AI Data Center Underwriting: Why Gross Compute Demand Fails Credit Screens

AI Data Center Underwriting: Why Gross Compute Demand Fails Credit Screens

Macro compute demand is a trap for project financiers. When HSBC Chief Executive Officer Georges Elhedery warned that artificial intelligence expansion will not turn every data center into a sound investment, the core flaw in market consensus is treating sector-wide compute growth as a blanket credit guarantee. Aggregate projections—such as McKinsey estimating up to $7 trillion in global data center spending by 2030 and JPMorgan projecting up to $4.1 trillion in AI-related debt issuance over the same horizon—measure capital appetite, not project-level solvency. A facility can sit inside a high-growth market and still destroy equity or trap debt holders if its lease duration lags debt amortization, its power tariff floats above contracted pass-throughs, or its tenant lacks investment-grade balance sheet durability.

A rigorous underwriting blueprint requires separating secular AI adoption from asset-level cash flow durability. First, lenders and equity sponsors must map how the underlying argument actually works and where it stops short. Second, they must evaluate facilities across four concrete risk dimensions: tenant creditworthiness and lease structure, power and thermal unit economics, capital expenditure timing versus debt service, and secondary liquidity or risk-transfer mechanics. Third, they must apply a disciplined credit screen before committing capital, while recognizing where models break down under hardware obsolescence and grid delays.

Deconstructing the Standard Caution

The public warning around indiscriminate data center lending centers on selectivity rather than skepticism toward artificial intelligence itself. Reported commentary from Elhedery treats AI as a genuine productivity driver while urging capital providers to examine tenancy agreements, contract pricing, and project financial strength. Alongside that caution, market observers have noted that large technology platforms have issued massive volumes of debt—rising from roughly $200 billion in bond financing in 2025 toward a projected $1 trillion pace in 2026—often behaving with relative insensitivity to yield and credit spreads because management views compute scale as existential.

Where standard commentary stops is on causal mechanics. Saying "contracts vary in quality" or "some projects face challenges" leaves three critical links unexamined:

  • Demand vs. Cash Flow Timing: Rapid token or training demand does not automatically convert into front-loaded free cash flow for a developer carrying construction debt before commissioning completes.
  • Borrower Spread Insensitivity vs. Asset Risk: A hyperscaler willing to pay wide spreads on corporate bonds to secure capacity is not identical to a merchant or specialized developer financing a single campus with project-level leverage.
  • Asset Durability vs. Hardware Churn: Physical shells, substations, and cooling plants last decades, whereas high-density racks and accelerator generations turn over on much shorter cycles, creating friction when lease terms and equipment useful lives diverge.

Distinguishing between known facts and educated hypotheses keeps underwriting grounded. Known facts include public statements urging selectivity on lease terms and pricing, third-party multi-trillion-dollar spending and debt projections through 2030, widening credit spreads on certain five-year single-A data center asset-backed securities from 1.95 percentage points in the first quarter of 2026 to an average of 2.57 percentage points over Treasurys over a subsequent three-month window, and bank efforts to transfer 20% to 50% of loan risk to insurers or sell troubled exposures at discounts. Educated hypotheses include how much merchant capacity will face re-leasing haircuts in the late 2020s, how fast grid interconnection queues will clear, and whether newer compute architectures will compress power density requirements per unit of output.

The Four Risk Dimensions of AI Infrastructure

Evaluating a data center as a standalone credit requires decomposing it into four interacting systems rather than pricing it off headline megawatts.

1. Tenant Quality and Contract Mechanics

In data center finance, lease structure is the bridge between digital demand and fixed-income repayment. A facility leased to a high-cash-flow cloud operator under a long-term, non-cancellable commitment behaves very differently from a facility reliant on short-term commitments, unhedged renewals, or early-stage AI laboratories whose revenue depends on continued equity or debt funding.

Lenders must look past headline lease length to test five contractual details:

  1. Termination and Force Majeure Clauses: Provisions allowing delays, payment deferrals, or walk-aways when permitting, substation delivery, or grid tie-ins slip shift schedule risk directly onto the sponsor.
  2. Evergreen vs. Cliff Renewal Risk: A ten- or fifteen-year initial term only protects long-dated debt if renewal mechanics or amortization schedules prevent a refinancing cliff when the initial lease expires.
  3. Counterparty Concentration: Single-tenant campuses concentrate renewal risk on one balance sheet, whereas multi-tenant facilities diversify credit exposure at the cost of higher leasing and downtime friction.
  4. Escalation vs. Inflation: Fixed rent escalators that trail operating cost inflation erode net operating income over time.
  5. Credit Support: Parent guarantees, letters of credit, or step-in rights determine whether a tenant's covenant holds when market pricing softens.

2. Power, Thermal Density, and Unit Economics

Data center economics are fundamentally energy conversion and heat removal problems governed by thermodynamics and utility tariffs. High-density AI clusters draw far higher rack power than legacy enterprise colocation, increasing demands on medium-voltage switchgear, liquid or hybrid cooling loops, and backup generation.

The cost function of operating the asset turns on three operational variables. First is power availability and lead time: securing firm megawatts and utility interconnection often takes years, meaning stranded capital sits idle if grid energization lags building completion. Second is tariff pass-through fidelity: if electricity rates spike or peak demand charges reset, any lag or cap in passing power costs to the tenant compresses operating margins. Third is power usage effectiveness and thermal headroom: higher ambient temperatures or suboptimal cooling designs force compressors and chillers to work harder, consuming power that cannot be billed as revenue-generating IT load.

3. Capital Expenditure Front-Loading and Debt Service

Construction outlays occur upfront—land, civil works, transformers, generators, and electrical distribution—while lease cash flows begin only after commissioning and tenant acceptance. When interest rates are elevated or credit spreads widen, debt service consumes operating cash flow earlier and leaves less margin for cost overruns or delayed ramp-up.

Consider a stylized developer financing a multi-phase campus with staged construction debt. Phase 1 delivers on schedule and begins collecting rent, but Phase 2 encounters a permitting delay or utility interconnection holdback. Even if long-term demand for AI compute remains intact, interest carries on drawn commitments while uncommissioned capacity generates zero rent, compressing debt service coverage ratios and forcing either sponsor equity injections or covenant waivers.

4. Liquidity, Syndication, and Risk Transfer

Many infrastructure lenders historically assumed deep secondary liquidity or smooth securitization exits for stabilized data center assets. When market sentiment turns cautious or spreads widen on asset-backed paper, exits narrow. Construction loans often stay on balance sheets until completion milestones or pooled sell-down restrictions expire, and secondary sales of troubled or delayed exposures can require discounts.

To manage balance sheet limits, banks increasingly use credit insurance or risk-transfer structures to pass a slice of exposure—often reported around 20% to 50% on comparable transactions—to insurers. Risk transfer reduces net retained loss on a given facility, yet it introduces counterparty dependence on the insurer and does not fix an underlying project whose cash flows fail to cover debt service.

Operational Underwriting Protocol

Practitioners underwriting or advising on data center debt and equity should enforce a sequential gate before approving term sheets or allocating capital:

  1. Audit Legal and Schedule Risk First: Review power purchase agreements, interconnection agreements, permitting status, and lease force majeure or termination triggers before modeling cash flows.
  2. Underwrite Downside Cash Flows: Stress-test debt service coverage against construction delays, slower tenant ramp-up, and partial re-leasing discounts rather than assuming uninterrupted full utilization.
  3. Verify Cost Pass-Throughs: Map every utility charge—energy, demand charges, transmission fees, and fuel—against lease reimbursement language to confirm no unrecovered leakage.
  4. Size Debt to Contracted Cash Flow, Not Speculative Upside: Cap leverage based on binding, credit-backed revenue streams rather than projected future market rents or uncontracted expansion phases.
  5. Define Exit and Transfer Mechanics Early: Establish clear covenants, reserve accounts, and syndication or risk-transfer options before closing so the lender is not trapped if market spreads move against the sector.

Where the Framework Has Limits

No underwriting checklist eliminates structural uncertainty in an infrastructure cycle moving at hardware speed. Three real-world limitations constrain any model:

  • Architectural Obsolescence: Rapid shifts in chip power draw, rack density, and cooling requirements can make older facility designs less attractive to frontier tenants even when the building itself is structurally sound.
  • Regulatory and Community Friction: Local opposition, water-use scrutiny, and grid allocation rules can stall or derail projects late in development regardless of financial modeling.
  • Opaque Private Contracts: Exact lease pricing, tenant credits, and side letters are frequently non-public, forcing outside analysts to infer risk from market pricing signals such as bond and securitization spreads.

Strategic Play for Capital Allocators

Treat data center financing as an asset-level credit and execution discipline rather than a thematic bet on artificial intelligence. Concentrating exposure in facilities backed by investment-grade covenants, firm grid delivery, clean pass-through mechanics, and conservative leverage—while walking away from construction-heavy or weakly contracted builds—is the most durable way to capture real infrastructure demand without inheriting avoidable credit losses.

AM

Amelia Miller

Amelia Miller has built a reputation for clear, engaging writing that transforms complex subjects into stories readers can connect with and understand.