DepreciatingCollateral

1.1The scale of investment needs

Projected investment in data centres is far above recent levels of infrastructure spending. In the central scenario of McKinsey, total investment of USD 6.7 trillion is needed by 2030, of which USD 5.2 trillion is for AI workloads (Figure 1.1).

Estimates across scenarios range from about USD 3 trillion to USD 8 trillion, which reflects high uncertainty about future AI demand.

The composition of this investment matters for credit risk. Around 60% of AI-related investment, or USD 3.1 trillion, goes to firms that produce chips and computing hardware (McKinsey, 2025). These assets can become obsolete much faster than buildings and energy infrastructure.

1.2A growing weight in debt markets

Hyperscalers increasingly finance their investment with debt. According to the OECD Global Debt Report 2026, nine large hyperscalers issued USD 122 billion in bonds in 2025, and their planned capital expenditure for 2026-2030 reaches USD 4.1 trillion (OECD, 2026a).

Data from Bank of America show that debt issuance by the five largest hyperscalers in 2025 was about four times the annual average of the previous five years (Figure 1.2, Panel A).

This trend also affects the structure of bond markets. The IMF reports that the share of hyperscalers in global investment-grade bond issuance has risen from less than 1% to 10% since 2024, while their free cash flow has declined (IMF, 2026).

The OECD Economic Outlook notes that risks could be repriced if expected returns from AI do not materialise, and that forced sales by leveraged non-bank financial institutions could amplify this process (OECD, 2025).

2.1A chain with five links

The financing of AI infrastructure can be described as a chain with five links (Figure 2.1).

Long-term investors provide funds to data centre owners through instruments such as bonds, special purpose vehicles (SPVs) and credit funds. These assets are leased to hyperscalers and AI developers that buy capacity under long-term contracts. The whole chain ultimately depends on what end users pay for AI services.

The BIS finds that off-balance-sheet structures strengthen the links between hyperscalers and non-bank investors such as private credit vehicles and insurers (BIS, 2026b). The BIS also notes that the terms of these arrangements are often poorly disclosed, and that the same asset may be pledged as collateral more than once.

2.2An example transaction: TPU lease financing

A transaction of about USD 35 billion closed in June 2026 shows several links of the chain in one structure (Figure 2.2). Private credit arranged by Apollo and Blackstone finances the purchase of Google-designed TPUs that are leased to Anthropic.

Two senior tranches are backstopped by Broadcom, which co-develops the chips with Google, and carry a coupon of about 5.75%. A junior tranche of about USD 4.5 billion carries the credit risk of the tenant and was sold at about 8.5% (The Paypers, 2026).

The structure illustrates two points. First, the rating often depends on the credit quality of the guarantor or buyer rather than on the value of the collateral. KBRA rated the junior tranche BB+, based on limited information selected by the company (FT, 2026). Second, Broadcom’s liability under the backstop is calculated as 85% of amounts outstanding minus the sale value of the hardware (Implicator, 2026a). The market value of the collateral therefore determines directly how losses would be shared.

3.1Hardware obsolescence

Debt maturities span several chip generations (Figure 3.1).

Nvidia plans to release a new data centre chip about every year.

By contrast, GPU-backed loans usually have maturities of three to five years, and some data centre SPV bonds run until 2049.

One market analysis finds that the main risk in this type of financing has shifted from obtaining GPUs to managing their rapid obsolescence (ENKI, 2026).

3.2Revenue deflation and margin pressure

Rental prices and token prices are moving in opposite directions (Figure 3.2).

The rental price of H100 GPUs fell sharply from its peak in late 2024 to mid-2025, but rose by about 44% between December 2025 and October 2026.

Over the same period, token prices fell by about 50% year on year (FourWeekMBA, 2026a).

This suggests that the value of the collateral has not collapsed, but that margins are narrowing for firms that rent hardware and sell tokens.

Price data have limited reliability. Different GPU price indices can show values up to eight times apart for the same chip (Implicator, 2026b), which makes it harder to verify the assumptions used to value collateral. Plans by CME and ICE to launch futures contracts on computing capacity could provide a more transparent price reference over time.

3.3Slow returns for enterprise users

Revenues at the end of the chain depend on the returns that enterprise users obtain from AI.

In a McKinsey survey of 1 719 executives, 89% of organisations use AI, but only 37% attribute any impact on earnings before interest and taxes (EBIT) to it (Figure 3.3).

Figure 3.3Respondents reporting higher individual productivity

Figure 3.3Organisations scaling AI across the enterprise

Figure 3.3Organisations attributing any EBIT impact to AI

The share of organisations attributing at least 5% of EBIT to AI stood at 6%, unchanged since 2025 (McKinsey, 2026). Users that cannot show a return may reduce spending or move to cheaper models.

4.1Borrowing costs are diverging

Borrowing costs within the same ecosystem differ widely depending on who stands behind the debt (Figure 4.1).

Debt backed by an investment-grade supplier is priced at about 5.75%, while junior tranches stand at 8.5-8.75%.

The cost of unsecured neocloud debt rose from 9.625% in June 2026 to an implied 15-16% in October 2026 (Crypto Briefing, 2026).

4.2Developments in October and November 2026

Events since mid-October 2026 show the risk moving from equity to credit markets (Figure 4.2).

The IMF identified an AI-related market correction as one of three salient risks to global financial stability (IMF, 2026).

Figure 4.214 OctImplied yields on CoreWeave notes maturing in 2030-2032 approach 15-16%

Figure 4.230 OctNasdaq falls 2.9%; S&P Global revises Meta’s outlook to negative

At the end of October, CoreWeave, Nebius and Nscale were effectively shut out of the bond market, and their shares fell by at least 30% during the month (WSJ, 2026a; CNBC, 2026b). In private funding markets, OpenAI raised about USD 11 billion against a target of USD 30 billion (WSJ, 2026b).

Figure 4.22 NovAnthropic’s amended S-1 discloses about USD 895 billion in commitments

Figure 4.24 NovOECD Pensions Outlook 2026: pension assets reach USD 78 trillion at end-2025

In the first week of November, ING offered about USD 500 million of data centre loans for sale. This illustrates a tendency for banks to transfer this risk to private credit funds, asset managers and insurers (CNBC, 2026a). Expectations that some US states may tighten permitting for data centre construction could also affect the financing of new projects (CNBC, 2026b).

5.1A rapidly growing role for private credit

Private credit has become a main source of funding for AI infrastructure.

According to the BIS, annual lending by private credit funds to AI-related firms rose from about USD 3 billion in 2010 to more than USD 40 billion in 2025 (Figure 5.1).

The outstanding stock exceeds USD 200 billion, and the share of AI-related loans in private credit has risen from less than 1% to about 8%. This stock could reach USD 300-600 billion by 2030 (BIS, 2026a).

The direct role of banks has remained relatively limited, in part because capital rules may assign higher risk weights to fast-depreciating hardware collateral. However, indirect bank exposure is rising. The IMF reports that loans from US banks to non-bank financial institutions increased by 18% in 2026, and their exposures to private equity and private credit firms by 30% (IMF, 2026).

5.2Insurers and pension funds

Long-term investors such as insurers and pension funds are among the final holders of a large part of this debt.

Private credit investments of EU insurers reached EUR 514 billion at the end of 2024, or 5.1% of their assets, while occupational pension funds held EUR 128 billion, or 4.4% (EIOPA, 2025). In the euro area, private credit accounts for 2.3% of assets for insurers, 1.4% for pension funds and 0.2% for banks (Figure 5.2, Panel A). An ECB stress exercise finds that insurers and pension funds would be the most affected sectors in a severe private credit shock (ECB, 2026).

Interest from these institutions continues to grow. In a 2026 survey, 57% of insurers plan to increase private credit allocations over the next 12 to 24 months, compared with 32% in 2024 (Mercer, 2026).

OECD data show that pension assets reached USD 78 trillion at the end of 2025, or 106% of the combined GDP of OECD countries (OECD, 2026d). At this scale, even a small portfolio share can represent a large amount.

6.1Off-balance-sheet obligations

Leases signed but not yet commenced appear in the notes rather than on the balance sheet, and the amounts are very large.

At 30 September 2026, these obligations exceeded USD 1.4 trillion in total at five companies alone (Figure 6.1).

Companies disclose the start dates, terms and cancellation conditions of these obligations in different ways, which makes it harder for investors to compare total leverage.

6.2The double exposure of financial institutions

Financial institutions are connected to the same small group of AI providers in two ways (Figure 6.2).

First, they use the models and cloud services of these firms. The OECD notes that supervisors see dependence on third-party AI providers that are outside the supervisory perimeter as a significant challenge (OECD, 2026b).

Second, they hold the bonds, loans and securitised debt of the same firms, directly or through funds.

These two exposures are usually monitored by different units and under different frameworks.

In early 2026, the ECB started to review both bank lending to the AI sector and banks’ use of GenAI, which shows that the two dimensions can be considered together (Bloomberg, 2026).

Financial distress at one provider could lead at the same time to service disruption and to credit losses.

Taken together, the five risk channels and five policy tools show that disclosure and stress testing have the widest coverage (Figure 7.1).

The considerations below are intended for discussion in the OECD Committee on Financial Markets.