Every day, one story that matters, broken down into three numbered scenarios, each with its own probability. Never fixed — it shifts as the situation changes.
Scénario
Friday, science
AI: US Tech's $725 Billion Bet
Published on September 11, 2026
Share:
The question at hand
US tech giants are pledging hundreds of billions of dollars to artificial intelligence: is the revolution catching up with the spending, or are markets on the brink of a crash like in 2000?
The facts
On 10 September 2026, analysis firm Capital Economics published a note that landed like a bombshell: according to its market economist James Reilly, "most indicators suggest the AI stock market boom is nearing its end." Of the eight signals he uses to spot a stock bubble running out of steam — valuations, index concentration, share issuance, foreign interest in US stocks — most, he says, already point to levels close to the dot-com speculative bubble* of 2000. His forecast: the S&P 500 would keep climbing until the end of 2026, before falling sharply in 2027.
This isn't an isolated warning. Since 2023, four American companies — Microsoft, Amazon, Google, and Meta — have more than tripled their annual spending on AI infrastructure: $725 billion planned for 2026, up from $410 billion in 2025, a 77% jump in a single year. These giants, nicknamed "hyperscalers*" for the size of their data centers, finance nearly all the chips, buildings, and electricity needed to run the most powerful AI models. Industry analysts are already calling it "the largest infrastructure investment cycle concentrated in a single year in the whole history of technology."
On the other side, revenue is struggling to keep pace. OpenAI, the maker of ChatGPT, has signed more than $1.4 trillion in commitments to buy computing power over the next ten years — including $500 billion for the Stargate project with Oracle and SoftBank alone. Its annualized revenue, meanwhile, stood at around $25 billion in mid-2026, with roughly $14 billion in losses expected for the year. We already reported, in early September, how OpenAI was aiming for a $1 trillion valuation* for its stock market listing, despite this imbalance. Some players, though, remain solidly profitable: chipmaker Nvidia still posts a 53% net margin, a level very few companies from the 2000s dot-com bubble could claim.
Yet this financial strength among the biggest players isn't enough to erase a deeper question: even if AI really does change the economy, does today's spending stand a chance of staying useful long enough to pay for itself?
Understanding it
The dot-com bubble of 2000 burst, but it also left behind infrastructure that went on to serve for twenty years.
Between 1996 and 2000, telecom companies laid hundreds of thousands of kilometers of fiber-optic cable, betting on demand that didn't exist yet. The bubble burst, many companies went bankrupt — but the fiber kept working. It ended up carrying the entire internet of the 2010s. AI's problem is different: a graphics processor loses most of its economic value in 2 to 3 years, not 20. If demand slows, today's data centers risk becoming obsolete long before they're paid off.
Investor Michael Burry, made famous for predicting the subprime crisis, has made exactly this lifespan gap his cause. In his view, big tech groups depreciate* their servers over 5 to 6 years, while their real lifespan is closer to 2 to 3 years. That gap, he argues, artificially inflates the sector's reported profits by roughly $176 billion between 2026 and 2028, by his own calculations. Nvidia has publicly rejected the criticism; Burry, for his part, has since sold off his bets against the stock, without that ending the debate over the accounting method itself.
A second, more discreet mechanism is precisely what worries financial authorities: how a good chunk of this money is being financed. Since the start of the year, Nvidia, OpenAI, Oracle, and AMD have signed a web of interlocking deals: Nvidia invests in OpenAI, which in turn commits to buying a comparable amount of Nvidia chips; Oracle is building $300 billion worth of computing capacity for OpenAI under the Stargate project; AMD, for its part, is offering OpenAI the option to buy its own shares at a token price, in exchange for large-scale chip purchases. The International Monetary Fund and the Bank of England both warned, as early as October 2025, that this kind of circular financing* posed a risk to global financial stability. Here, from biggest to smallest, are the main deals already signed in 2026:
The biggest circular AI bets of 2026
1OpenAI – Oracle (Stargate) — $300BSigned in July 2026 · up to 4.5 GW of computing capacity reserved for OpenAI over several years.
2Nvidia – OpenAI — up to $100B signedWithin an announced $600B framework · Nvidia invests in OpenAI, which commits to buying back a comparable amount of Nvidia chips.
3Off-balance-sheet private debt*, cumulative — over $120BMeta, xAI, Oracle, CoreWeave · shifted since late 2025 to private investors (Blue Owl, Pimco, BlackRock, Apollo...).
4Meta – Blue Owl (SPV "Beignet Investor") — $30BOctober 2025 · Hyperion data center in Louisiana, 90% off-balance-sheet financed by private debt.
5AMD – OpenAI — warrant agreementOpenAI can buy up to 10% of AMD's equity at a token price, in exchange for large-scale chip purchases.
Five different structures, one same principle: the supplier becomes the investor (or vice versa), circulating the same money among a small number of players rather than bringing in genuinely new capital.
The problem isn't that these deals are illegal or hidden — they're public and announced with great fanfare. It's the way part of this money is accounted for that worries regulators.
Understanding it
Circular financing isn't a scam, but it can give a false impression of solidity.
When a company borrows through a special-purpose vehicle* rather than directly, the debt doesn't show up on its own books — only the rent it later pays appears there. Meta, for instance, borrowed $27 billion for its Louisiana data center without that sum appearing as debt on its balance sheet. The real risk doesn't change, though: if AI demand disappoints, the same infrastructure loses value, just financed by creditors less visible to anyone looking only at the official balance sheet.
This climate of financial distrust doesn't tell the whole story about AI's real usefulness, though. A MIT study published in summer 2025 found that 95% of corporate generative AI pilot projects were still generating no measurable gain in results — but it also noted that the 5% that do work, often back-office automation of repetitive tasks rather than flashy customer-facing uses, deliver genuine gains. In other words: the question isn't whether AI is useful for anything — it already is, in places — but whether it's useful for enough things, fast enough, to justify a bill running several hundred billion dollars a year.
Combined AI capex of the 4 US giants (Microsoft, Amazon, Google, Meta)$725B (2026, +77% year-on-year)
Weight of the "Magnificent Seven*" in the S&P 500≈ 34% (late 2025 – mid-2026)
Favorable, stable, or degraded
Will the AI bubble eventually burst?
What we're assessing
FavorableAI revenue gradually catches up with spending, without a major financial shock.
StableSpending and warning signs keep growing together, with no crash in the immediate term.
DegradedA trigger event sets off a broad market correction.
Favorable
20%
Unlikely
AI revenue catches up with spending, no crash
By mid-2026, OpenAI posts annualized revenue near $25 billion, sharply up year-on-year, and the professional uses that already work (back-office automation, customer service) start delivering real, measurable productivity gains. The tech giants, for their part, stay profitable: Nvidia still posts a 53% net margin, worlds away from the revenue-less companies of the dot-com bubble. Growth in usage gradually catches up with spending, without a financial shock interrupting the momentum.
This is the least likely of the three scenarios, because it requires the already huge gap between $1.4 trillion in commitments and the much more modest revenue of the main AI players to close quickly, when a recent study shows 95% of corporate AI pilot projects still deliver no measurable gain. Less likely than stable, which requires no sudden improvement in usage. Also less likely than degraded, since warning signs have multiplied in recent months rather than easing.
Indicators affected
Combined AI capex of the 4 US giants$800-850B by end of 2027 (more moderate rise)↑(vs $725B in 2026)
Weight of the Magnificent Seven in the S&P 50028-30% (sharp pullback, broadening market)↓(vs ≈ 34% today)
The France angle
An orderly normalization benefits French savings invested in international equities (life insurance, PER retirement plans) via global funds, without the shock of a broad crash. ↑ Rather favorable for France.
Stable
45%
Likely
The bubble keeps growing, without bursting right away
US tech giants maintain, or even raise further, their 2027 spending forecasts, buoyed by very real profits that reassure part of the market. The most fragile financing structures — loans backed by data centers, cross-guarantees between suppliers and customers — keep piling up, with no specific event triggering a broad correction. The gap between AI spending and real AI revenue keeps widening, but markets choose, for now, to look past it.
This is the most likely of the three scenarios, because it extends exactly the trajectory observed since 2023, with no precise date settling things one way or the other. More likely than favorable, which would require a much faster revenue catch-up than current figures show. Also more likely than degraded, since the tech giants involved remain, for now, profitable enough to keep absorbing the shock without being forced to.
Indicators affected
Combined AI capex of the 4 US giants$950B - $1T by end of 2027 (further rise)↑(vs $725B in 2026)
Weight of the Magnificent Seven in the S&P 50034-36% (near-stable, slight rise)→(vs ≈ 34% today)
The France angle
American giants capture most of the industrial gains, while French savings remain exposed to a bubble that keeps growing without any French player really benefiting. ↓ Rather unfavorable for France.
Degraded
35%
Likely
A crash wipes out some of AI's stock market gains
One link in the circular financing chain — a default on a data-center-backed loan, or the revelation of an overly fragile private-debt structure — triggers a broad loss of confidence. The tech giants, who already account for about a third of the S&P 500, drag the entire stock market down with them. Infrastructure spending planned for 2027 is abruptly revised downward, and credit-financed data centers become assets that are hard to resell.
This scenario remains less likely than stable, since it requires a genuine trigger event — a payment default or an accounting revelation — not just a buildup of signals the markets already know about. But it is more likely than favorable, because the scale of private debt already committed (over $120 billion off-balance-sheet) now weighs more heavily than a merely theoretical risk.
Indicators affected
Combined AI capex of the 4 US giants$500-550B by end of 2027 (sharp pullback)↓(vs $725B in 2026)
Weight of the Magnificent Seven in the S&P 50020-24% (marked drop)↓(vs ≈ 34% today)
The France angle
A crash in tech stocks would hit the unit-linked savings of French households and could trigger a global economic slowdown that France, already fiscally fragile, would struggle to recover from. ↓ Rather unfavorable for France.
Indicative orders of magnitude for the 3 scenarios above, estimated with the information available at publication and re-assessed if the situation changes — never guaranteed forecasts. Learn more about our method →
Key takeaways
US tech giants are pledging hundreds of billions of dollars to artificial intelligence: is this revolution catching up with the spending, or are markets on the brink of a crash?
The four largest US tech groups will spend $725 billion in 2026 to build AI infrastructure, up 77% year-on-year, while the IMF and the Bank of England already compare stock valuations to those of the dot-com bubble of 2000.
The most likely outcome (45%): the bubble keeps growing without bursting right away, buoyed by very real profits at the tech giants, despite increasingly visible signs of fragility in the riskiest financing structures.
Signal to watch: the quarterly results of US tech giants, especially Q4 2026 (published late January-February 2027), when investors will judge whether AI-related revenue growth finally justifies these sums.
Fairly negative
Our assessment of the impact for France: fairly negative. The stable scenario, the most likely (45%), leaves French savings exposed to a growing bubble without capturing its benefits; the degraded scenario (35%) would expose them directly to a crash. Together, these two scenarios carry an 80% probability, against just 20% for an orderly, favorable normalization.
A rise in the price of an asset (stocks, real estate...) largely disconnected from its real value, fueled by the expectation that prices will keep rising — until confidence turns.
Hyperscalers
The cloud and AI giants (Microsoft, Amazon, Google, Meta...) that build the data centers AI needs, at massive scale.
Valuation
The total estimated worth of a company, calculated during a funding round or stock market listing — not actual revenue or profit collected.
Depreciation
In accounting, the way a company spreads the cost of equipment over several years rather than counting it all at once — the longer the chosen period, the higher each year's reported profit appears.
Circular financing
A structure where a company invests in one of its own customers or suppliers, so that the money injected ends up coming back to it as purchases — which can artificially inflate the reported revenue on both sides.
Special-purpose vehicle
A legally separate company, created solely to carry the debt of a specific project (here, a data center) — which lets the company that benefits from it keep that debt off its own books.
Private debt
Loans granted by investment funds rather than traditional banks, often less regulated and less visible to financial authorities.
Magnificent Seven
The nickname for seven American tech giants (Nvidia, Apple, Microsoft, Alphabet, Amazon, Meta, Tesla), whose weight in stock indices has grown so large that they alone drive a large share of market performance.
