Bond traders published the number this week that production architects should have had months ago: roughly $70 billion in contingent liabilities sit behind the AI API market, unrecorded on any vendor balance sheet and undisclosed to enterprise customers. Bloomberg's August 15 investigation identified these as "residual value guarantee" backstops—off-balance-sheet pledges made by chipmakers and hyperscalers to cover AI model companies' compute bills if they stop paying.

A special-purpose vehicle borrows money to purchase chips, secured by cashflow from a compute contract with an AI company. If the AI firm stops paying, the assets get re-leased or liquidated. Any shortfall triggers the backstop: the chip vendor or big-tech co-signer covers the gap. Meta built the template for data centers with a ~$27 billion deal ("Beignet") for its Hyperion campus in rural Louisiana and a ~$13 billion deal ("Sopaipilla") in Texas. Both structures engage residual value support only if Meta exits early. Meta's SEC filing states: "RVG payments are not probable, and therefore no liability has been recorded to date."

SPV / Residual Value Guarantee structure: an SPV buys chips with lender capital and leases compute to an AI company; the backstop provider covers any default shortfall.
FIG. 02 SPV / Residual Value Guarantee structure: an SPV buys chips with lender capital and leases compute to an AI company; the backstop provider covers any default shortfall. — Bloomberg, August 15 investigation; article body

Broadcom extended this model to chip financing with Project Big Sky—a $35 billion deal where Apollo Global Management and Blackstone financed custom AI chips leased to Anthropic, with Broadcom providing the backstop. Chip financings amortize in roughly five years to match hardware depreciation. Jensen Huang announced this week that Nvidia will extend residual value support for up to 25% of individual opportunities, assessed case by case, as part of a $500 billion financing partnership. CreditSights: Nvidia is "writing a put." Their assessment: "This is pro-cyclical and exacerbates boom-bust potential. The guarantee is nearly costless in the boom phase, but becomes most relevant in a severe, abrupt downturn, if/when customers are defaulting and market value for hardware is falling."

Deal / ProjectBackstop ProviderLenders / PartnersDeal SizeAsset TypeAI Beneficiary
BeignetMeta~$27 billionHyperion data center campus, LouisianaMeta
SopaipillaMeta~$13 billionData center, TexasMeta
Project Big SkyBroadcomApollo Global Management, Blackstone$35 billionCustom AI chips (~5-yr amortization)Anthropic
Nvidia RVG PartnershipNvidia (up to 25% per deal)$500 billion (financing program)Nvidia AI chipsVarious customers
FIG. 03 Major residual value guarantee (RVG) deals identified in the AI compute financing market — Bloomberg; article body (b1, b2)

Architects now face counterparties underwriting their inference stack that are exposed to the credit quality of unrated model companies. Moody's flagged $1.2 trillion in 2026 datacenter commitments and $460 billion in direct debt across six hyperscalers, expecting spend to reach $1 trillion in 2027. A large share traces to OpenAI and Anthropic. S&P downgraded Oracle to BBB- (one notch above junk) because roughly half of Oracle's $638 billion in remaining performance obligations comes from OpenAI. S&P warned: "If OpenAI were unable to pay Oracle, Oracle could be left with massive data center leases it might be unable to exit or have to re-lease under less-favorable terms." A July 2026 Nikkei Asia investigation of SEC filings found Oracle's off-balance-sheet commitments grew more than 2,900% in four years—to roughly $273 billion—as the company raced to build Stargate infrastructure for OpenAI. That figure now exceeds Oracle's reported on-balance-sheet debt.

EntityMetricValueNotes
AI API market (aggregate)Estimated off-balance-sheet contingent liabilities~$70 billionBloomberg, August 15
Datacenter sector (6 hyperscalers)2026 capex commitments$1.2 trillionMoody's warning
Datacenter sector (6 hyperscalers)Direct on-balance-sheet debt$460 billionMoody's, 2026
Datacenter sector (projected)2027 spend~$1 trillionMoody's forecast
OracleRemaining performance obligations$638 billion~half attributable to OpenAI
OracleOff-balance-sheet commitments~$273 billion2,900%+ growth over 4 years (Nikkei Asia, July 2026)
OracleS&P credit ratingBBB− (one notch above junk)Downgraded due to OpenAI concentration risk
OpenAIReported debt$0As of Q1 2026
OpenAIQ1 2026 capex (buildings & equipment)$46 millionOn-balance-sheet
OpenAIReported purchase commitments$665 billionLives in SEC filing footnotes
FIG. 04 Key AI-related financial exposure figures across major entities, 2026 — Moody's; S&P; Nikkei Asia, July 2026; SEC filings; article body (b3, b4)

H100 compute rates fell from roughly $8 per hour in early 2024 to $2–$4 per hour by late 2025, per Silicon Data's H100 Rental Index and Introl market data—a 64% decline. The neocloud 1-year contract price troughed at roughly $1.70 per hour in late 2024 before recovering, per SemiAnalysis. That trough was partly funded by a subsidy embedded in this credit structure: the gap between what AI APIs cost to run and what developers actually pay is financed by off-balance-sheet vehicles. OpenAI illustrates the asymmetry: $0 in reported debt and $46 million in Q1 2026 capex on buildings and equipment, against $665 billion in reported purchase commitments. That number lives in the footnotes.

H100 compute rates collapsed ~64% from ~$8/hr (early 2024) to $2–$4/hr (late 2025); the neocloud 1-year contract price troughed at ~$1.70/hr in late 2024.
FIG. 05 H100 compute rates collapsed ~64% from ~$8/hr (early 2024) to $2–$4/hr (late 2025); the neocloud 1-year contract price troughed at ~$1.70/hr in late 2024. — Silicon Data H100 Rental Index; Introl market data; SemiAnalysis

None of this is disclosed to API customers. Fitch issued a credit risk warning on July 27; the NAIC opened a review of data center debt sitting on insurance balance sheets. Rating agencies are beginning to factor off-balance-sheet exposure into their analysis. Architects will learn about credit stress the same way they learn about outages—after the fact.

Treat your top-three API providers as counterparties with undisclosed leverage. Build provider-abstraction layers before you need them. Model a 2–3x cost increase scenario on your highest-volume inference routes as a planning assumption, not a tail risk.