This is debated. Analyzing the revenue side of large models faces significant uncertainty. Different application scenarios vary greatly; the direct economic benefits of large models and indirect economic benefits are difficult to measure and aggregate. One drawback is that large language models can provide inaccurate answers, limiting their value in many business applications.

Taking a longer view, some macroeconomic research estimates AI's impact on economic growth, presenting both optimistic and pessimistic views. Synthesizing the literature, AI's contribution to annual productivity growth likely falls between 0.08 and 1.24 percentage points, with its midpoint comparable to extrapolations from the IT and electricity revolutions. CICC Research's 2024 report 'The Economics of AI' estimated that, by 2035, AI adoption will bring an additional ~9.8% to China's GDP, corresponding to an additional annualized growth rate of approximately 0.8%.

III. Economies and Diseconomies of Scale

DeepSeek's breakthrough lies in using algorithmic architecture improvements to compensate for computing power constraints, simply put, using 7nm chips to do the work of 4nm chips. Regarding the implications for leading semiconductor companies, two views emerged initially. The pessimistic view argued that algorithmic improvements substitute for advanced chips. The optimistic camp cited the historical Jevons Paradox, where technological progress that increases the efficiency of resource use leads to an income effect outweighing the substitution effect, resulting in increased overall demand.

Applying the coal logic to chips, we must ask: as factor inputs, how do the economic attributes of chips and coal differ? Chips, as manufactured goods, exhibit economies of scale, meaning unit costs decrease as production scale increases. Coal, as a natural resource endowment, exhibits diseconomies of scale; the mining supply is inelastic, so increased demand leads to price increases.

This logic can be extended from chips to large models. We can view large models as inputs that create value through applications. The scaling law implies diminishing returns to scale for factor inputs like compute, data, and electricity. The input threshold implied by the scaling law places giant tech firms in an advantageous position. The question is, how long can such monopoly profits last?

China's open-source model significantly impacts the global AI competitive landscape, pushing the market towards greater fairness. DeepSeek's dynamic sparse architecture was incorporated into an international standard by IEEE, and many startups in Europe and the US are adopting Chinese open-source models. The low-cost advantage of open-source attracts global developers, helps developing countries build localized AI applications, and creates a more balanced global AI innovation force.

IV. Creative Destruction

Synthesizing the above discussion, the current high valuations of US AI-related stocks may rest on two pillars. First, investors may be overly optimistic about long-term future earnings growth, causing stock prices to deviate significantly from current earnings. Second, current earnings themselves may be unsustainable.

A bubble burst could manifest through the effects of economies of scale in the chip industry, particularly combined with changes in the competitive landscape, such as the development of China's advanced semiconductor industry or improvements in algorithmic architectures that enhance chip output efficiency, leading to broken technological barriers and falling advanced chip prices.

Finally, it is worth noting that tech bubbles differ from real estate bubbles. The bursting of the former causes significant short-term disruption but, in the long run, constitutes creative destruction. Economies of scale and positive externalities mean that periods of overinvestment in specific sectors, while unsustainable, may ultimately benefit long-term, macro-level technological progress and innovation.