OCC Warns Banks Financing AI Data Centers Face Credit Risks From Opaque Deals and Rapid Depreciation

The Office of the Comptroller of the Currency has cautioned banks providing financing for AI data center construction that they face a category of credit risk the existing lending framework was not designed to handle — including opaque deal structures, rapid GPU hardware depreciation, and growing community opposition to data center siting that can halt or delay projects mid-construction. The warning is the clearest signal yet from federal regulators that AI infrastructure lending requires its own specialized risk frameworks, separate from standard commercial real estate or infrastructure lending.

The timing is not coincidental. Three of the world's largest banks have recently purchased credit insurance to hedge exposure to AI infrastructure loans — an unusual step that signals internal risk managers have concluded that standard credit assessment tools are inadequate for this asset class.

What the OCC Is Warning About

The OCC's concerns cluster around three structural problems in the AI data center lending market:

Opaque funding structures. Many AI data center projects are financed through complex multi-party arrangements — involving hyperscaler commitments, real estate developers, independent power producers, and specialized AI infrastructure operators — where the creditworthiness of any single counterparty is difficult to assess independently. Banks extending loans to these structures may hold collateral that is difficult to value and difficult to liquidate if a project fails mid-construction.

Rapid GPU depreciation. The hardware at the heart of AI data centers — NVIDIA's H100 and H200 GPU clusters, with B100 and B200 deployments expanding — depreciates faster than conventional data center equipment. A GPU generation cycle now runs approximately 18 to 24 months. Banks that accept GPU infrastructure as collateral on five or seven-year loans may find the underlying asset worth a fraction of its appraised value before the loan matures. This is a genuinely novel risk: the collateral degrades faster than the debt.

Community opposition and siting risk. AI data centers require substantial power draw, water for cooling, and significant land footprint. Projects have faced organized local opposition in several US markets — from residential communities concerned about noise, water use, and transmission infrastructure — that has delayed or terminated projects after significant construction investment. Lenders that have disbursed funds against anticipated completion schedules can be exposed when siting disputes extend timelines by years.

The Credit Insurance Signal

The report that three major banks have purchased credit insurance against AI infrastructure exposure is the most concrete evidence that internal risk management teams are treating AI data center loans as a distinct risk category requiring hedging. Credit insurance — in which a third party agrees to cover losses if a borrower defaults — is typically used when lenders hold concentrations of risk they cannot easily diversify. Its use here implies that at least some banks' AI lending exposure is concentrated enough to warrant the cost of insurance.

Credit insurance also has a secondary effect: it surfaces risk pricing that the broader market can observe. If insurance premiums for AI data center loan portfolios rise over the next 12 months, that becomes a leading indicator of deteriorating confidence in the asset class — information that affects lending terms and project economics across the sector.

Why Banks Are Still Lending

Despite these risks, lending to AI data center construction has expanded significantly. The economics are clear: AI infrastructure buildout is one of the highest-capital-expenditure segments of the current economy, with committed spending from hyperscalers running into the hundreds of billions annually. Banks that sit out the lending opportunity cede significant revenue to competitors.

The OCC's intervention is not a prohibition — it is a signal that banks need to build specialized underwriting capacity for this asset class rather than extending standard commercial real estate or infrastructure lending frameworks without modification. That means assessing GPU depreciation schedules explicitly, understanding the contractual structures behind hyperscaler commitments, and evaluating siting risk on a project-by-project basis rather than relying on general market assumptions about data center development.

The Broader Regulatory Picture

The OCC's warning lands as AI infrastructure spending accelerates globally. European banking regulators have begun similar conversations about concentration risk in AI-related lending, and the Basel Committee on Banking Supervision has flagged emerging technology asset classes as an area requiring closer attention in future framework updates.

For US banks specifically, the OCC guidance creates an expectation — not yet a requirement — that AI data center loan portfolios receive enhanced scrutiny. Banks that have been applying standard underwriting criteria to these loans may face examiner questions about how they accounted for GPU depreciation schedules and siting risk. Getting ahead of that conversation is easier now than during an examination.

What to Watch

The OCC is expected to follow this guidance with more formal supervisory expectations as AI infrastructure lending continues to grow. For banks already holding significant AI data center loan portfolios, the immediate task is model validation: stress-testing those portfolios against scenarios involving GPU depreciation ahead of schedule, construction delays from community opposition, and hyperscaler commitments that do not materialize at the expected scale. The lenders best positioned for this environment are those who build internal expertise to underwrite AI infrastructure as its own asset class before regulators make that a formal requirement.