OpenAI’s credit rating could end Nvidia’s guarantee

AI finance · Credit ratings · Guarantees

An OpenAI subsidiary is due to lease a future data-centre campus in Ohio. Nvidia has agreed to cover certain losses if that company stops paying, subject to the contract’s conditions. ^[1] That protection ends if the subsidiary or its parent obtains the credit rating required by the contract. A credit rating assesses a company’s ability to pay its debts. Nvidia has kept the required rating confidential: the public document shows [***] in its place. ^[2][5]

That gives an unusually concrete meaning to a report published by the Financial Times on 8 September 2026. Bankers advising OpenAI and Anthropic are seeking investment-grade ratings after the companies’ potential initial public offerings, arguing that fresh capital would strengthen their finances. The report describes discussions, not rating decisions. ^[4]

Better credit can mean cheaper borrowing. Under some contracts, it can also allow another company to withdraw its financial support. The question is whether the stronger borrower makes up for the departing guarantor. The Nvidia–OpenAI arrangement provides public documents with which to examine that trade-off, without pretending to know what the rating agencies will decide.

What the $105 billion actually covers

On 17 August 2026, Nvidia announced guarantees connected to leases at SB Energy’s PORTS Technology Campus in Ohio. An affiliate of OpenAI Group PBC is to be the tenant. The initial commitment concerns approximately 4.25 gigawatts of IT load and carries a cumulative $105 billion payment cap. That is neither a loan advanced to OpenAI nor an estimate of Nvidia’s eventual loss. ^[1]

Nvidia’s quarterly report for the period ended 26 July 2026 describes the subsequent August transaction; the reporting date is not the signing date. It sets out nine construction phases, with guarantees generally becoming effective when the relevant leases start. Coverage is limited to defined lease and power obligations. Initial service is expected from Nvidia’s fiscal 2029, consistent with the 2028 start described in the 8-K. ^[1][3]

The arrangement is a residual value guarantee. Following a covered default, Nvidia has options that include taking over the lease, seeking a replacement tenant or initiating a sale. Its payment is generally described as the shortfall between a guaranteed minimum value and proceeds recovered through a replacement lease or disposal. Some detailed terms are withheld. ^[1][2]

The sequence matters. A phase has to reach the relevant stage; the default must satisfy the agreement; and recoveries affect the outcome. Adding the whole cap to Nvidia’s funded debt would discard those conditions. Treating the guarantee as irrelevant merely because no payment is immediately due would be no more useful.

A rating can change the landlord’s protection

The SEC exhibit is headed Form of Residual Value Guaranty. It is a template with blanks and redactions, not a fully completed public copy of every executed lease package. Section 13(a)(iv) provides for automatic termination when the tenant or its parent reaches the undisclosed rating. Another provision allows certain replacement credit-support instruments, subject to conditions that are also partly confidential. ^[2]

The missing threshold cannot simply be filled in as “BBB−”. That is the familiar boundary for investment grade on some rating scales, including the one illustrated by the SEC. It tells us nothing about the threshold negotiated for this contract. A rating is an assessment of credit risk, not a promise by the agency to pay the creditor. ^[5]

An IPO and the release trigger are therefore separate events. The visible clause makes the rating the trigger. Going public might help the company achieve the rating. The document does not establish that an IPO is either necessary or sufficient.

For the landlord, the practical question is who remains liable after the release. The tenant’s own obligations do not disappear when the guarantor’s end. But an additional source of recovery is removed on the contract’s terms. The tenant’s financial strength, the protections still in place and the property’s value then have to be considered together.

The landlord should not be confused with investors in bonds that OpenAI might issue directly. Lending to the company and funding a vehicle that owns a campus do not automatically confer the same rights. We have not reconstructed the collateral, ranking and covenants of every financing associated with the Ohio project. Describing Nvidia’s release as an immediate transfer of $105 billion to pension funds would invent the missing links.

The support also serves Nvidia’s business

There is an explicit commercial exchange. Nvidia’s quarterly report says the campus will exclusively host its AI infrastructure, subject to limited exceptions. The company also explains that financing constraints can hold back some customers’ deployments. Supporting credit can therefore enable demand for the supplier’s own products. ^[3]

This is not insurance supplied by a party with no interest in the underlying trade. It connects the customer’s ability to build with the manufacturer’s interest in selling. That does not establish fictitious revenue or concealed debt. It identifies a dependence worth testing: genuine demand for computing may still need financial assistance from the vendor.

The 8-K also says OpenAI has agreed to reimburse and indemnify Nvidia for amounts actually paid to the landlord. A reimbursement claim is a contractual right, not cash already in hand. If Nvidia has to pay because the customer is in financial difficulty, collecting from that same customer becomes a separate question. ^[1]

More independent access to funding could consequently benefit both companies. It would provide the customer with resources and might remove an obligation from the supplier. Their interests can align without proving that rating agencies have accepted their bankers’ case.

S&P already counts the guarantee

On 18 August 2026, the day after the announcement, S&P Global Ratings affirmed Nvidia’s AA rating with a stable outlook. It nevertheless projected a guarantee-related debt adjustment of $4.2 billion in 2028, rising to approximately $37.7 billion in 2031, after tax. The calculation incorporates the guaranteed minimum value and the agency’s own assessment of property recoveries. These are analytical projections, not recorded losses. ^[6]

S&P’s favourable conclusion also depends on the cash generation it expects from Nvidia. The support is therefore not being treated as costless to credit quality. In the agency’s scenario, the guarantor has the financial capacity to carry it. ^[6]

If the guarantee terminates, the implications for Nvidia and for the beneficiary’s financing may differ. The supplier sheds an obligation. The landlord loses support but may have a better tenant. Calling the development an “upgrade” does not, by itself, settle the combined effect.

How much IPO cash reaches the company

The bankers’ argument deserves a fair hearing. An equity raise that brings in usable cash can extend a company’s ability to meet expenses and absorb losses ahead of its creditors. Credit quality can improve before a business becomes profitable. The first task is to calculate how much of the offering actually reaches it.

An IPO may contain newly issued shares, which fund the company, and shares sold by existing owners, whose proceeds go to those sellers. The SEC makes that distinction explicit. The market valuation of the whole company is a third, separate number. ^[8]

Consider a wholly fictional $50 billion offering. Suppose $10 billion goes to existing shareholders selling their stock and the company bears $1 billion of issuance expenses. The issuer receives $39 billion net. The $50 billion describes the transaction; $39 billion describes the net addition to company cash. These assumptions are not estimates for an OpenAI or Anthropic IPO.

The same care is needed before a listing. On 31 March 2026, OpenAI announced a financing round involving $122 billion of committed capital, at a $852 billion post-money valuation. That announcement is not a cash statement as of 9 September. Establishing available liquidity would require the payments received, any remaining conditions and the spending since the round. ^[7]

Even a large cash inflow must be matched against its uses. Money raised to meet already agreed investment commitments may not be a free reserve available to repay bondholders. The answer depends on when payments fall due and whether spending can genuinely be deferred.

The payment schedule matters as much as the raise

S&P’s corporate liquidity methodology compares sources and uses of cash. “Adequate” liquidity is expected to withstand adverse circumstances over the following twelve months and is described as ratings-neutral, rather than an automatic enhancement. The framework also distinguishes committed investment from some growth projects that are easier to curtail. ^[9]

For an AI lab, that means building a period-by-period cash budget after the proposed transaction. Compute payments, rent and other expenses belong in it, alongside sufficiently predictable receipts and spending that can actually be postponed. We do not have the consolidated schedule needed to assign our own rating to either OpenAI or Anthropic.

An annualised revenue figure cannot fill that gap. Multiplying one month’s sales by twelve indicates a pace. It does not produce twelve months of recognised revenue, still less twelve months of net cash receipts. A creditor needs to understand the margin after providing the service and the timing of cash inflows relative to payment obligations.

Suppose capacity costs are partly fixed while customers pay according to usage. Rising utilisation could improve coverage of those costs. Lower prices or a delayed rollout could instead reduce receipts without cutting the payment schedule proportionately. This is a transmission scenario, not a measurement of the two labs’ current finances.

S&P’s broader corporate framework also examines competitive position, business risk and cash flow relative to leverage. The amount raised cannot substitute for those other dimensions. ^[10]

Anthropic’s financing follows different terms

On 9 June 2026, Broadcom, Apollo and Blackstone announced a financing platform whose initial $35 billion transaction was intended to support more than one gigawatt of additional capacity for Anthropic. That is the transaction’s announced size, not a full Broadcom guarantee. ^[11]

In its 10-Q signed that day, covering the period ended 3 May, Broadcom describes an arrangement entered into on 8 June. An investment partner takes over AI-rack purchase agreements and related leases with a customer that is not named in the passage. Broadcom provides a backstop for five-year lease obligations, with maximum exposure of $29 billion. It grows as equipment is deployed and falls as the customer pays. Remedies include assuming the lease or selling the racks. ^[12]

The dates and subject matter allow the disclosures to be read alongside each other, not merged into a single undifferentiated number. Nor do they establish that Broadcom has the same rating-based release as Nvidia. We have not identified such a clause in these documents. The June figures are not being presented as a measurement of net exposure in September.

The useful comparison concerns the industrial role. A manufacturer helps fund infrastructure that uses its components. The value of the equipment, its ability to be redeployed and the customer’s finances then matter to both the supplier and outside investors.

Different creditors can depend on the same cash flow

Research published by the Bank for International Settlements describes vehicles that borrow to fund data centres while technology companies commit to leases or capacity purchases. The vehicle’s borrowing is separate from its customer’s balance sheet, but debt service depends on payments from that customer. The authors identify connections to private-credit funds, insurers and banks. Their general analysis is not an audit of the Ohio campus. ^[13]

Avoiding double counting requires following the same dollar through the contracts. A customer payment becomes the property owner’s revenue; some of that cash may then pay lenders. A guarantee adds a possible recovery route if the chain fails. Summing debt, rent commitments and the guarantee cap without reconciling them could count overlapping economic exposures more than once.

Dependence can still be concentrated. If final demand for computing disappoints, a customer might struggle to pay, a building might be harder to relet and a chipmaker might face weaker sales. These difficulties could reinforce one another rather than behave as three independent risks. The public information reviewed here does not establish their joint probability.

A rating does not eliminate this analysis. It concerns a defined issuer or instrument, at a particular time, under stated assumptions and protections. As the SEC explains, it does not cover every market or liquidity risk of a security. A bond can lose value before any default, including when the yield investors demand rises. ^[5]

Assess the financing after the guarantor leaves

The question for the agencies is specific: which protections do their scenarios retain once the release conditions are met? It would be inconsistent to count continuing support that disappears within the very scenario being assessed. We have found no evidence that agencies are doing so. The uncertainty calls for disclosure, not an allegation of accommodating ratings.

A sound outcome remains possible. The company receives enough equity, cash flows become more predictable and payment coverage improves. Initial vendor support has helped it graduate to more independent funding. Removing the guarantee need not weaken the project if the stronger tenant compensates for its loss.

A less favourable outcome would be a mainly temporary liquidity improvement, followed by rapid spending, after a creditor has already lost a recovery route. Distinguishing between the two requires net issuance proceeds, payment commitments, rating analyses and the protections that apply after termination. The treatment of a later downgrade also needs checking: the visible clause does not reconstruct every other agreement that might be relevant.

The redacted threshold prevents a prediction of Nvidia’s release. The published clause is enough to establish that a rating can change who remains contractually responsible. A credit opinion can become a contractual event. What matters is the borrower’s ability to pay and the remedies left if the opinion proves too optimistic.

---

Scope. Documentary analysis with an evidence cut-off of 9 September 2026. Rating discussions are attributed to news reporting; no future rating is assumed. Company figures are attributed to their disclosures. The guarantee template is partly redacted. S&P projections and the fictional IPO calculation are separate from observed facts. No interviews, recoveries or legal audit of the entire contractual package are claimed.

Related l0g reporting: the $500 billion financing announcement and how data-centre debt reaches investment funds.

Sources and documents

1. NVIDIA / SEC EDGAR · 2026-08-17. Form 8-K · 17 August 2026. Item 1.01, Residual Value Guaranties ; Item 2.03. Initial arrangement, tenant, $105bn cap, 4.25 GW, service conditions, default remedies and OpenAI reimbursement undertaking. First service is expected from 2028, not already achieved.
2. NVIDIA / SEC EDGAR. Exhibit 10.1 · Form of Residual Value Guaranty. Préambule A–F ; §§ 1(d), 2, 12 et 13(a)(iii)–(iv). Section 13(a)(iv) identifies a tenant or parent rating trigger. The required rating is redacted. This is a template, not the entire executed package. The terms of any subsequent reinstatement have not been established.
3. NVIDIA / SEC EDGAR. Form 10-Q · quarter ended 26 July 2026. Note 10, Guarantees / SB Energy Corp. guarantees (p. 19) ; facteurs de risque, land, power, and shell commitments. Sets out nine phases, the exposure profile, covered obligations and the commercial exchange. The August guarantees must not be described as already contracted at the July reporting date.
4. Financial Times; attributed republication in Folha; Reuters · 2026-09-08. Anthropic and OpenAI bankers push for top-tier credit ratings post-IPO. FT reporting on bankers’ discussions with rating agencies. The original requires a subscription; the Folha republication credited to the Financial Times was consulted. A Reuters dispatch distributed by Boursorama also relays the story and states it could not independently verify it. These republications are not independent confirmations. No assigned rating or future decision is inferred from the report.
5. SEC · Investor.gov · 2017-10-12. Updated Investor Bulletin: The ABCs of Credit Ratings. What is a credit rating? ; What a credit rating is not. Definition, BBB− threshold example, issuer/instrument distinction and limitations. This rating-scale boundary does not disclose Nvidia’s contractual threshold.
6. S&P Global Ratings · 2026-08-18. Research Update: Nvidia Corp. To Provide Financial Support For SB Energy AI Infrastructure Campus Buildout; Ratings Affirmed. Rating Action Overview ; paragraphe « In our view, the RVG is potentially a significant liability for NVIDIA » ; Ratings List. AA/stable affirmed; tax-affected debt adjustments of $4.2bn in 2028 and approximately $37.7bn in 2031. These are not expected or recorded losses. The two cited years do not provide a complete annual time series.
7. OpenAI · 2026-03-31. OpenAI raises $122 billion to accelerate the next phase of AI. Premier paragraphe. $122bn of committed capital and an $852bn post-money valuation. Neither measures independently verified cash available in September.
8. SEC · Investor.gov. Updated Investor Bulletin: Investing in an IPO. Selling shareholders. Proceeds from existing-share sales go to sellers, not the issuer. The 50 − 10 − 1 = 39 example is fictional teaching arithmetic, not SEC data or an IPO forecast.
9. S&P Global Ratings · 2014-12-16. Methodology And Assumptions: Liquidity Descriptors For Global Corporate Issuers. §§ 9–11, 31, 37, 58–60. Liquidity sources/uses, ratings-neutral adequate liquidity, committed versus some reducible investment. No numerical rating test is applied to either lab in the absence of data.
10. S&P Global Ratings · 2024-01-07. Criteria | Corporates | General: Corporate Methodology. Paragraphs 3–7 and 11–16. Business risk, competitive position, cash flow relative to leverage and modifiers including liquidity. General corporate criteria, distinct from project-finance methodology; they assign no rating to OpenAI or Anthropic.
11. Broadcom, Apollo et Blackstone · 2026-06-09. Broadcom, Apollo, and Blackstone Establish Landmark Strategic Platform to Accelerate More Than 20 Gigawatts of Global AI Deployments. Titre secondaire et premiers paragraphes. Announced $35bn initial transaction supporting more than 1 GW for Anthropic. That figure is not the maximum backstop exposure in the 10-Q.
12. Broadcom / SEC EDGAR · 2026-06-09. Form 10-Q · quarter ended 3 May 2026. Note 11, Subsequent Events, p. 22 ; signature du 9 juin 2026. Five-year lease backstop, $29bn maximum, variable exposure and rack-related remedies. The customer is unnamed in this passage. A rating-triggered release is not established here, and current net exposure is not calculated.
13. Banque des règlements internationaux / Bank for International Settlements · 2026. Financing the AI infrastructure boom: on- and off-balance sheet borrowing. Paragraphes sur les véhicules, les loyers/capacity offtake et les liens avec les investisseurs non bancaires. General framework separating vehicle debt from customer payments. Authors state that their views need not represent the BIS or its member central banks. Not automatically applicable to PORTS.