The Stock Market Is In a Bubble, When Will It Burst?

We are now ‘in’ an AI-centric market bubble, though not at the end of it in the sense of there being a ‘mania’ proper. In the short term, we might well see a wobble in AI stocks.

---

I’ve spent a bit of the summer re-reading the work of Charles Kindleberger, an important economist whose career intersected monetary systems and stock market bubbles, two issues that are top of mind for investors and economists today.

Kindleberger had an interesting career. His PhD advisor at Columbia was Henry Parker Willis (a key architect of the Federal Reserve Act in 1913 and the first Secretary of the Federal Reserve Board that became what we now know as the ‘Fed’). Then in one of his first jobs at the US Treasury, Kindleberger worked for Harry Dexter White, the interlocutor and ‘rival’ of JM Keynes at the Breton Woods conference.

To that end, Kindleberger had a very strong sense of the creation of the monetary infrastructure that has built today’s economic world, and he then had the opportunity to play a role in this as one of the architects of the Marshall Plan, which did so much to spur growth in post-war Europe and to cement the view of the USA as the benevolent world power. Benn Steil’s book The Marshall Plan: The Dawn of the Cold War is worth a read.

Beyond his policy work, Kindleberger is best known for ‘Mania’s, Panics and Crashes’, the best outline of how asset bubbles form and are followed by crashes.

It is highly pertinent today because a variety of stock market valuation indicators (the long-term ‘Shiller Price Earnings ratio’, the ‘Buffet Indicator’ as well as measures of market concentration – the largest ten stocks account for 75% of the entire market), point to the kind of market behaviour seen only in market bubbles (like 2001). Consistent with this, various investors, as well as entrepreneurs like OpenAI’s Sam Altman are warning of a ‘bubble’.

One test of the bubble thesis is to follow Kindleberger’s theory that asset price bubbles follow a common speculative cycle. According to Kindleberger, bubbles often start with an innovation – in a technology (i.e. railways) or a financial policy or market structure, or even a growth ‘miracle’ (note all the Tiger economies from Hong Kong to Ireland have seen boom/bust cycles) towards which investors channel capital, and then even more as asset prices rise and a narrative around the ‘mania’ begins to build.

This effect helps to loosen the strings of the overall economy and financial sector, but around this point asset prices are reaching incredulous levels as investor euphoria intensifies, drawing in further speculation, until prices then turn down, and the house of cards collapses in a crash. The collapse is always greater when households, institutions and individuals have borrowed on the back of high asset prices, and logically there is greater contagion across the economy.

One lesson from the Kindleberger book is that ‘new’ things – inventions or economic policy liberalizations often provide the spark for a new speculative bubble, that can grow and destabilise a financial system if enough speculative capital is driven towards it.

AI is a case in point. For context, Morgan Stanley estimates that over USD 3 trillion will be invested in AI related infrastructure (data centres, energy), with about half of that coming from the cashflow of the large technology companies, and the rest from private credit. It is clear that the large technology firms (from Microsoft to Nvidia) are the driving force behind this boom, especially so in the context of the fiscal weakness of most Western governments.

The recent results season was instructive in this respect. Often in a bubble, the Dot.com one being a good example, the earnings associated with the bubble are only ‘prospective’ or somehow inflated. This is not the case with the large technology firms so far – by and large they report very strong earnings, which helps to soften the bubble argument. The catch is the circularity of the capital expenditure by the large technology firms – META for example is spending aggressively on data centres and chips and running down its cash levels. To that extent, the large tech firms are making a bold bet to get ahead in the AI game, but it is a concentrated and possibly existential bet.

If the strong cashflow position of the firms at the centre of this bubble is unusual in the context of the Kindleberger framework, two other factors also stand out as untypical.

The first is that short- and long-term interest rates in the major economies are close to ‘neutral’, neither too hot not too cold. Bubbles are often characterised, or preceded by ‘easy money’, though it has to be recognised that the last decade has been one of near continual stimulus (from QE to fiscal spending).

The other unusual factor is that the stock bubble is occurring against a backdrop of intense geopolitical and economic policy uncertainty – from great power competition to an unravelling trade order to a reconfiguring of America’s role in the world. The one way in which these dislocations fit the bubble narrative is that AI is a strategic asset, a ‘must-have’, that leads to an environment where for example the US government aims to take a strategic stake in Intel.

Having written in early 2024 of a ‘Bubble Brewing’, I think we are now ‘in’ an AI centric market bubble, though not at the end of it in the sense of there being a ‘mania’ proper. In the short term, we might well see a wobble in AI stocks, before they pick up again.

In a future note I will detail my own experiences of the dot.com and European real estate bubbles, which lead me to think that ‘we are not there yet’ in terms of the evolution of this bubble. It will end with the absurd – Nvidia encroaching on a USD 10 trn valuation, food companies publicly adopting AI and seeing their values double, wild predictions that AI can treble productivity and wipe out the debt, and on the flip side, the structural risks to energy and jobs markets that AI could pose.

But, we are not there yet.

By Mike O'Sullivan, Senior Contributor

---

Lessons From the Dot-Com Era

One should be just as cautious about predicting the imminent burst of an AI bubble as skeptical of the exaggerated hype currently surrounding artificial intelligence.

The AI Bubble and the Dot-com Era

There are concerning signs. The “Magnificent Seven” stocks (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla) make up more than a third of the S&P 500, with recent growth driven by an AI story. Investors are becoming uneasy with this level of concentration. At the peak of the dot-com bubble in 2000, the top technology stocks from the late 1990s (Cisco, Dell, Intel, Lucent and Microsoft) accounted for 15% of the index. Such concentration heightens risk.

The parallels do not stop there. A massive telecommunications infrastructure buildout ushered in the e-commerce era. The world needed the internet pipes to enable high-speed connectivity. This triggered an overly optimistic deployment of fiber optic networks, which led to catastrophic bankruptcies when the demand did not materialize in the short term.

Today, the leading AI companies are investing hundreds of billions of dollars in new data centers. The total capital spending in this area is being discussed in the trillions of dollars, figures that were once only associated with large countries’ GDPs. Will history repeat itself, causing an imminent collapse? Meanwhile, the connectivity boom and investments from a quarter century ago enabled the always-on world we live in today. They created opportunities for value creation beyond infrastructure, at the application level, and drove the transformation of the information technology industry through the shift to the cloud. Some might argue that data centers are now the new utilities required to provide on-demand information services for an increasingly connected world.

Will the Demand for AI Materialize?

Much of the current attention is focused on the consumer space. OpenAI’s ChatGPT website received over five billion visits during July. But that is not the whole story.

The true economic impact will be measured by consumer and enterprise adoption. The National Bureau of Economic Research started publishing its survey of generative AI adoption about a year ago. As of late 2024, about 40% of the U.S. population reported using generative AI, and 23% reported having used it for work at least once in the week before they were polled. When comparing the level of adoption since the initial product launch, generative AI at work is taking off faster than the personal computers or the internet, the study concludes. This underscores the potential of AI as what economists call a general-purpose technology, one with deep and pervasive impact on the economy.

But challenges remain. A group of MIT researchers surveyed over 300 publicly disclosed AI initiatives, more than 50 companies and hundreds of senior leaders from January to June 2025 to conclude that 95% were not getting any return for their investment. They were also able to identify three elements that made the remaining 5% successful. Successful companies are buying instead of building, executing within business units as opposed to central laboratories and choosing tools that integrate with their existing business workflows. While achieving the returns associated with business transformation is rare, adoption is high, with 90% seriously exploring buying an AI solution. This is a familiar pattern in enterprise technology adoption. It has been captured by what consultants call the hype cycle, tracking innovative technologies from their market entrance to when businesses are likely to benefit from them, and the technology has become mainstream.

Bank of America, the second-largest bank in the U.S., with a $4 billion budget for new technologies such as AI, is an example of the pattern identified in the MIT study: integrating AI and business workflows. The bank developed a tool to help bankers prepare for client meetings, retrieving information from multiple systems. Previously, a junior banker would have executed this process over multiple hours or days.

How Far Can The Current AI Models Take Us?

As AI usage increases, so does the debate about its ultimate potential and whether the current development model is sustainable.

Much of the progress to date has been made on the back of large language models that benefit from scale. Scale means that with more computing power and more data, one produces better outcomes. Richard Sutton, an AI pioneer, observed in 2019 that general methods leveraging computational power outperform those that rely on human ingenuity and complex heuristics (in what he coined “The Bitter Lesson” for humanity). He has recently criticized the industry’s fixation on scaling and called for a correction towards agents that learn continuously.

Gary Marcus, one of the most vocal critics of the artificial intelligence hype, commented on the mixed reviews received by OpenAI’s latest ChatGPT-5 release. He echoed the sentiment that a development model predicated on scaling is not the path forward, a position he has been sponsoring for decades.

These scientists’ deep skepticism about the current progress represents a technical word of caution. The hype conditions created by investors and large AI laboratories can lead to disappointment. Both, however, are believers in AI’s ultimate potential, while suggesting that alternative approaches are necessary. These may necessitate more, instead of less, investment in research and development.

Is There An AI Bubble?

One should pause when even Sam Altman, who helped spark the AI boom, warns that the market may be overheating. He and other investors mention soaring valuations, too much money chasing unproven business models, and the risk of building infrastructure faster than demand will justify. Like in the MIT report, they worry that much of the capital outlays are flowing into projects unlikely to deliver results soon. The concern is less about AI’s long-term promise and more about inflated expectations setting the stage for a sharp correction.

Binary thinking that swings between hype and the fear of an AI bubble may limit more nuanced analysis. AI’s long-term potential remains significant, but markets rarely move in straight lines. A correction could slow momentum in the short term while reinforcing the need for discipline. The next phase will depend on advancing research, improving model quality, and directing enterprise investments toward measurable economic value.

Paulo Carvão is a Senior Fellow at Harvard