AI should not become next asset bubble

Whether the artificial intelligence sector in the United States has become a bubble is stirring intense debate, with sharply divided opinions. Pessimists warn that AI is inflating into an unprecedented mega-bubble, potentially 17 times larger than the dot-com bubble of 2000 and four times bigger than the 2008 subprime mortgage crisis. But optimists argue that AI technology is redefining economic paradigms and could drive advances that eclipse those of the past century over the next decade.

What truly warrants attention is not whether an AI bubble exists, but whether AI has become excessively financialized. An objective assessment of this financialization is more important.

History shows that nearly every cutting-edge technology goes through a phase of capital frenzy — the "bubble" stage — in its early development. This is known as the Perez industrial cycle. A certain degree of exuberance in the initial phase of breakthrough technologies fits the pattern of technological industrialization and is not necessarily harmful.

The challenge is to determine whether the bubble is appropriately sized. AI is not immune to the Perez cycle, and some bubble is inevitable. Globally, there is consensus that AI is advancing from large language models to agents and embodied intelligence. But opinions are divided on the pace of this development, especially in its transformation into real-world industrial applications, where both depth and speed vary widely. Widespread adoption can absorb investment and mitigate the bubble, but a shortfall in real-world applications may depress returns on investment and lead to overheated capital influx.

In both China and the US, it is difficult to determine whether the computing power market is overheating. It is a seller's market driven by genuine demand. High-end training computing power is in short supply, and infrastructure-level resources such as graphics processing units and data centers are its hard currency. The significant supply-demand gap suggests that investment in computing power is addressing genuine demand rather than creating the overcapacity that results in a "bubble burst".

This boom is fueled not only by current shortages but also by expectations of future applications. For instance, Huawei estimates that nearly 900 billion intelligent agents could exist worldwide by 2035, implying an enormous demand for computing power as foundational infrastructure.

But the absence of a bubble in computing power does not mean there is no irrational exuberance in other areas of AI. Some AI application firms — such as those offering apps or wrapper models — are massively overvalued. The term "price-to-dream ratio" has been coined to describe these loss-making companies with exorbitant market capitalizations. The rapid rise of AI-related stocks in the US, Japan and the Republic of Korea calls for close scrutiny.

The deeper concern is whether AI has become over-financialized. This occurs when resources such as computing power, models, data, and equity in AI projects begin to function like financial assets. They acquire financial attributes such as price signals, tradability, speculative potential and leverage.

A moderate financialization of AI helps in raising capital and fostering technological progress. But excessive and rapid financialization inflates virtual value and diverts capital toward speculation instead of research and development. In such a scenario, participants focus less on the practical value of AI and more on the price the next buyer is willing to pay. This creates a self-reinforcing cycle where rising prices attract more capital, leading to even higher prices, ultimately resulting in a market with no effective brakes.

Without strong financial attributes, a bubble in the AI industry is less concerning because markets have self-correcting mechanisms. Even if the bubble bursts, the infrastructure it leaves behind, such as massive computing power centers, can serve as low-cost public resources for society's digital transformation. After the dot-com bubble burst in 2000, the optical fiber networks, internet protocols, skilled engineers and online consumer habits became the foundation for the subsequent growth of the digital economy.

In contrast, strong financial attributes can inflate bubbles and potentially trigger financial crises. The real estate sector is a cautionary example. When exchange value overshadows user value, it can lead to serious financial issues. The turmoil surrounding Fannie Mae and Freddie Mac during the 2008 global financial crisis highlighted the dangers of such imbalances.

Today, signs of excessive financialization are emerging in the AI sector in the US. Technological resources are increasingly being treated as tradable, speculative, and collateral assets, raising the risk of their decoupling from technological R&D and real-world applications.

China's AI sector is not immune from such risks. Some channels for transmitting these financial attributes from the US to China are already in place. As China expands financial support for AI development, it must also remain alert to risks.

While financial backing is important, China should proactively prevent AI from acquiring strong financial attributes, and guide the short-to-medium-term financing climate, valuation logic, and long-term development direction to ensure that financial support remains on the right track. The strategy for AI development should be "strong financial support but weak financial attributes".

If that baseline is maintained, a certain level of investment frenzy or minor bubbles can be managed. The value of an AI product should be determined by the genuine services or applications it provides, not by the expectations of a future buyer.

High-end computing power with genuine demand is strategic infrastructure for the future. Financial instruments such as data center REITs and intellectual property-backed financing are based on real assets and tangible demand, with their financial attributes closely tied to use value. Their development should continue in a sound and proper way.

At the same time, AI should be protected from the speculative dynamics that plagued the real estate sector. Four guardrails are particularly important. First, establish a rigorous verification mechanism to ensure the authenticity of AI-related assets and prevent fictitious computing power and virtual data from entering financial markets. Second, place a limit on the core AI resources held by any single capital entity to prevent monopolistic speculation. Third, regularly stress test AI-related financial assets to identify potential bubbles and mitigate risks proactively. Fourth, ensure all financial products are sold to suitable investors to maintain market confidence while respecting inherent risks.

AI's ultimate value lies in generating real productivity. Finance should return to its core role of serving the real economy, while AI should stay true to its essence of technological innovation.

The author is the chairman of the Chinese Society for Sustainable Development. The views do not necessarily reflect those of China Daily.