There is a number that capital markets have not fully metabolized: $105 billion. That is the ceiling of credit support NVIDIA has extended to OpenAI for a single data center complex in Ohio, according to regulatory filings reported by the Financial Times. Pause on the structure before you accept the business-development narrative. NVIDIA is not merely selling accelerators to OpenAI; it is financing those accelerators. The chip supplier has become the banker. The vendor has become the creditor. And the entire guarantee evaporates only when OpenAI obtains what the agreement calls a “satisfactory credit rating.”
Three rating agencies now hold an operational veto over one of the largest hardware commitments in corporate history. This is not hyperbole; it is capital structure laid bare in a single clause. I have audited financial and cryptographic infrastructure for 27 years. The blockchain remembers; the architect forgets. Somewhere inside the Ohio project, an architect forgot that the supplier who finances the client becomes the first creditor in a liquidation. Yet nobody in this chain is pricing that priority correctly.
This story was never about model benchmarks. It is about who owns the liability when the AI narrative collides with a balance sheet that cannot yet produce free cash flow.
The backdrop is a coordinated campaign. OpenAI and Anthropic have both been walking into the offices of Moody’s, S&P Global, and Fitch, accompanied by Morgan Stanley and Goldman Sachs, seeking investment-grade status. The Financial Times reports that these conversations remain early and the rating agencies are not yet convinced. Effective speculative-grade treatment is still in force. Neither company has demonstrated sustained positive free cash flow. Neither has released the unit economics of a single inference call. And neither has told the market how much capital the next model generation will consume before revenues catch up.
Why push for a rating at all? Because the IPO that both companies are approaching is no longer merely an equity event. It is the doorway to the public bond market. Investment-grade status would permit pension funds, insurance companies, and other fiduciary capital to buy AI infrastructure debt at interest rates that equity financing cannot imitate. It would also terminate the NVIDIA guarantee, releasing OpenAI from a financial embrace that is becoming politically and economically inconvenient.
History offers harsh benchmarks. Meta, Netflix, and Tesla all waited more than a decade after their IPOs to obtain investment-grade ratings. SpaceX moved faster, but its upgrade rested on government contracts and predictable, contracted cash flow. AI labs have neither. They have subscription revenue from corporate clients, API usage fees, and an expense line defined by chip purchases measured in the hundreds of billions.
The industry frames this as a liquidity milestone. It is not. It is a solvency test conducted in public, with the entire AI supply chain watching.
Begin with the loan that is not called a loan. NVIDIA’s $105 billion credit support for OpenAI is supplier finance at industrial scale. The arrangement lets OpenAI build data centers and procure GPUs without drawing down equity or issuing traditional debt. In exchange, NVIDIA secures a captive buyer for its most advanced silicon across a multi-year construction program. The structure converts a chip sale into a recurring balance-sheet claim. If OpenAI thrives, NVIDIA books the orders and collects the interest or equivalent compensation. If OpenAI stumbles, NVIDIA absorbs a credit event that will ripple through its own earnings, inventory, and investor sentiment.
NVIDIA is not OpenAI’s partner. It is OpenAI’s lender of first resort. The distinction matters because lenders eventually demand covenants, reporting, and repayment. Suppliers demand orders. When one company is both supplier and lender, the conflict of interest is hidden in plain sight. NVIDIA’s incentive to keep shipping chips may override its incentive to enforce credit discipline. That is how supply chains generate hidden leverage: the party with the strongest commercial interest in the borrower’s expansion is also the party holding the credit risk.
This arrangement will remain in force until OpenAI secures a rating that releases NVIDIA from its exposure. The filing says the guarantee terminates upon a satisfactory credit rating. That clause is the single most important sentence in the modern AI capital stack. Until it triggers, OpenAI carries a contingent liability that no spreadsheet can fully model, because its size depends on NVIDIA’s willingness to continue financing rather than to demand payment.
Rating agencies are not evaluating intelligence. They are evaluating the capacity to repay debt from cash flow. AI laboratories have spent years selling a vision of superintelligence; the rating agencies are asking for something more mundane: positive free cash flow, customer concentration data, and the durability of revenue contracts.
OpenAI and Anthropic cannot yet show those numbers. Their revenues are growing, but their capital expenditures are growing faster. They are spending on data centers, land, power, and interconnection before they have proven that their customers will pay enough per token to amortize the fixed costs. This is the classic infrastructure paradox. A company that must build ahead of demand will always appear capital-intensive before it appears profitable. Rating agencies have seen this movie before, and they rarely offer investment-grade ratings during the construction phase.
The agency calculus is further complicated by customer concentration. OpenAI depends on Microsoft for distribution and compute. Anthropic depends on Amazon and Google for cloud capacity. Those relationships provide strategic sponsorship, but they also create a concentrated receivable base. If a single cloud partner changes its terms, the revenue shock is immediate. The agencies will look through the headline revenue and ask what portion comes from three or four hyperscalers that could build competing models tomorrow. Strategic sponsorship is not contractual revenue. It is goodwill with an invoice attached.
The comparison to SpaceX is the most dangerous precedent in the current narrative, because it is misleading. SpaceX obtained investment-grade treatment because its cash flows were anchored by government launch contracts with scheduled milestones and contractual penalties. The U.S. government is a slow payer, but it is a certain payer. OpenAI and Anthropic sell to enterprises that can cancel subscriptions, reduce API usage, or switch to open-weight models that cost less. There is no committed counterparty in the AI stack that resembles the federal government. There is only a voluntary flow of corporate experimentation budgets.
The market should also examine the historical pattern that the rating agencies have internalized. Meta, Netflix, and Tesla each required a decade or more after their IPOs to earn investment grade. All three survived near-death experiences during that period. All three saw their equity holders diluted and their debt holders protected. All three ultimately achieved upgrade because they converted narrative into recurring cash flow: advertising clicks, subscription fees, and vehicle deliveries. OpenAI and Anthropic have not yet demonstrated that their products command the same pricing power. The rating agencies are not being contrarian by holding them at speculative grade. They are being historically literate.
Now add the competitive dynamic. OpenAI and Anthropic are pursuing the same rating, through the same investment banks, at the same moment. The simultaneity is not a coincidence. Both teams know that the first laboratory to reach investment grade will obtain a structural cost-of-capital advantage that the other cannot easily offset. The first entrant will issue bonds to fund the next generation of compute, lock in supply agreements with chip vendors, and lengthen its cash runway while holding equity dilution down. The laggard will face higher interest costs, shorter supply visibility, and a harder conversation with its own board.
This is where the balance sheet becomes a competitive weapon. In the early years of AI development, leadership was defined by paper quality and benchmark scores. Then it shifted to revenue and market share. The next phase will be defined by interest expense. A 200-basis-point differential on $50 billion of debt is a $1 billion annual advantage. That is not a rounding error on a research budget; it is a decisive reallocation of resources toward compute, talent, or dividends.
There is a systemic dimension that the market has not priced. NVIDIA’s exposure to OpenAI is not isolated. If OpenAI’s debt becomes distressed, NVIDIA will face both a credit loss and an order book collapse. That dual hit would transform NVIDIA from a stable compounder into a cyclical credit story, and its own rating could come under pressure. The AI trade has therefore created a correlation that the rating agencies have not yet adjusted for: the largest chip supplier and its largest buyer are now financially inseparable. When investors buy NVIDIA equity, they are buying OpenAI’s credit risk. When they buy OpenAI’s future bonds, they are buying NVIDIA’s willingness to keep financing. This circularity is precisely the kind of hidden correlation that produces systemic surprises.
I have spent my career looking for the failure vectors that polite analysis ignores. Let me apply my standard pre-mortem. Three events would break this structure. First, a major cloud customer builds its own models and redirects compute budgets internally, starving OpenAI and Anthropic of revenue precisely when their data center debt matures. Second, a security or safety incident triggers regulatory intervention, slowing deployment and delaying the revenue curve by eighteen months. Third, NVIDIA itself resets its terms, demanding upfront payment for chips and pushing the financing burden back onto laboratories that cannot self-fund. Each vector is plausible. Combined, they constitute the tail risk that rating agencies are paid to price.
I also apply a sustainability stress test borrowed from my analysis of algorithmic stablecoins. In the Terra collapse, the model demanded infinite growth to survive. The AI infrastructure model is not identical, but it shares the same mathematical shape: break-even depends on continuous utilization growth at declining unit costs. If utilization stalls, the fixed costs do not disappear. They merely move across the income statement as depreciation, interest, and storage charges. A laboratory can survive a revenue miss. It cannot survive a utilization miss when its debt is structured for a seventy percent utilization forecast.
The bulls deserve a hearing, because the bear case is too comfortable. First, the strategic sponsorship argument has real weight. Microsoft, Amazon, and Google have existential reasons to keep OpenAI and Anthropic alive. The financing they have already committed resembles a government contract in everything except name. If the liquidity event occurs, those sponsors will likely convert credit exposure into equity or asset purchases, cushioning the bond market’s losses.
Second, the AI revenue cycle may be faster than historical precedent. These laboratories have crossed $1 billion in annualized revenue more quickly than most software companies in history. Their pricing power is unproven, but their top-line growth is real. If Europe and Asia fund sovereign AI champions, the data center operating rates will find a floor that private usage cannot supply.
Third, the rating agencies may be forced to adapt. If sovereign wealth funds and strategic buyers treat AI compute as infrastructure rather than optional consumption, the asset class will acquire the characteristics of a utility. Utilities receive investment-grade ratings because their cash flows are regulated and predictable. AI infrastructure may eventually earn the same treatment by government decree rather than commercial proof. That is the optimistic path, and it cannot be dismissed.
Even on the optimistic path, the harder question remains: what happens when the capital markets discover that two companies with speculative-grade balance sheets are building the largest physical infrastructure expenditures since the construction of the interstate highway system? The moment the first downgrade arrives, the systemic risk will not stay inside the AI sector. It will spill into cloud providers, semiconductor vendors, power utilities, and asset-backed securitizations that package data center debt for yield-seeking institutions.
A final note on process. The financing game that OpenAI and Anthropic are playing is a legitimate response to a brutal capital cycle. They have no choice but to pursue investment-grade ratings, because the alternative is permanent dependence on supplier credit and strategic shareholders. But the discipline that rating agencies impose is the discipline of the margin statement. Until these companies produce free cash flow that covers depreciation, interest, and the next generation of capital expenditure, the upgrade path remains blocked by arithmetic rather than skepticism.
The blockchain remembers; the architect forgets. The memory of this period will be written not in press releases but in bond indentures, rating committee minutes, and the contractual clause that turns a chip supplier into a banker. When the next cycle arrives, the architects of the AI build-out will deny the leverage they created. The debt markets will remember it perfectly. The only question is which side of that memory the market will be forced to price first: the upgrade that unlocks cheap capital, or the downgrade that reveals how much of this build-out was financed by hope rather than cash.
I do not envy the rating analyst who receives that file. The model card for the newest AI system is an engineering document. The credit memo for its builder is a confession. Vibe coding has reached the capital structure, and the capital structure does not care how intelligent the model claims to be.
