Assessment: systemic financial risk from the AI investment boom is increasing, but it has not yet become acute.

The important change is no longer merely that AI shares have entered a correction. New official US data now show how quickly advanced-computing borrowing has grown, while Anthropic’s flotation prospectus gives an unusually clear view of the scale and duration of one frontier laboratory’s compute commitments. Together, these developments strengthen the evidence that the AI boom is becoming more deeply connected to public debt markets, private credit, project finance, leases and contingent obligations.

The correction has deepened

The Nasdaq Composite entered a technical correction on 5 November, falling more than 10 per cent from its 6 October record. Reuters estimated that more than $2.5 trillion was erased from the index during October as investors reassessed whether rapidly rising AI capital expenditure can produce returns quickly enough.

By 16 November, Nvidia had lost about $1.9 trillion in market value since early October. Several AI infrastructure providers had fallen by at least 40 per cent from the October peak, while banks and credit investors were reported to be applying greater scrutiny to new AI-related debt deals. This remains an orderly repricing rather than a funding seizure, but the scale and persistence of the decline show that investors are no longer pricing near-perfect execution across chips, data centres, power supply and AI services.

The Federal Reserve quantifies the debt buildout

The Federal Reserve’s Financial Stability Report, published on 13 November, provides the strongest new official evidence. It reports that debt issuance used to finance advanced-computing infrastructure exceeded $680 billion by October 2026, compared with about $130 billion during all of 2023. Advanced-computing borrowing rose from less than 6 per cent of non-financial investment-grade corporate debt issuance in 2023 to about 31 per cent in 2026. Outstanding public debt associated with advanced computing was estimated at more than $1.1 trillion.

The composition is also becoming more complex. Financing now includes public and private debt, structured credit and project finance, alongside rapidly growing lease commitments, backstop guarantees and other contingent obligations. This broadens the number of investors exposed to an AI disappointment and makes it harder to identify where losses would ultimately sit.

Anthropic exposes the commitment risk

Anthropic’s November flotation prospectus illustrates the execution risk at company level. The company reported approximately $900.9 billion of purchase commitments and lease agreements as of and after 30 September 2026. About $451.6 billion is expected to be incurred before the end of 2030, with substantially all of the remainder expected between 2031 and 2045. Roughly $70.7 billion falls due during the twelve months after 30 September 2026.

These are commitments rather than ordinary funded debt, and Anthropic also reported substantial liquidity, financing capacity and anticipated flotation proceeds. They should not therefore be treated as an immediate solvency deficit. They do, however, demonstrate how frontier AI firms must commit to chips, facilities and power long before the associated revenue is known. That timing mismatch becomes dangerous if demand, pricing or technical advantage develops more slowly than expected.

Debt and off-balance-sheet exposure are still expanding

The Bank of England reported in September that global debt issuance by AI-related firms had reached about $450 billion by early September 2026, more than twice the total for all of 2025. It cited projections of roughly $4.1 trillion in debt-funded AI capital expenditure from 2026 to 2030 and about $700 billion of private-credit financing for data centres between 2026 and 2028.

The Bank highlighted opacity, circular arrangements and the risk that borrowing structures could obscure where losses ultimately sit. Analysis from the Bank for International Settlements likewise found growing use of long-term debt and off-balance-sheet structures to finance AI infrastructure. Spillovers into US Treasury and investment-grade corporate markets have so far been small and short-lived, but sustained issuance could make that channel more important.

Semiconductor demand remains the counterweight

The correction has not yet been accompanied by evidence of collapsing semiconductor demand. Nvidia’s most recent reported quarter showed data-centre revenue of $89 billion, up 117 per cent from a year earlier, and the company expected further sequential growth. The next earnings report is therefore an important test: strong realised demand would support the case that the correction is removing excess valuation rather than signalling a collapse in the underlying investment cycle.

Some market analysis also points to moderating capital-expenditure growth and tighter credit standards as potentially stabilising forces. A slower, more selective buildout could improve underwriting discipline and spread projects over a longer period. That interpretation is plausible, but it does not remove the refinancing, concentration and collateral-life risks already embedded in the system.

What the evidence does not yet show

There is still no evidence of frozen core credit markets, forced bank deleveraging or a 2008-style funding shock. The Federal Reserve and Bank of England evidence describes rising vulnerability, not an active systemic crisis. Much of the debt has been issued by investment-grade technology companies with large revenues, and demand for advanced chips and data-centre capacity remains strong.

Verdict: risk is increasing. The deeper equity correction, the Federal Reserve’s new debt figures and Anthropic’s extraordinary long-term compute commitments all strengthen the case that an AI disappointment could now travel through more parts of the financial system. The evidence still supports a judgement of rising fragility rather than present systemic failure.