Commentary: Why AI Stocks Are Defying Gravity — And What Could Bring Them Down
Since the release of ChatGPT in late 2022, the stock prices of leading U.S. artificial intelligence companies — the so-called Magnificent Seven — have far outpaced the broader market. Following the emergence of DeepSeek in early 2025, China’s top AI firms, primarily listed in Hong Kong, also posted notable gains. In the U.S., although corporate earnings have grown rapidly, risk premiums remain extremely low, reflecting optimistic investor sentiment. These lofty valuations have sparked growing debate about whether AI-related assets are in bubble territory. Rather than attempting to technically define a bubble, this article explores the relationship between asset prices, innovation and the macroeconomy.
Cause and effect
One way to absorb high stock valuations is through falling interest rates, leading some investors to hope for rate cuts by the U.S. Federal Reserve. Traditionally, a seesaw relationship exists between interest rates and risk asset prices: lower rates tend to lift stocks. But this theory is challenged by the recent surge in equity prices despite rising U.S. interest rates.
There are three possible dynamics between interest rates and stock markets. First, the classic view: rates influence prices. Second, the reverse: stock performance influences interest rates. The AI-driven equity rally has supported aggregate demand, fueling inflationary pressure and prompting the Fed to keep rates elevated. As of this year, AI-related capital expenditure has contributed about one-third of U.S. GDP growth. Meanwhile, the wealthiest 10% of Americans — who hold 85% of U.S. equities — now account for half of all consumption, the highest level on record. Rising stock valuations boost consumption and reduce savings, while AI capital investment boosts demand, together pushing up the natural rate of interest.
Third, both interest rates and stock prices may be driven by a third force: global capital inflows. U.S. Treasury data showed that as of September 2025, foreign investors held $21.2 trillion in U.S. stocks — 31.3% of the total market. This is the highest level since World War II. International capital bets on U.S. tech, reinforcing stock gains and supporting demand, which in turn supports higher rates.
The second and third dynamics suggest that high valuations and high interest rates can coexist. What matters going forward is whether a drop in interest rates stems from weakening investor confidence. If enthusiasm around AI fades and capital spending slows, it could weaken demand and bring down the natural interest rate. In such a case, falling rates would reflect market pessimism — not provide relief. Digital platforms and generative AI are allowing more retail investors to participate in the AI investment narrative, which could amplify both upside momentum and downside risk.
Another way to absorb valuations is through earnings growth. The concern is that gains have been concentrated in a few mega-cap names. From an optimistic view, AI is a general-purpose technology, like electricity or the steam engine, that will eventually impact all sectors. But today’s investors must consider the cost-benefit equation of AI and assess the efficiency and output of these applications.
Cost and benefit
The current AI boom features immature application technology but high expectations. This requires ample support from capital markets, especially equity. The first consideration is R&D costs, especially for large models. These costs include computing power, personnel, electricity and data-related expenses. Application costs, notably power consumption for inference, also matter.
A notable trend is the shift from capital-light software models to capital-intensive hardware and infrastructure, led by tech giants. These firms now back major AI startups, taking the place of traditional venture capital. Large companies can better harness AI to reduce uncertainty, boosting expected returns.
Still, returns from large models are uncertain. Effectiveness varies across use cases, and it’s hard to aggregate cost savings, efficiency gains or indirect benefits. Language models often provide inaccurate answers, limiting commercial utility. Their value lies in being fine-tuned on company data, but most firms are not yet prepared.
Looking long-term, research offers both optimistic and conservative projections. Nobel laureate Philippe Aghion has reviewed studies using two methods. Extrapolations from past tech revolutions suggest AI could add 0.8 to 1.3 percentage points to annual GDP growth over the next decade. Task-based models, like those used by Daron Acemoglu, estimate just 0.07 percentage points. Aghion believes the latter may underestimate AI’s impact, given cost declines and improving capabilities. A comprehensive review suggests AI could boost productivity growth by 0.08 to 1.24 percentage points annually.
A 2024 research report published by CICC Global Institute estimates AI will add 9.8% to China’s GDP by 2035, or about 0.8% annually. The challenge for companies in the near term is bridging the gap between today’s massive costs and uncertain future profits. This has led to worries of overheating. Economic analysis, particularly of scale economies, can help assess these risks.
Economies and diseconomies of scale
DeepSeek’s breakthrough involved algorithmic upgrades to compensate for China’s lack of advanced chips due to U.S. export restrictions. In essence, it simulated 4-nanometer chip performance using 7-nanometer chips. Pessimists feared this would undercut chipmakers like Nvidia. But the historical Jevons Paradox offers a different view: better efficiency often leads to more demand, not less.
The original Jevons Paradox described how James Watt’s improved steam engine led to higher coal consumption overall. Despite more efficient usage per unit, total usage soared as more industries adopted steam power.
Following DeepSeek, Chinese AI firms continued making algorithmic and system gains. Does the Jevons Paradox still hold? Chips exhibit economies of scale: unit costs fall as output grows. Coal, a natural resource, displays diseconomies of scale — rising costs with greater demand. This difference matters. With increased demand, chipmakers may benefit from volume, but prices fall. Coal prices rise, benefiting miners both from quantity and price.
While chipmaking is not perfectly competitive, major players may reap monopoly profits for a time. Still, the long-term trend reflects falling costs and prices.
This logic extends to large AI models. Inputs like compute and electricity face diminishing returns, meaning improvements require more input. Yet strong demand supports upstream investment. Diminishing returns alongside rising profits indicate pricing power. Scale barriers benefit tech giants. But how long will these monopoly profits last? Antitrust regulation remains a question, as does geopolitical rivalry.
Open-source models offer an alternative. China’s open-source efforts have shifted global dynamics, forcing rivals like OpenAI and Meta to adapt. DeepSeek’s sparse architecture has been adopted in IEEE standards. Open-source tools, with lower inference costs, challenge closed systems. They empower developers worldwide, including in emerging markets, fostering broader innovation.
Energy consumption is another constraint. Power comes from fossil fuels and clean energy. China leads in green energy, while the U.S. supports fossil fuel use. Fossil fuels face diseconomies of scale; green energy benefits from manufacturing economies of scale. This divergence could matter more if large models become ubiquitous.
Creative destruction
In summary, today’s lofty AI stock prices rest on two pillars: long-term earnings optimism and a current capital spending surge. Both may prove unsustainable.
Valuations could tumble if chip economies of scale erode barriers or if AI applications fall short of expectations. Key questions include whether algorithmic progress can offset diminishing returns, and whether AI delivers increasing returns at scale. These questions will take time to resolve.
Unlike real estate bubbles, tech bubbles can be productive. While painful in the short term, they often accelerate innovation. Overinvestment, if excessive, can still drive long-term economic benefits through scale and positive spillovers. By contrast, real estate — with its diseconomies and negative externalities — poses broader systemic risks when it collapses.
caixinglobal.com is the English-language online news portal of Chinese financial and business news media group Caixin. Global Neighbours is authorized to reprint this article.
Image: Nattawat – stock.adobe.com