ECB Classifies AI as Systemic Risk for First Time: $4 Trillion in Off-Balance-Sheet Commitments Sounds Subprime Alarm

The European Central Bank’s annual Financial Stability Report ranks AI valuation-bubble risk as the top threat to euro area financial stability and traces

On November 18, 2026, the European Central Bank published its annual Financial Stability Report, using the most authoritative institutional language to date to rank AI valuation-bubble risk as the top of three core threats to euro area financial stability. What makes the report unusual is that it precisely describes a systemic contagion path running from tech stocks to sovereign debt and then to banks and non-bank institutions, as well as a structural flaw that has never appeared in this form in history.

The Numbers Alone Are Striking Enough

The core data cited in the report is daunting in magnitude: from 2025 to 2032, AI infrastructure investment needs will exceed 3% of US GDP each year; cumulative capital expenditure by major US tech companies is projected to reach $5.5 trillion by 2030; and by the second quarter of 2026, Big Tech’s off-balance-sheet lease and purchase commitments had exceeded $4 trillion, far higher than their total on-balance-sheet liabilities.

More critical are the cash flow data. The ECB report notes that in 2026 the free cash flow of most US hyperscale cloud companies has turned negative—capital expenditure has exceeded their operating cash flow, forcing companies to turn on a large scale to the bond market, private credit, and other external financing channels. The report specifically names Nvidia and Broadcom as exceptions, as both still have strong cash-generation capacity; Nvidia’s quarterly revenue of $115.8 billion reported this week corroborates this assessment.

ECB Vice-President Boris Vujčić was blunt when the report was released: “Given persistent geoeconomic uncertainty and growing exposure to AI, financial stability risks in the euro area remain high.”

Why Summon Two Ghosts at Once

In the report, the ECB rarely cites two of the most devastating bubble precedents in history at the same time: the 2000 dot-com crash and the subprime bubble that triggered the global financial crisis in 2008. This is not rhetorical exaggeration but a technical distinction with specific meaning.

The two analogies point to different layers of risk. The similarity with the dot-com bubble lies in the scale of investment: a new technology attracted a capital influx surpassing any historical precedent in a short period, while the pace of commercialization lagged far behind. But the dot-com bubble was largely equity-financed; when it burst, losses were borne by shareholders, and the impact on the real economy was relatively limited—that recession was shallow.

The subprime analogy points to the more hidden and more dangerous structural features of this AI boom: debt expansion, off-balance-sheet opacity, and circular self-financing.

The report identifies three specific mechanisms. First, chipmakers are financing their own customers’ purchases, which artificially inflates apparent demand and makes the supply chain look more robust than it is. Second, hyperscalers hold equity stakes in AI developers, and rising valuations of the latter in turn lift the former’s book gains, forming a closed-loop, self-reinforcing valuation circuit. Third, much AI infrastructure debt matures far later than the one-to-three-year economic useful life of AI chips—meaning companies may still be repaying debt borrowed to buy these assets after the assets have already depreciated and been retired.

The report also specifically notes that US accounting standards do not count these off-balance-sheet commitments as liabilities, leaving the true risk exposure nearly opaque to outside observers. This opacity was one of the key conditions that enabled the 2008 subprime crisis to build to systemic scale while the market broadly underestimated risk.

Contagion Chain: From Tech Stocks to Euro Area Sovereign Debt

The risk transmission path depicted by the ECB spans multiple markets, making it fundamentally different from past cyclical adjustments in the tech industry.

The path begins with a rupture in valuation logic. The current market capitalization of hyperscalers rests on “grand profit expectations”; once the market begins to doubt whether the pace of AI commercialization and monetization can support these expectations, or competitive pressure compresses these companies’ pricing power, repricing will occur. Because AI infrastructure is increasingly debt-financed, losses in the stock market will spread to the bond market through the balance sheets of the companies holding these positions—precisely the key difference from the dot-com model of purely equity financing.

Euro area non-bank financial institutions’ equity exposure to US tech and AI corporate securities accounts for roughly 16% to 26% of their total sample assets under management. Banks are exposed to this risk both through direct holdings and through exposure to non-bank institutions. Meanwhile, sovereign debt markets are also under strain: debt levels have risen in several euro area countries, and news-sensitive and leveraged investors account for a growing share of the sovereign bond investor base, raising the probability of abrupt repricing in tandem. Vujčić explicitly warned that sustained high exposure to AI could simultaneously hit banks’ and non-banks’ AI holdings and sovereign debt, triggering cross-contagion.

The end of this chain is not the share price of any tech company but the stability of the euro area financial system as a whole.

Anthropic’s IPO: The Paradox Made Concrete

The day before the report was released, Anthropic had just completed a historic IPO, raising about $81.3 billion and reaching a market value of about $1.5 trillion. The coincidence of timing is highly symbolic.

According to India Today, Anthropic’s prospectus itself contained candid disclosure of risks across the AI industry: the company is investing heavily, data centers are under unprecedented pressure from compute demand, and its models were even involved in a security incident involving access to unauthorized systems. Yet the $1.5 trillion valuation corresponds to a company still exploring its path to profitability.

The core issue is not Anthropic itself but the industry logic reflected by this valuation: the market is currently willing to pay a premium for the proposition that “AI will change everything,” regardless of when that change arrives, in what form, or how much cost companies must bear. This is exactly the “fault line between present value and earnings expectations” described by the ECB.

Banks Are Sound, but Non-Bank Institutions Are the Weak Link

In the report, the ECB explicitly distinguishes the risk status of two types of institutions. The euro area banking system currently has ample capital buffers, its non-performing loan ratio remains at a low 2.2%, and liquidity is abundant; it is deemed capable of withstanding potential shocks.

But the situation of non-bank institutions is more complex. Hedge funds, money market funds, insurers, and pension funds have relatively high exposure to AI corporate securities, while macroprudential regulatory frameworks in these areas remain incomplete. The report notes that increasingly close links between banks and non-bank institutions could become a channel for cross-sector transmission and amplification of shocks.

This means that even if the banking system itself can withstand the first wave of shocks, the fragility of non-bank institutions could still pressure banks in turn through asset sales and liquidity strains, making the overall system’s resilience weaker than any single institution’s capital adequacy ratio suggests.

Independent Judgment: What This Report Really Changes

The ECB report does not precisely predict when the AI bubble will burst, and the report itself explicitly states that it cannot provide a timeline for repricing. It is a risk assessment document, not a prophecy.

But what it changes is the discursive landscape around the “legitimacy narrative” of AI infrastructure investment. Over the past two years, the AI industry’s justification for compute investment has roughly been: larger scale means stronger models, stronger models mean higher commercial value, and therefore any investment is justified. This logic has internal coherence in an environment dominated by equity financing, because losses are ultimately absorbed by equity investors with a higher risk appetite.

But when the financing structure shifts from equity to debt, when off-balance-sheet commitments exceed $4 trillion and are opaque to outsiders, and when a structural mismatch emerges between asset lives and debt maturities, this logic no longer holds. The ECB’s first publicization of this judgment in the form of a Financial Stability Report means that the AI industry will face more institutional constraints in the future regulatory environment and financing costs.