Search the site
Warnings of an AI stock market bubble abound. Should investors and policymakers be concerned? This column argues that innovation-driven bubbles can yield real benefits when financed by equity rather than debt. Even if investors face large losses, policymakers have little reason for concern, unless it drives risk appetite elsewhere in the system and AI investment relies on bank credit.
The stock market has seen an extended bull market since 2009, with the S&P 500 up 63% this decade despite the damaging Covid shock. The primary driver of stock prices today is investment in artificial intelligence (AI). Is the market in a bubble – and should we be concerned?
While a bubble might concern only those mis-valuing their investments (Albori et al. 2025), it can become a broader issue for two reasons: first, if it affects risk appetite elsewhere in the system, encouraging uneconomic investments; and second, if it is fuelled by credit, especially bank credit, providing a direct causal chain between the bubble and traditional channels of financial fragility. So policymakers face a dilemma: they could risk killing innovation, or let the boom run and risk a harsher correction later.
A cautionary example is China, which displayed most of the conventional signs of a bubble in the 2000s. If its policymakers had burst the bubble, China would not be where it is today. And even if the financial authorities would want to deflate bubbles, the political leadership would probably disagree. For example, even if the Fed concluded AI was in a bubble and wanted to deflate it, the current US administration would probably not allow it to do so.
While many factors cause bubbles, momentum plays a central role. Investors see prices going up, which makes them buy. And because they buy, prices go up further, all in a happy feedback loop. Even those suspecting a bubble may play along and invest, hoping to get out in time.
The way we perceive risk is also strongly linked to bubble dynamics. As momentum drives prices up, volatility declines mechanically, since steadily rising prices appear stable. Yet such fair-weather measures of risk miss the accumulation of hidden fragility. The true risk – the probability of collapse – is hidden and endogenous (Danielsson et al. 2009). It rises quietly during the calm phase while measured volatility falls. The risk we measure is usually not the risk we care about; measured risk only shoots up after a crisis hits, when it is too late to trade out.
This implies that volatility, the way most investors and policymakers identify excessive risk, misleads us. And so will all the other risk measures, such as credit spreads, expected shortfall, and most measures of systemic risk. They tell us what we want to hear while concealing the real danger.
The risk to society is that when a bubble collapses, sharply reduced wealth not only leads to reduced spending by the entities most exposed, it also suppress risk appetite through the system, leading to sharp drops in economic activity and employment. When bubbles are financed by credit, and especially bank credit, the consequence can also be a systemic event.
There are plenty of warnings that the market might have overheated, such as those from the IMF and the Bank of England, and from Jamie Dimon in the private sector. Joseph P. Kennedy, the father of President John F. Kennedy, suggested a way to identify bubbles: "When the shoeshine boys are giving stock tips, it is time to get out of the market."
But a more fundamental approach asks whether the assets behind the bubble have intrinsic value or if, like bitcoin or stamps, their value is solely what a future investor might pay. We can look at the funding assumptions for all the server farms being built. Jason Furman (reported in Financial Times 2025) estimates that almost all the economic growth in the US this year has arisen from investment in IT and related sectors.
History shows that periods of major innovation – railways in the 1840s, electricity in the 1890s, the internet in the late 1990s – often coincide with speculative booms generating spectacular profits and losses. This is good for society and fantastic for a few investors, even though most investors will suffer large losses. In 1866, the largest bank in the world caused the 19th century’s biggest financial crisis by betting heavily on the hottest technology of its day – shipping – and losing.
Consider the dot-com bubble in the late 1990s. The best example of dot-com-era excess was Yahoo!, which had a market capitalisation of 128ドル billion on 3 January 2000. At that time, it had about 180–200 million users worldwide, suggesting that the market valued each user at about 680ドル. Back then, Yahoo! was the AI of its era – a must-own name with a price implying that it would dominate the digital future. But it didn’t. One dollar invested in Yahoo! at the height of the dot-com boom would now be worth about 0ドル.04.
At the same time, Amazon reached a peak valuation of 39ドル billion. Anyone buying then would have seen the share price collapse by over 90% within two years. Yet those investors who bought at the top and held onto the stock until today would have seen their position rise almost fifty-fold – an annualised return of about 15% over 25 years. Even a purchase at the top of the bubble would have been remarkably profitable in the very long run.
The details change, but the pattern of extrapolating early success into perpetual growth is familiar. Every generation seems to find its Yahoo!s and Amazons – the firms everyone assumes will own the future.
CUDA made Nvidia what it is, and investors value Nvidia as if it will continue to increase its dominance. But its price-earnings ratio is about 53. If CUDA loses relevance, then in 2050 Nvidia will not be what it is today. Similarly, the half a trillion valuation of OpenAI is based on its being the company that will end up owning AI – Amazon, not Yahoo!.
At the peak of the dot-com bubble, many companies were priced on similar expectations, and most failed. Yahoo! failed because Google came up with better search technology; a couple of people with a clever idea was all it took.
If AI is a bubble, it does not appear to be a dangerous one. On the contrary. We should all appreciate all the investors providing the money to fuel the AI boom that will make our lives so much more comfortable in the years to come.
There does not seem to be an immediate cause for public concern, but that could change for two reasons. First, if AI investment drives risk appetite elsewhere in the system, it can lead to Minsky-type investment boom-to-bust cycles in GDP and crises, as we document empirically in Danielsson et al. (2018, 2022).
Second, as long as the funding for AI data centres comes from equity investors, as did the investment in the dot-com bubble, there is little reason for social concern. But although that was initially the case, the situation is changing. AI firms are increasingly relying on debt-based and circular financing structures, as reported recently by the Financial Times (2025b) and Bloomberg (2025). Circular funding loops of this nature imply that the same capital appears several times on different balance sheets. The consequence is hidden fragility, misleading investors, creditors and regulators.
The authorities face a difficult problem. While their primary obligation is not to suppress bubbles, they need to identify excesses in the financial system and act before it becomes too late, which they did not do in the years before the Global Financial Crisis in 2008. They also need to ensure that if a bubble bursts, the rest of the system will not be at risk.
Actual danger emerges when bank lending joins the funding cycle. At that point, the AI bubble ceases to be a matter for investors alone and becomes a genuine policy concern.
Albori, M, V Nispi Landi and M Taboga (2025), "Unpacking US tech valuations: An agnostic assessment," VoxEU.org, 8 September.
Bloomberg News (2025), "OpenAI’s Nvidia, AMD Deals Boost 1ドル Trillion AI Boom With Circular Deals", 14 October.
Danielsson, J, H S Shin and J-P Zigrand (2009), "Modelling financial turmoil through endogenous risk", VoxEU.org, 1 March.
Danielsson, J, M Valenzuela and I Zer (2018), "Low risk as a predictor of financial crises", VoxEU.org, 26 March.
Danielsson, J, M Valenzuela and I Zer (2022), "How global risk perceptions affect economic growth", VoxEU.org, 13 January.
Financial Times (2025a), "Does GDP growth minus AI capex equal zero?", 1 October.
Financial Times (2025b), "Measuring risk in the AI financing boom: A shift towards debt raises the potential fallout from the data centre spending spree", 13 October.