A trillion-dollar game of corporate chicken: what happens when the AI bubble bursts?
And who is already paying the price for investment mania?
The last few weeks have seen a growing number of mainstream commentators and major financial institutions warning about the multiplying signs that stock markets are overheating on the hype around AI investments. The FT is one of the media outlets sounding the alarm about what happens when the AI bubble pops. Commentators such as Gillian Tett have called out the emergence of a “cargo cult” around expectations that AI will deliver unprecedented gains in productivity. Meanwhile, the FT’s journalists have begun to investigate how the corporations with the most at stake in the competition to secure trillions of dollars in investment, such as OpenAI, are starting to rely on complex financial engineering schemes which direct flows of capital in a loop between companies working together at different stages in the AI value chain.
As Grace Blakely explains in a detailed post here, this ‘circular financing’ is inflating a bubble of expectations of profitability which, sooner or later, will burst:
“Imagine if a car manufacturer invested in a taxi firm, which used the money to buy only that car manufacturer’s vehicles. The carmaker could claim rising sales, the taxi firm could say it was growing, and investors might conclude the transport sector was booming. Everyone would start to throw money at these firms for fear of missing out on the boom, creating a self-fulfilling cycle of rising valuations. But there’s no one at the end of the chain buying the product – taxi rides. Instead, much of the activity is being financed by the carmaker’s own money. Unless real customers start paying lots of money for taxi rides, the flow of money between these companies will eventually reach an abrupt halt.”
How does this phenomenon relate to the structure of the AI industry I outlined a couple of weeks ago? In that post I argued that the race to produce ‘foundation models’ dependent on hyperscale computation, gigantic datasets, AI models with trillions of parameters and eye-wateringly expensive engineering talent has dominated the last decade of the industry’s development. The ‘foundation model’ paradigm is a story of the commercialisation of what was previously an academic field. This can be seen in the graph below from Our World in Data.
That story has been defined by success in attracting record levels of investment throughout the whole stack of technologies which is used to produce and deploy foundation models. This has included development of specialist semiconductors for data centres, to training dataset production, to technical innovations in model architecture such as the Transformer designed by Google engineers in 2017. One of the causes of the frantic levels of capital expenditure by the major companies in the US AI industry is the fact that they need to hedge their bets over which of these areas will deliver a decisive advantage.
This could easily come about through denying competitors access to costly equipment or scarce power supplies, or forcing them to overextend themselves with purchases they can’t sustain, rather than actually improving the performance of AI products.
In other words, the largest players among the US tech giants (such as Amazon, Google, Meta, xAI, OpenAI and Microsoft) are engaged in a game of corporate chicken with trillions of dollars of investment at stake.
Many analysts have looked back to previous technology-fuelled financial bubbles as a possible guide for what might happen next. Last year, Reuters columnist Edward Chancellor argued that the railway boom and bust in Victorian Britain is replete with lessons about today’s AI bubble. Capital surged into railway construction only to collapse as overproduction became obvious.
“It turned out that the revenue projections provided by so-called “traffic takers” were wildly overoptimistic. Railway engineers underestimated costs. The vogue for constructing direct lines between large urban centres proved mistaken, as most traffic turned out to be local. As a result, Britain’s rail network was plagued with overcapacity. There were three separate lines connecting Liverpool with Leeds and London with Peterborough. Railway companies cut dividends as returns on invested capital fell to 3%. By the end of the decade, the index of railway stocks was down 65% from its 1845 peak.”
The fall-out from the dot.com boom at the turn of the millennium and the crash which followed it is also an important comparator. Some of the big companies jostling to stay ahead in the tech investment arms race weathered that storm too. Grace Blakely argues that they are unlikely to go under when the deluge hits this time around.
“The AI companies, meanwhile, will be betting on the fact that they’re big enough to avoid going under regardless of which way the way the market goes. Amazon, Google, Microsoft, and the other tech companies that have come to define our generation survived the dot-com bubble. They’ll all survive this one too – joined by the new hegemons, like OpenAI and Anthropic. So, who is going to be left picking up the pieces?”
It is true that size is often a protective factor, particularly if your company is simply “too big to fail” and can therefore turn to the government to help with a bail-out. However, a more detailed look at the history of the companies which dominated the overinflated tech stocks twenty-five years ago shows the picture is more complicated than this.
Microsoft was badly damaged by the crash, taking around a decade and a half to recover the value it lost when the bubble burst. The company’s pivot towards building cloud infrastructure played a major role in its recovery, paving the way for the AI boom as it was Microsoft’s partnership with OpenAI which allowed Sam Altman’s company to launch ChatGPT in 2022. So yes, Microsoft made it through to the next upward sweep of the rollercoaster ride after 2000 and might do it again.
But the fate of Sun Microsystems offers a cautionary tale. Sun’s business model was more like Nvidia’s in that it concentrated on very expensive hardware for servers, rather than offering products in both home computing and enterprise IT services like Microsoft at the time. Thor Johannsson, a former Sun technical consultant put it like this a few years ago:
“They were making really cool servers, which cost more than cars and some even cost more than houses with nice swimming pools. Not exactly your home PC variant.”
At the height of the stock market bubble this propelled Sun’s valuation to then unheard-of levels as companies scrambled to create internet infrastructure. But this business model also left the company much more overexposed than Microsoft as the oversupply of servers and cables relative to demand became clear. Microsoft then took over large parts of Sun’s business. As Johannsson explains, “the wealth that the home computers had generated, paid for the same companies to take over server rooms.” The company limped on for a few years after 2000 before being bought by Oracle for a fraction of its former value.
Another significant victim of the dot.com crash was Cisco. The company had a value of $15 billion in 1995 but this had skyrocketed to $550 billion five years later as investors bet on soaring demand for network routers and switches. In the crash it lost 80 percent of its value and its stock price has never fully recovered.
Ironically, Google – which was still a small(ish) start-up in 2000 – owed its existence not to the foresight of executives at the companies which survived the dot.com crash, but to Andy Bechtolsheim, co-founder of Sun and later a vice-president at Cisco. In 1998 Bechtolsheim took a risky bet as angel investor on a then-untested technology – the PageRank algorithm which Larry Page had developed at the heart of his fledgling search engine business with partner Sergey Brin. The $100,000 cheque he wrote allowed Page and Brin to move out of their college rooms into a garage at Menlo Park and start hiring staff.
The early years of Google also underline how much more is at stake in this round of hype, overproduction and speculation. By 2001, just three years after its formation, the company was reporting profits, had hired Eric Schmidt (another former Sun executive) and had laid the technical basis for the advertising business which would power its rise to monopoly status in the search engine market. OpenAI, by contrast, doesn’t expect to be profitable until 2029, a decade after transitioning from being a non-profit. In the meantime, the company has taken on hundreds of billions of dollars of obligations.
Marxist economist Michael Roberts points out that the AI investment bubble isn’t just a risk for OpenAI but for the US economy as a whole. Without AI-related investments the US would either be in recession or close to it. Moreover, as a recent report from McKinsey shows, AI is reshaping global investment flows, as foreign direct investment (FDI) has surged into the US.
And there are now significant knock-on effects from the AI infrastructure boom in areas such as energy systems which also require lots of expensive, specialist machinery and therefore often “lock in” investment for long periods of commitment to particular types of technology. A recent report in the FT highlights how the AI data centre boom is driving up demand for gas turbines. This is another market dominated by a small number of large companies (Mitsubishi Heavy Industries and Siemens Energy are two key ones). Customers in the US now face 3-year waiting lists for gas turbines, while those in Asia will have to wait at least 5. This could propel customers towards accelerating the adoption of renewables, but it could also lead towards the expansion of coal power to fill the gap. China’s domestic energy market is dominated by a mixture of coal and renewables, and China accounted for 93 percent of new global coal construction in 2024.
The climate costs of the intense concentration and centralisation of capital reflected in the current AI investment mania underline that it won’t just be the tech bros who pay the price when this game of corporate chicken ends in disaster.