The alarm no longer comes only from the skeptics. In June 2026, the Bank for International Settlements compared the AI boom to the great financial bubbles of the past. Oracle has slumped, the hyperscalers are committing hundreds of billions with no guarantee of a return, and business customers themselves are starting to worry about their bills. A clear-eyed look, without the doom.
On 28 June 2026, the Bank for International Settlements (BIS), the central bankers' bank, published its annual economic report. The chapter is titled « Progress and peril ». In it, the BIS openly compares today's AI euphoria to four historical manias: the canal mania of the 1830s, the British railway mania of the 1840s, the electrification exuberance of the 1920s and the dot-com bubble of 2000. All ended with a reversal in investment and a recession.
Its figures are blunt: the five largest hyperscalers are set to spend more than 1,000 billion dollars on AI-related capital expenditure between 2025 and 2026, a pace that is « outpacing their earnings and free cash flow ». The risk, the BIS warns, is that disappointment over returns turns this capex boom into a « protracted investment bust », with knock-on effects on financial conditions.
The BIS is not alone. The Bank of England calls valuations « stretched », on some measures close to the peak of the dot-com bubble (October 2025). The IMF, the ECB and the Federal Reserve all warned in late 2025 of the risk of a « sharp correction ». In the Fed's May 2026 survey, nearly one professional in two now cites AI as a major risk to financial stability, up from one in three six months earlier.
The ECB stresses it: unlike the stars of the dot-com bubble, today's hyperscalers show high margins, solid growth, little debt and diversified revenues. That is why most institutions speak of « stretched valuations » and « concentration » rather than a « bubble » in the strict sense. The nuance matters: the risk is not that these giants collapse, but that the market abruptly resets its expectations.
No company embodies the mood swing better than Oracle. On 10 September 2025, its share jumped 36% in a single session, its best since 1992, lifted by a cloud order book that had ballooned to 455 billion dollars. Larry Ellison briefly became the richest man in the world.
Ten months later, the wind has turned. In late June 2026, Oracle posted its worst stock-market week since 2001 and the bursting of the dot-com bubble, with a roughly 19% plunge over five days. From its peak, the share has lost close to 58%. The reason: worry over the debt taken on to finance its data centers, and over the solvency of its AI customers. Oracle's debt climbed to around 130 billion dollars, its free cash flow turned negative, and Moody's cut its outlook to negative, citing « counterparty risk ».
The heart of the risk lies in a single contract. According to the Wall Street Journal (10 September 2025), Oracle and OpenAI signed a computing deal worth about 300 billion dollars over five years, starting in 2027, or nearly half of Oracle's order book. Yet OpenAI booked around 13 billion dollars of revenue in 2025 and expects to lose money through 2028. The whole question is whether the customer can pay for the capacity the supplier is building for it.
The BIS flags a worrying arrangement: chipmakers and hyperscalers take equity stakes in the AI labs, which then buy chips and cloud back from them. Nvidia invests in OpenAI, owns a slice of GPU renters, and so on. The line doing the rounds captures the fragility: « if OpenAI cannot pay Oracle, Oracle cannot pay Nvidia ». The BIS notes that these deals are « poorly disclosed », with the risk of the same asset being pledged several times.
Amazon, Microsoft, Google and Meta together invested on the order of 400 billion dollars in infrastructure in 2025, and are aiming much higher in 2026: about 200 billion for Amazon alone, 180 to 190 billion for Google, 125 to 145 billion for Meta. All of it without certainty as to the timing, or even the existence, of a return to match the sums committed.
Their leaders no longer hide it. Mark Zuckerberg admits that « misspending a couple of hundred billion would be unfortunate », while judging the risk « higher on the other side », that of building too slowly. Sundar Pichai acknowledges « elements of irrationality » and warns that « no company is going to be immune, including us » if the bubble bursts. Sam Altman himself reckoned back in August 2025 that investors were « overexcited » and that « someone is going to lose a phenomenal amount of money ».
If a bubble bursts, not everyone is in the same boat. The large cloud providers would probably survive: they hold net cash, diversified revenues, long contracts and control of the scarce resources (power, land, permits, grid connections). That is the key difference with the debt-laden telecom operators of the year 2000.
The most exposed sit downstream: the pure GPU renters (the « neoclouds », one of which carries more than 21 billion in debt and a share price more than halved from its peak), Oracle, loss-making start-ups, hardware makers, subcontractors, and the customers who bet everything on a single supplier. A correction would cleanse valuations, but it would claim very real victims.
The doubt no longer concerns only the financial markets. On the customer side, one concern is growing: the lack of control, transparency and cost predictability of proprietary models. According to the FinOps Foundation, 73% of companies say their AI spending has exceeded their forecasts. In practice a bill can spiral quickly, and lock-in looms as soon as processes are tuned for one specific provider.
Add to that the availability risk: a model can be withdrawn, restricted or made more expensive overnight, for regulatory or commercial reasons the customer does not control. We documented this in our article on the suspension of Fable 5 and Mythos 5.
Even the will to spend runs into physical reality. Data centers accounted for about 1.5% of global electricity in 2024, a share set to double by 2030 according to the International Energy Agency. In the United States they could reach up to 12% of national consumption by 2028. Yet building a high-voltage line takes four to eight years, and lead times for transformers and gas turbines have exploded.
On top of that comes local opposition. A March 2026 Gallup poll shows that 71% of Americans reject a data center near their home. In Virginia, the « Digital Gateway » mega-project (nearly 25 billion dollars) was definitively killed off in early July 2026. In Ireland, a de facto moratorium on connections in Dublin was only lifted, under strict conditions, in December 2025. These constraints make expansion plans far more uncertain than they look on investor slides.
Most observers converge on the same scenario: a phase of tightening, a « squeeze », in which AI players will have to prove their profitability or raise prices sharply. Consultancy Bain estimates that 2,000 billion dollars in annual revenue would be needed by 2030 to fund the planned capacity, and that 800 billion could be missing. The Sequoia fund had already posed its « 600 billion dollar question » back in 2024.
The paradox is that prices face pressure in both directions. Upward: the premium plans at 200 dollars a month and usage-based billing are multiplying. Downward: the arrival of very cheap models, such as DeepSeek in early 2025, set off a price war. The « squeeze » arises precisely from this collision between the need to earn margins and the commoditization of compute.
History calls for nuance. After the dot-com bubble burst, telecom operators had massively over-invested in fiber optics. Many went bankrupt, including WorldCom, but the infrastructure laid down became the backbone of the modern internet. The nineteenth-century railway mania ruined shareholders, with railway stocks falling 65%, while it tripled the British network. Good for society, painful for the impatient investor.
An honest caveat: unlike fiber, which is passive and useful for decades, AI chips depreciate fast (on the order of 20% a year) and age within a few years. What will survive intact is mostly the « shell »: buildings, the power grid, fiber, land and permits, not necessarily the silicon inside them today.
That is the thrust of Jeff Bezos's remark, describing an « industrial bubble » rather than a purely financial one: « when the dust settles and you see who the winners are, society benefits. AI is real, and it is going to change every industry. » A bubble and a real technology are not mutually exclusive.
You do not have to settle the « bubble or not » debate. Four reflexes stay winning whatever the scenario.
Adopt a multi-model architecture, keep your prompts and your data portable. The freedom to switch providers is your best insurance.
Caps, alerts, tracking by use case, quarterly review. A FinOps approach avoids nasty surprises on the bill.
Clauses on service continuity, exit and data retention. Beware of technical as much as commercial lock-in.
Not on the hype. A pilot that produces no measurable value should be stopped or reworked, not extended out of habit.
The bubble may be real. So is the technology. The organizations that will weather a possible correction best are those that treat AI as a critical supplier, measure its value, and keep their freedom to manoeuvre.
Central banks (BIS, ECB, Bank of England) mostly speak of « stretched valuations » and correction risk rather than a bubble in the strict sense. The nuance: unlike 2000, the big players are highly profitable and carry little debt.
The large cloud providers would probably survive. The most exposed are GPU renters, suppliers, subcontractors and customers reliant on a single player. For a business user, the real risk is rising prices and availability.
Four reflexes: do not depend on a single provider, manage costs like a cloud budget (FinOps), demand reversibility and contractual transparency, and judge each use case on its real ROI.
Molderez Consult SRL helps Belgian organizations build AI that is profitable, portable and well-governed, whatever the market climate.
Assess my AI strategyTransparency: this article was drafted with the help of artificial intelligence and reviewed by Molderez Consult SRL.