The AI Buildout Is Real. The Credit Structure Is the Risk.

The AI buildout is both a technology cycle and a credit cycle. Physical power and construction constraints support real demand through at least 2028, while the largest hyperscalers carry investment-grade cash flow independent of AI. The risk lies in the financing wrapper: it converts gradual demand normalization into discrete, correlated credit events.

The Tell Is the Plumbing

Roughly $3 trillion of data-center capital expenditure is required through 2028. Hyperscaler operating cash flow covers about $1.4 trillion, leaving more than $1.5 trillion for credit markets. AI-linked debt issuance could approach $570 billion in 2026, more than twice the 2025 total.

The marginal financing dollar is moving from public investment-grade bonds toward private credit, off-balance-sheet ventures, and securitization. Fixed-income investors are therefore accumulating correlated exposure to sustained AI-compute demand.

Each financing layer transfers risk away from hyperscaler equity and toward creditors underwriting GPU utilization, contract renewal, and residual hardware value.

The Offtaker Is the Credit Instrument

CoreWeave’s DDTL 4.0 and DDTL 5.0 isolate the decisive variable: offtaker credit quality.

Otherwise similar structures produced a four-notch rating difference and a 225-basis-point pricing gap. The hardware provides a recovery floor; the customer contract determines the credit quality.

An investment-grade take-or-pay commitment can make GPU infrastructure resemble investment-grade project debt. Replace that commitment with non-investment-grade customers and the same collateral becomes high-yield.

How Demand Risk Transfers

Each step increases the distance between the underlying asset and the hyperscaler’s audited balance sheet. Investors several layers removed may have limited visibility into utilization, renewal probability, or residual GPU value.

The systemic problem is not merely debt volume. It is opacity combined with concentrated exposure to the same offtakers and operating assumptions.

Utilization Is the Load-Bearing Assumption

Estimated breakeven utilization for a fully costed, debt-financed GPU cluster is approximately 55–70%.

Below roughly 60%, a representative 1,024-GPU H100 cluster may lose about $330,000 per month. Above 80–85%, it can produce comparable positive cash flow because revenue scales more directly with utilization than operating costs.

Bare-metal GPU rental margins may reach 55–65% before depreciation, but fully loaded neocloud margins fall to roughly 14–16%. A modest utilization shortfall can therefore turn a facility cash-flow negative.

Debt-service coverage compounds the sensitivity:

A 1.35x DSCR requires operating income to remain 35% above debt service. That covenant embeds assumptions about utilization, rental rates, contract durability, and hardware life.

GPU depreciation remains unresolved. Most operators use five- to six-year accounting lives. Critics argue that annual architecture refreshes imply two to three years of economic life. Counterevidence suggests older A100 GPUs can retain strong contract rates by shifting from training to inference.

Both outcomes may exist across different hardware cohorts. That variability is precisely the credit risk: depreciation is an underwriting assumption, not a settled fact.

Where Cracks Appear First

CDS is the most liquid public signal. Oracle’s five-year CDS reportedly rose from about 75bp to more than 200bp between early and mid-2026. Sustained widening in Oracle and Nvidia spreads, independent of company-specific news, would indicate broader repricing of AI credit risk.

SRT pricing may move earlier. Those investors receive loan-level data and directly absorb tail risk, making spread widening a cleaner warning than a later rating action or default.

The Bull Case

Physical capacity remains scarce. Many announced projects lack secured power, financing, or construction progress. Project filtering suggests only about 54% of announced gigawatts may be delivered by 2030, with deliveries peaking around 2027.

Power, permitting, and supply-chain constraints limit actual construction. Scarcity supports utilization and pricing for completed assets even if announced projects are canceled.

The strongest offtakers also have diversified cash flow. Meta, Microsoft, and Amazon contracts are materially different credits from commitments by unrated AI laboratories.

High-quality underwriting still shows discipline. Power access, permitting, counterparties, and supply-chain execution constrain loan sizing rather than merely decorating credit memoranda.

The Decoupling Is the Risk

This is not the telecom cycle repeated mechanically. AI demand exists, capacity is physically constrained, and major offtakers possess real cash flow.

But the financing cycle has developed its own failure mechanics:

1. SPVs keep debt away from consolidated credit metrics.
2. Offtaker substitution can turn GPU debt from high-yield into investment-grade without improving hardware economics.
3. SRTs transfer concentrated AI exposure from banks to insurers and pension funds.
4. Continued lending supports the revenue assumptions used to justify additional lending.

A technology cycle can absorb weaker demand through lower prices, reduced utilization, and equity losses. Leveraged finance cannot adjust as smoothly. DSCR covenants, maturity mismatches, rating thresholds, and residual-value guarantees transform gradual deterioration into covenant breaches, forced sales, refinancing failures, and appraisal disputes.

The practical monitoring framework is therefore simple:

The technology thesis can remain correct while the credit structure fails. Utilization, coverage, contract quality, and collateral life will reveal that failure before the headline default.