Nvidia Is Now Cosigning the AI Boom It Supplies

Monday offered a clean split-screen on the AI trade. In Seoul, the Kospi plunged 8 percent and triggered its eighth marketwide circuit breaker of 2026¹, the mandatory 20-minute halt that kicks in when the index falls that far, as Samsung Electronics and SK Hynix sank in a global semiconductor rout. The proximate spark was Chinese memory maker CXMT, whose shares soared 466 percent² on their Shanghai debut and revived fears about memory pricing and Chinese competition. Meanwhile, in California, Nvidia announced a $5 billion equity investment³ in Safe Superintelligence, the research lab founded by former OpenAI chief scientist Ilya Sutskever, a company that has never released a product or booked revenue⁴. And The Wall Street Journal reported that Nvidia is negotiating a roughly $250 billion guarantee⁵ so OpenAI can lease a 10-gigawatt data center campus in southern Ohio, part of a project expected to cost more than $500 billion.

The world's most valuable chipmaker deepened its financial entanglement with its own customers on the very day markets delivered their harshest verdict yet on whether those customers' spending will pay off. So the question is worth stating precisely: how much of the AI infrastructure boom reflects end users paying for intelligence, and how much is a small cluster of firms underwriting one another's ability to keep buying?

Start with the mechanics of the Ohio deal, because they are more revealing than the headline number. Nvidia would pledge its balance sheet so that lenders treat OpenAI as a safe counterparty, something OpenAI's non-investment-grade credit profile⁶ cannot accomplish on its own. The guarantee would cover the lease and construction debt but not the chips, and Nvidia is separately discussing financing up to $350 billion of OpenAI's chip purchases⁵. That is vendor financing in the plain meaning of the term: a supplier deploying its own credit so a customer can buy the supplier's product. For scale, every partner lease guarantee Nvidia has disclosed in its most recent quarterly filing adds up to $3.5 billion⁷. The arrangement under discussion is roughly seventy times larger than everything currently on the books.

The bearish reading goes straight to 1999. During the telecom bubble, Lucent committed $8.1 billion, Nortel $3.1 billion, and Cisco $2.4 billion⁸ in loans to customers buying their own equipment. When those customers' revenue failed to materialize, 47 competitive carriers went bankrupt between 2000 and 2003, the vendors wrote off the loans, and Cisco's stock never regained its peak. On this view, Nvidia's lattice of stakes and backstops, $30 billion already invested in OpenAI⁷, positions in CoreWeave and xAI, and now SSI, is the same machine with bigger numbers.

I spent a while trying to make that analogy hold, and it half fails. The dot-com carriers were thinly capitalized shells building fiber that sat dark. Today's compute is not sitting dark; it is being consumed. OpenAI's annual recurring revenue grew from $2 billion in 2023 to more than $20 billion in 2025⁹, tracking almost exactly with its compute expansion from 0.2 to 1.9 gigawatts, growth its CFO called "never-before-seen growth at such scale." Microsoft, one of the hyperscalers (the giant cloud providers that buy GPUs by the hundreds of thousands), is spending roughly $190 billion on capital expenditure this fiscal year¹⁰ while reporting 40 percent Azure growth, and has told investors it expects to remain capacity constrained¹¹ through at least the end of its fiscal year. When your problem is unpowered GPUs waiting for grid connections, you do not need to manufacture demand.

But the analogy only fails where the customer is Microsoft. Where the customer is OpenAI, it tightens uncomfortably. OpenAI has raised its projected compute spending to around $750 billion through 2030⁷ against that $20 billion revenue run rate. Sequoia's David Cahn, who has tracked this arithmetic for three years, now estimates that 2026's $1.5 trillion in AI infrastructure spending requires roughly $3 trillion in revenue¹² to pencil out, while OpenAI and Anthropic combined generate on the order of $80 billion. Allianz Research puts the divergence between AI capex growth and revenue growth at roughly 46 percent, wider than the 32 percent gap¹³ recorded during the 2001 telecom cycle. Real revenue, growing fast, can still be catastrophically insufficient for the commitments stacked on top of it.

And here is the structural problem that survives every rebuttal: supplier credit corrupts the demand signal even when no sham revenue is ever booked. Consider the CoreWeave arrangement. Under a $6.3 billion order disclosed to the SEC, Nvidia is "obligated to purchase the residual unsold capacity"¹⁴ of CoreWeave's data centers through April 2032 whenever CoreWeave's own customers do not fill them. Nvidia, in other words, has agreed to become the buyer of last resort for the downstream output of its own chips, a backstop that helped CoreWeave reassure its lenders¹⁵. Each deal is defensible in isolation. Taken together, the equity stakes, the capacity backstops, and now a prospective quarter-trillion-dollar cosign mean that Nvidia's order book has stopped being independent evidence of demand. Outside investors can no longer easily distinguish customers who could buy without Nvidia's help from customers who exist at this scale because of it.

The credit markets have started pricing this. On July 9, S&P cut Oracle's rating to BBB-, one notch above junk, explicitly because OpenAI accounts for roughly half of Oracle's $638 billion backlog¹⁶ of contracted cloud obligations, of which only about 12 percent converts to revenue in the next year. A single private, unprofitable company's growth trajectory is now embedded in Oracle's credit rating, CoreWeave's business model, SoftBank's Ohio buildout, and Nvidia's prospective guarantees. That is the honest verdict on the circularity charge: the demand is real, but the growth rate of the demand is increasingly cosigned, and the cosigners are all exposed to the same borrower.

Which is also why Monday's carnage, for all its violence, was repricing rather than proof of a demand shock. Nothing in the Seoul selloff involved a cancelled order, a falling utilization rate, or a hyperscaler cutting capex guidance. An actual demand shock has a specific signature: Microsoft or its peers guiding capex down, Nvidia's customer advances and deferred revenue reversing, Oracle's backlog conversion stalling, GPU utilization sliding toward the sub-50 percent zone where telecom-style writedowns become plausible¹⁷. None of that has happened yet. What has happened is that equity investors realized they cannot verify how much of the boom would survive without the vendor's checkbook, and they repriced the uncertainty itself.

The document that will settle this is not a stock chart but Nvidia's next 10-Q. If the Ohio guarantee is signed, the disclosed guarantee exposure jumps from $3.5 billion to a figure with two more zeroes, and at that point Nvidia's revenue and its customers' solvency formally become the same trade. OpenAI's revenue has roughly tripled each year alongside its compute; it needs to keep doing approximately that to grow into $750 billion of commitments. The year it merely doubles is the year the cosigner finds out what the loan was really worth.