AI Debt Goes Systemic: BoE Counts $450B Issued and $700B of Data-Centre Capex Routed Through Private Credit as Carlyle Flags Subprime-Style Underwriting

AI Debt Goes Systemic: BoE Counts $450B Issued and $700B of Data-Centre Capex Routed Through Private Credit as Carlyle Flags Subprime-Style Underwriting

The week of September 28 – October 5, 2026 marked the moment AI-infrastructure financing stopped being a private-credit growth story and became a financial-stability story. The Bank of England's Financial Policy Committee (meeting Sept 25, record published Sept 30) put hard numbers on the pipeline, and one of the industry's own — Carlyle's head of global research — publicly compared its underwriting to pre-2008 subprime mortgages.

The BoE FPC's numbers (primary source):

• Global AI-related debt issuance totaled around $450 billion as of early September (Morgan Stanley estimate) — "more than double the total issuance in all of 2025."
• JPMorgan analysts estimate AI capex financed through debt issuance will total around $4.1 trillion between 2026 and 2030.
• Morgan Stanley estimates $700 billion of data centre capital expenditure between 2026 and 2028 will be financed by private credit — the first time a major regulator has quantified the private-credit share of the buildout.
• The FPC's core warning: "The increasing indebtedness of AI firms combined with opacity and, at times, 'circular arrangements' that can be associated with this financing, can complicate the assessment of risks and could amplify losses if expectations disappoint." It added that AI-issuance expectations are now entangled with sovereign debt markets, since growth and fiscal outlooks depend on AI productivity gains materializing.

Carlyle's inside warning. Jason Thomas, Carlyle's head of global research and investment strategy, told Bloomberg Television in mid-September that AI project financing is often underwritten "against the investment-grade credit of the sponsor, the hyperscaler or big tech backer, rather than against the actual economics of the data center itself" — the same shortcut that let subprime mortgages be rated as if the underlying house didn't matter. Carlyle estimates the buildout needs roughly $1 trillion in financing, more than half of all the money private credit currently manages. Carlyle is not a bystander: it was part of the syndicate (with Blackstone, BlackRock, and DigitalBridge) that assembled CoreWeave's $7.5 billion private loan package.

The market is already pricing marginal projects. Two data points from late September show where the financing stack is cracking:

• SB Energy (Nvidia-backed, central to the Ohio campus serving OpenAI-linked demand): a proposed $4.9 billion debt issue drew weak interest at prospective yields around 10%, and marketing was postponed on Sept 25 while the sponsor addressed SEC questions about valuation and OpenAI reliance. Nvidia has invested $1.5B and committed another $1.5B at IPO price.
• Nscale closed $3.36 billion of pre-IPO convertible financing on Sept 25 — Third Point led, with funds managed by Apollo, Citadel, and ADIA participating, and Nvidia committing a $1 billion tranche funding mid-November. The Nvidia-led consortium (announced in August with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR to mobilize $500B+) is thus producing its first concrete closings — via convertible and venture-style structures, not the senior direct-lending product the market expected.
• Goldman Sachs research circulating the same week: investors have provided ~$500 billion of financing to AI-linked groups, with private debt funds' loans to the tech sector at ~$450 billion, up ~$100 billion since early 2024.
• Nvidia is now sharing GPU depreciation and compute-price data with insurers (FT) to make GPU-collateralized loans insurable — an attempt to manufacture a market for the collateral's residual value.

Australia's regulator struck a similar note. The RBA's October Financial Stability Review (published ~Oct 2) found "risks in international private credit markets are growing, although spillovers have been limited," and its AI box flagged that "a rise in off-balance sheet financing via special purpose vehicles increases hidden risks" — the same circularity the FPC cited.

What it means for a strategist: Three numbers frame the risk: $450B issued, $700B of data-centre capex routed through private credit by 2028, $4.1T of debt-financed AI capex projected by 2030. The underwriting question Carlyle poses — project economics vs. sponsor balance sheet — is the same one the FPC raised as "circular arrangements." The 10% yields SB Energy reportedly faced suggest the marginal AI project borrower is no longer getting 2024-era pricing; the cost of the next data centre is rising even as headline issuance doubles. This converges with the software-borrower stress tracked in AI and SaaS Concentration Risk in Private Credit — Underwriting Deterioration and the Medallia Takeover Case Study and the valuation debate in Apollo's Zito: "I Literally Think All the Marks Are Wrong" — Managers Concede the Valuation Problem as Proof Arrives From Banks, Short Sellers and Data Vendors.