The artificial intelligence (AI) bubble is receiving a growing amount of attention. The Bank for International Settlements (BIS), in its December quarterly magazine, offers both reassurance and caution. It appreciates the strong earnings
of the sector, which, in reality, presented mixed results in the third quarter, with a few business groups advancing and others treading water, while one of the frontrunners, OpenAI, forecasts losses until 2030. It was Nvidia, with its strong profits, that revived the sector's euphoria.
After three years of acceleration, which raised the weight of the Magnificent Seven
from 20% to 35% on Wall Street, the BIS sees signs of a retrenchment
due to wariness about stretched valuations
and episodes of volatility
. It considers the optimistic expectations
to be well-founded and, in this respect, the AI trend – which the bank never refers to as a bubble
– is different from the dot-com bubble of the late 1990s, which was largely fuelled by over-optimistic expectations that were not underpinned by realised earnings growth
. The BIS notes that, despite the turbulence
in tariffs, high market volatility, renewed tensions between the US and China, the bankruptcies of First Brands and Tricolor, and fraud uncovered at two regional banks, there have been no contagion effects.
The BIS further certifies the good health of AI by comparing the Magnificent Seven
with China's Terrific Ten
, which experienced a sharp rally through 2020 and early 2021, fuelled by strong earnings momentum and the pandemic-driven surge in digital adoption
, but then suffered a prolonged correction due to regulatory tightening, weak domestic demand, and a shift in global investor appetite away from Chinese assets
.
Season of bubbles
The debate on the bubble
covers a wide range of positions: from those who deny its existence, to those who think it is a potential risk, to those who counter the doubters with future steps towards artificial superintelligence; to those who believe that the economy is teeming with processes that are growing unsustainably
. For Brad Stone of Bloomberg, bubbles are pretty much everywhere you look. [...] There may be a bubble in gold, whose price has soared almost 64% in the year [to December 12th], and one in government debt, according to Børge Brende, chief executive officer of the World Economic Forum, who recently observed that nations collectively have not operated this deeply in the red since World War II
. There are deflated bubbles such as that of Bitcoin, an early cryptocurrency, and that of commemorative memecoins ($170 billion at their peak in January), driven up and then down by Donald Trump and Melania Trump's junk coins, which lost respectively 88% and 99% of their value
in a year.
The distinction between speculative fads that promise fool's gold
and bubbles generated by large secular investments (railways, electrification, major public works), which require the centralisation of myriad small capital investments, is true, but only up to a point. The centralisation of monetary capital, so that it can be transformed into productive capital, takes place through the issuance of interest-bearing securities (shares and bonds), through the credit system, and sometimes through taxation. Stock market bubbles, real estate bubbles, and technology bubbles differ in terms of the intensity of capital absorbed, but none can do without productive interest-bearing capital and the plethora of rentiers that it brings with it. In these processes, investors and fools, rational
and irrational
euphoria mix and merge.
JPMorgan's vision
A report published in December by Jacob Manoukian, head of US investment strategy at JPMorgan, takes us to the upper echelons of America's leading financial group. The executive asks the question: Is AI a bubble?
. For the banker, the question is practical: on one hand, to feel comfortable allocating capital to stocks, we must feel confident that we are not about to see a bubble burst
. On the other hand, the bubble begins to form in part because credit is widely available
. Herein lies the banker's dilemma: he does not like explosive
bubbles, but his credit can also be explosive; credit increases leverage and creates a disconnect between economic fundamentals and market valuations. More and more investors join the crowd – until fundamentals finally prevail and the bubble bursts
. It follows that timing matters
.
First, the profitability and productivity of the investment must be considered. Manoukian cites the railway boom of the mid-19th century and the Internet boom of the late 1990s. Between 1843-1853, the number of kilometres of railway built in the United Kingdom quadrupled, but railway revenues per mile remained flat. By mid-2001, telecommunications companies had installed 39 million km of fibre optic cable, but only one-tenth of the fibre was activated and only one-tenth of the wavelengths of the active fibre was being used. Clearly, in both cases there was a large gap between the social utility of the investment and the profit earned: a gap called excess capacity
– capacity that was not justified by concurrent consumer demand
. In the AI cycle, Manoukian does not see any excess capacity today: three-quarters of the data centres under construction are pre-leased, and in all sectors (IT, energy, data centres), components are scarce relative to demand.
Secondly, the bubble expands when the cost of capital decreases. Manoukian cites four examples: the tulip mania in 17th century Holland, fuelled by Amsterdam's deep capital markets; the Japanese speculative bubble of the 1980s, inflated by bank loans secured by artificially high stock values; the US property bubble of the early 2000s, fuelled by subprime mortgages spread through an interconnected shadow banking system; and the shale energy bubble of the 2010s, boosted by zero interest rates after the financial crisis.
Manoukian argues that today the decline in the cost of capital consists of three elements: Oracle's entry into the field with massive borrowing largely covered by the markets; the beginning of interest-rate cuts by the Federal Reserve; and $500 billion in dry powder
– i.e., committed, but unallocated capital – from private credit.
More bonds and fewer constraints
Thirdly, according to the JPMorgan strategist, there are financial techniques to accelerate investment expansion. One example is the South Sea bubble in the early 18th century, which ushered in the transformation of debt into shares. Today, tech giants with low debt levels would only need to bring themselves up to the average debt level of other groups to have an additional trillion in capital. It should be noted that this also applies symmetrically to banking giants, which are obtaining lower capital requirements, despite the Basel rules, with the government's approval and the Fed's consent.
The banker concludes: Exuberance is building, but it would need to reach much higher levels before we would grow more cautious. When we consider the evidence, it seems clear that the ingredients for a market bubble are present. [...] We think the risk that a bubble will form in the future is greater than the risk that we may be at the height of one right now. [...] No one knows who will ultimately capture the value from the AI revolution. [...] But the biggest risk, to us, is not having exposure to this transformational technology
.
This analysis partly motivates JPMorgan's decision to strengthen its ten-year strategic investment fund (Security and Resiliency Initiative
) from $1,000 billion to $1,500 billion. But the banking group's alignment with the White House and Pentagon's rearmament plans also weighs heavily.
Goldman Sachs and BNP Paribas
Goldman Sachs is among the optimistic groups on Wall Street but differentiates the AI universe based on share prices. It warns that investors aren't willing to reward all AI big spenders the same
. Doubts are growing about companies that were betting on rapid productivity gains and those most exposed to infrastructure construction, with meagre profits and increased debt. There is greater confidence in companies more focused on marketing and revenue growth (semiconductors, data centre operators, technology hardware, and electricity companies). For hyperscaler manufacturers, growth is expected to slow from 75% in the third quarter to 25% by the end of 2026. However, Goldman itself doubts the validity of forecasts that have so far proved too low and credits a potential doubling of capital expenditure across the sector, currently only 0.8% of GDP compared to 1.3% at the peak of the technology booms of the last 150 years.
BNP Paribas, Europe's leading banking group in terms of assets, paints two contrasting pictures. One reflects scepticism dictated by Europe's lag in AI, highlighting the risks; the other assesses its potential as an investment area. It concludes that AI has not yet entered bubble territory. It highlights the risks: the race for supremacy among high-tech giants may lead to infrastructure oversupply; huge initial investments with future revenues and profits create uncertainty about the real return on invested capital; growing debt financing, recourse to private credit, and the rapid deterioration of graphics processing units (GPUs) combine to create high levels of risk; and the circular relationships
between suppliers and customers are causing alarm. We are concerned about the systematic risk associated with the interrelated financial dependencies
, reads BNP's Investment Outlook 2026.
However, the bank also makes generous concessions. First, the players are large, rational companies with solid balance sheets and good liquidity. Second, the AI market is still young; only 16% of companies that have adopted it have implemented it widely, and new applications are in the pipeline. Third, the existing infrastructure is consistent enough to allow for immediate commercial applications. This was not the case during the dot-com bubble, when optical fibre ran years ahead of last-mile
broadband access networks and smartphones. ChatGPT has already reached an estimated 800 million weekly users in less than three years, compared to the thirteen years it took for the Internet to be adopted.
Debt and Asia challenges
The artificial intelligence bubble is being examined by bankers and technologists in terms of its internal mechanisms, bottlenecks, and financing models. But external factors are just as decisive or influential, as demonstrated by the chill that has descended on the green economy. Ruchir Sharma, writing in the Financial Times, notes that if federal debt continues to worsen, buyers of Treasury bonds will demand higher yields, which would be reflected in the cost of capital: public debt could therefore be the short fuse that will cause the AI industrial bubble
to burst.
The Economist points out that there is also a race for AI in Asia (Japan, South Korea, Taiwan), with technology stocks less expensive than their peers around the world, and considerably cheaper than those in America
. In Northeast Asia, average share prices are less than three times their book values, compared to more than five times in the United States. Competition here, even more than on products and technologies, is on the cost of capital.
Chatham House, a British think tank, warns of the danger that American high-tech companies are relying too heavily on government contracts. This dependence could potentially cede large portions of the global AI market to China, which has proven itself capable of filling gaps, as it did with mobile telephony, solar energy, and electric vehicles.