What a correction would not correct: the AI bubble and European dependence

On 31 October 2026, Martine Orange published an analysis in Mediapart under the title “Bulle de l’IA : le troisième avertissement des marchés”¹. The question in that headline now occupies the reports of financial institutions as much as the market pages: do artificial-intelligence valuations rest on revenues that do not yet exist, and will we know before the answer is administered by the markets?

We shall not settle that question, for a reason having less to do with caution than with competence: predicting the timing of a correction is not a legal exercise, and the economists who attempt it disagree. We would rather pose another question, one the financial framing tends to obscure and which does fall within institutional analysis: assume the correction happens. What would it change about European digital dependence?

The commonly held answer, usually implicit, is that it would loosen it. A sobered sector, cheap computing capacity, weakened giants, a rebalanced relationship, a political window opening for the European legislature. We shall argue the opposite: a downturn would tighten European dependence rather than loosen it, and it would do so through three mechanisms, none of which turns on the price of assets.

What the last week of October showed

Let us begin with the facts, which are more precise than the vocabulary of bubbles. On 30 October, artificial-intelligence shares fell as investors questioned the returns on the spending committed². Meta had lost 7% following its results of 21 October, even though its revenue beat expectations; the company announced a financing programme of some 95 billion dollars to fund its artificial-intelligence projects, 35 billion through an underwritten equity offering and 60 billion through an at-the-market programme; on 30 October, S&P Global revised the outlook on its rating downwards, citing cash flow and capital expenditure³. Intel fell 25% on disappointing guidance, and the S&P 500 shed close to 2% that Friday⁴.

Microsoft is the most instructive case, precisely because it is not a case of poor results. Azure revenue grew 48% year on year and the stock held at record levels. In the same quarter, the company’s free cash flow fell 27%, to 18.63 billion dollars, and the group announced that it would turn to the bond market in order to continue its investment in artificial-intelligence infrastructure⁵.

This point deserves to be dwelt on, because it governs what follows. What worried the markets at the end of October 2026 was not a shortfall in revenue. It was the way capital expenditure is now being financed. For 2026, the four largest buyers of computing capacity were together approaching 700 billion dollars: Alphabet raised its own forecast by 15 billion in July, to between 195 and 205 billion; Amazon held its own at around 200 billion; Meta raised its to between 125 and 145 billion; analysts expected a total above one trillion for 2027⁶. Sums of that order cannot be drawn indefinitely from operations.

A shift in structure, not an excess of valuation

This is precisely what the International Monetary Fund describes in its Global Financial Stability Report of October 2026, which identifies three salient risks and places first among them a market correction driven by artificial intelligence. The report notes that such a correction “could be exacerbated by rising leverage, concentrated exposures, and circular financing arrangements”⁷.

Two figures from that report deserve to be retained, because they measure not a valuation but a structure. Since 2024, the share of global investment-grade bond issuance accounted for by the largest data-centre operators has risen from less than 1% to 10%⁸. And the Fund counts fourteen publicly disclosed transactions since 2023, representing approximately 300 billion dollars of equity investment led by those same operators in financially weaker artificial-intelligence firms, supplemented by 650 billion of debt, in exchange for 1.4 trillion in future revenue commitments; roughly 320 billion of that debt is carried by the investee companies themselves, including debt backed by investor guarantees and debt held off balance sheet⁹.

What these figures describe is not an expensive market. It is a shift from equity financing to financing by debt and by crossed commitments. The distinction matters, because the two do not behave alike in a downturn. A loss of capital falls on the shareholder and stops there. A debt commitment survives the correction and continues to bind: it must be serviced, and to service it one must sell. An equity bubble deflates; a credit bubble is liquidated.

The contrary argument, and why it holds only halfway

There is a serious argument for the opposing view, and it ought to be taken in its strong form rather than its convenient one. Excess capacity lowers the price of access. The precedent is well known: between 1999 and 2001, telecommunications operators laid down, largely on credit, optical-fibre capacity calibrated on traffic that arrived only years later. The operators went bankrupt; the glass stayed in the ground. For a decade bandwidth was cheap, and universities and research networks were able to build backbones by acquiring unused fibre at liquidation prices¹⁰. Transposed, the reasoning runs as follows: an excess of data-centre capacity would lower the cost of computation for European buyers, who are not the ones who financed it.

The argument is right about price. It is a good deal less right about ownership, and this is where the same precedent turns around. The assets did not vanish: they changed hands. Global Crossing filed for bankruptcy in January 2002; its network was eventually acquired by Level 3 Communications in October 2011, for three billion dollars including the assumption of 1.1 billion of debt¹¹. A bankruptcy does not dissolve infrastructure. It transfers it to whoever still holds cash at the moment the price gives way. The useful question is therefore not whether computation will become cheaper, but who will own it afterwards — and that is a balance-sheet question.

The first asymmetry: European balance sheets

Put in those terms, the comparison is unambiguous. In September 2026, Mistral raised three billion euros at a post-money valuation above 21 billion, in a round led by Samsung, alongside the Scaleup Europe Fund — an EU-backed vehicle managed by EQT — and PSG Equity. The funds are intended for infrastructure, including data centres of its own, its chief executive stating that the company intends in time to rely entirely on capacity it builds itself¹². Set against the roughly 700 billion dollars of annual spending mentioned above, three billion euros is not one order of magnitude apart: it is three.

But the difference in structure matters more than the difference in amount. Mistral’s investment programme is financed by funding rounds, that is, by the market sentiment a correction destroys first. Microsoft’s, whose free cash flow fell 27%, will be financed by bond issuance. One can issue when sentiment turns; the other must wait for it to return. Arthur Mensch himself observed, at the time of the round, that “the volatility in this space is actually quite extreme”¹³. The observation is accurate and it does not work in his favour: volatility is a far heavier constraint on a company financed by equity than on one financed by rated debt.

The second asymmetry: a timetable tied to the cycle

The European instrument most directly concerned displays the same fragility, in budgetary form. On 30 July 2026 the EuroHPC Joint Undertaking launched its call for tenders for artificial-intelligence gigafactories, with a submission deadline of 12 November 2026, for up to seven sites located in at least seven Member States; selection is expected in early 2027 and operations within the eighteen months that follow¹⁴.

Two features of this arrangement deserve notice. The first is that public support takes the form not of a subsidy but of a purchase: the Joint Undertaking and several participating States “will jointly procure compute access time from the selected AI Gigafactories”, public money acting as an anchor customer intended to draw in private investment expected to exceed twenty billion euros. The second is that, on the Joint Undertaking’s own account, only part of that public financing falls under the current multiannual financial framework, the bulk of it to come from the next one, covering 2028 to 2035¹⁵.

The structure therefore rests on private commitments to be subscribed in 2027 and on appropriations to be entered from 2028 onwards. A tightening of financing conditions in 2027 would strike precisely that joint. We do not claim the scheme would fail; we observe that it is not insulated from the cycle but calibrated on its continuation, and that a downturn would hit it at exactly its thinnest point.

And it would do so from a distribution already settled. European providers held less than 15% of their own cloud market in 2025, while Amazon, Microsoft and Google controlled more than 70% of it; the two largest European firms in the sector, SAP and Deutsche Telekom, account for 2% each; and a European Parliament report establishes that at least 59% of the European enterprise-software market is held by American companies¹⁶. A correction operates on that distribution, not on a blank page. Proportional losses leave proportions intact.

The third asymmetry: law relaxes in a contraction

There remains the least discussed mechanism, and it is the one that falls most directly within our own field. One might expect a downturn to strengthen the legislature’s hand: a discredited sector, a less deferential public, a political window. The recent European sequence suggests the contrary.

The obligations applicable to high-risk systems under the artificial-intelligence regulation have already been deferred. For the stand-alone systems of Annex III, the deadline moves from 2 August 2026 to 2 December 2027; for systems embedded in products already governed by the harmonisation legislation of Annex I, from 2 August 2027 to 2 August 2028¹⁷. That deferral was granted in a period of euphoria, in the name of competitiveness and simplification. Yet if the ground is competitiveness, a contraction reinforces it: it is harder to refuse a delay to a sector doing badly than to one doing well. We do not predict a second deferral. We note that the argument which produced the first would be stronger, not weaker, in a downturn.

The coincidence of calendars is then worth looking at, because it is striking. The new deadlines fall in December 2027 and August 2028. The gigafactories are to be selected in early 2027 and brought into service eighteen months later, with the bulk of their public financing falling under the 2028-2035 framework. And the horizon of fragility described by the International Monetary Fund is the near term. These three sets of dates overlap. We draw no scenario from this; we note that the moment when European computing capacity is to be financed is also the moment when the deferred obligations become applicable and the moment the financial authorities designate as the most exposed. A policy that had treated these three calendars as independent would be poorly prepared for their meeting.

What a correction would not correct

We have argued elsewhere that the decisive variable in digital sovereignty is neither ownership nor location, but jurisdictional attachment and the reversibility of dependences¹⁸. The bubble debate allows that criterion to be tested, since it puts to the proof what asset prices can and cannot reach.

A correction alters valuations. It does not alter section 734.13 of the American Export Administration Regulations, nor the jurisdiction to which a model is attached, nor what a contract provides on portability. The interruption of June 2026 owed nothing to valuations and would have unfolded identically in a bear market. Reversibility is not a function of asset prices.

The one thing a downturn would alter in this register would work the wrong way. It would reduce the number of available suppliers, through bankruptcy and acquisition, and the existence of an alternative is the material condition of reversibility. Dependence on a supplier with three competitors is not equivalent to dependence on a supplier with none, even at the same price. It is in this sense that a bursting bubble does not emancipate: it concentrates.

Hence a practical conclusion, the same as the previous one, which the bubble debate leaves untouched. The measures that make a dependence reversible — portability clauses, documentation permitting genuine migration, a verified fallback, the capacity of an administration to keep functioning without its usual supplier — cost little and do not depend on the cycle. Their value does not vary according to whether the bubble exists, which is their chief merit at a moment when nobody can say. We wrote that it is quicker to organise an exit than to build a factory. A correction would not contradict that proposition. It would make the second term harder, and leave the first unchanged.

Martine Orange, “Bulle de l’IA : le troisième avertissement des marchés”, Mediapart, 31 October 2026. The article is behind a paywall and we have not been able to consult its full text: we take up the question it formulates, without discussing an argument we have not read. All the market facts reported below come from accessible sources, cited in their place. ↩
CNBC, “AI stocks sell off as investors question AI spending returns”, 30 October 2026. ↩
On the fall in the share price after the results of 21 October and on the announced financing programme, see CNBC, “Meta stock trade as AI spending surges”, 29 October 2026. On the outlook revision by S&P Global, see CNBC, “Meta outlook revised downward by S&P on cash flow and capex concerns”, 30 October 2026. ↩
CNBC, “Stocks started the week with promise but ended in despair”, 31 October 2026. ↩
CNBC, “Microsoft’s strong quarter backs up its stock rebound”, 28 October 2026, and “Tech giants need legacy businesses to keep growing as they pour trillions into the AI buildout”, 30 October 2026. ↩
On Alphabet’s raised forecast, see CNBC, “Alphabet raises the stakes on AI spending”, 24 July 2026; on the total approaching 700 billion for 2026, “Tech AI spending may approach $700 billion this year”, 6 February 2026; on the expectation of a total above one trillion in 2027, “Big Tech capital expenditures now seen topping $1 trillion in 2027”, 30 April 2026. These totals aggregate company guidance and analyst estimates; we retain them as orders of magnitude, not as accounts. ↩
International Monetary Fund, Global Financial Stability Report — Resilience, Risk, and Transformation, October 2026, overview page and executive summary. The three salient risks identified are an AI-related asset-price correction, a sovereign bond-yield shock and cyber threats. ↩
Ibid., chapter 1: “Since 2024, hyperscalers’ share of global investment-grade bond issuance has increased to 10 percent from less than 1 percent.” ↩
Ibid.: “Fourteen publicly disclosed transactions since 2023 account for approximately $300 billion of aggregate equity investments led by hyperscalers in financially weaker AI firms, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future revenue commitments.” The same chapter states that about 320 of the 650 billion in debt is taken on by the investee companies, including debt backed by investor guarantees, contingent financing and other off-balance-sheet debt. ↩
On the optical-fibre overcapacity produced by the investment wave of 1999-2001, its bankruptcies and the subsequent acquisition of unused fibre at low prices by university and research networks, see the account in Technostatecraft. Estimates of the proportion of fibre that remained unlit vary widely between sources and we retain none of them. ↩
Global Crossing filed for bankruptcy in January 2002; the acquisition by Level 3 Communications, announced in April 2011, was completed on 3 October 2011 for three billion dollars including the assumption of 1.1 billion of debt. These points of chronology should be confirmed against a primary source before being relied upon elsewhere. ↩
CNBC, “Mistral bags $24 billion valuation as Samsung leads funding for Europe’s AI champion”, 8 September 2026. The company announces a post-money valuation above 21 billion euros; the article’s headline gives 24 billion dollars, a figure the body of the text does not repeat. We retain the euro figure, which is the one communicated by the company. ↩
Ibid. Remarks by Arthur Mensch as reported by CNBC. ↩
EuroHPC Joint Undertaking, “The EuroHPC Joint Undertaking launches the AI Gigafactories Call”, 30 July 2026. The Joint Undertaking’s mandate was extended to the deployment of gigafactories by Council Regulation (EU) 2026/150 of 16 January 2026. See also the European Commission page devoted to the scheme, which records 77 expressions of interest covering 60 sites across 16 Member States. ↩
Ibid. The press release states that the Joint Undertaking and several participating States will jointly procure compute access time from the selected gigafactories, the joint public financing acting as an anchor customer and expected to unlock more than twenty billion euros of private investment; it specifies that part of the public financing falls under the current multiannual financial framework and that the bulk will come from the 2028-2035 framework. We note an ambiguity in the official communications: the Commission page presents the twenty billion euros announced in February 2025 as an amount to be mobilised under the InvestAI initiative, whereas the EuroHPC release uses a figure of the same order to designate the private investment expected. The two readings do not coincide and we do not resolve them. ↩
CNBC, “These four charts show how reliant Europe is on U.S. digital infrastructure”, 13 February 2026, drawing on Synergy Research Group data for cloud market shares and a European Parliament report for enterprise software. John Dinsdale, Synergy’s chief analyst, observes there that it will be “incredibly difficult” for European cloud providers meaningfully to reverse that trend. ↩
On the deferral of the deadlines applicable to high-risk systems by the so-called digital omnibus regulation on artificial intelligence, adopted by the European Parliament on 16 June 2026 and by the Council on 29 June 2026, see Future of Privacy Forum, “The AI Act implementation timeline: what changes under the AI Omnibus?”, 28 July 2026. The exact Official Journal reference could not be verified directly on EUR-Lex at the time of writing and should be before any citation in litigation. Besides these deferrals, the same text makes substantive changes — notably a new prohibition under Article 5, applicable from 2 December 2026, and a softening of the artificial-intelligence literacy duty under Article 4. ↩
Nicolas Heynderickx, “Sovereignty is not a market share: Europe, artificial intelligence and the question of jurisdiction”, 23 October 2026. On the episode of June 2026 and on section 734.13 of the Export Administration Regulations, see the references gathered there. ↩