On Oct. 21, insuring $10 million of CoreWeave’s debt against default for five years cost about $1.145 million a year, according to CNBC, and the company that sells CoreWeave its chips, owns about a tenth of its stock and has agreed to buy up to $6.3 billion of its unsold computing time is Nvidia.
Over the past year Nvidia has lent its balance sheet to the A.I. buildout, and as CNBC’s Kif Leswing and Isabel O’Brien reported on Saturday, that role now faces its first real test. The scale of it is in Nvidia’s quarterly filing for the period ended July 26. It lists $36 billion of what the company calls A.I. cloud agreements, commitments to buy computing time from the GPU rental companies known as neoclouds. It lists $20 billion of data center leases Nvidia signed as tenant and expects to reassign to others. And it puts Nvidia’s maximum exposure on guarantees at $108.5 billion, nearly all of it for one OpenAI campus in Ohio.
That comes to $164.5 billion, equal to 76 percent of the $216 billion in revenue Nvidia reported for its last fiscal year. Its annual report, filed in February, did not mention A.I. cloud agreements at all. Beyond those commitments sit a $99 billion portfolio of stakes in companies including OpenAI, Anthropic and CoreWeave, and a residual-value program with Wall Street firms that could put up to $125 billion more on the line if fully used. Lucent’s customer financing, the cautionary case from the last infrastructure boom, peaked at 24 percent of its revenue.
The record supports a narrower worry than the selling suggests. Nvidia pays on almost none of this unless a customer defaults. It generated $70 billion of free cash flow in the first half of the year, and the GPUs it stands behind rent for more today than they did a year ago. The danger is in the timing. The guarantees come due only if demand for A.I. computing disappoints, which is when Nvidia’s own sales would be falling too. Nvidia declined to comment to CNBC.
What the filing shows
Most of the $36 billion comes from a program Nvidia started on July 1. Nvidia agrees to rent a neocloud’s GPUs at a floor price if the operator can’t find other customers, which gives the operator a guaranteed income to borrow against. In return Nvidia takes a share of the rental revenue above that floor, reportedly as much as half in some proposed six-year deals, The Next Web reported, citing The Wall Street Journal.
SemiAnalysis, a chip research firm, traced $27.5 billion of the total to three deals: $21.1 billion for Firmus’s 360-megawatt campus on the Indonesian island of Batam, $4.2 billion for Australia’s SharonAI and an estimated $2.2 billion for GMI. It put the average floor at about $2.35 per GPU-hour for Nvidia’s GB300 systems, against $4.50 to $4.60 for five-year contracts on the open market.
The fine print shows which way the deal cuts. Nvidia’s commitment shrinks whenever someone else rents the capacity, and the clouds “can unilaterally stop providing to us and sell to third-party customers at more advantageous rates,” according to the 10-Q. In a strong market, the obligation shrinks toward nothing. In a weak one, Nvidia keeps paying the floor.
The program ran into resistance within two months. The Journal reported on Aug. 27 that Nvidia had paused some deals after employees raised antitrust concerns and partners objected to Nvidia approving their end customers. Nvidia denied the report, and a spokesperson said the model “is still in place and continues to evolve due to high demand.”
SemiAnalysis’s Sept. 11 note, by Daniel Nishball, Oliver Kennon and Terence Ong, added up six lines of the filing and got $530 billion of gross off-balance-sheet commitments, up from $184 billion a quarter earlier and against $91 billion of liabilities on the balance sheet. The biggest piece, $279 billion, is supply and capacity purchases Nvidia makes to build its own products. The firm posted the note on X under a subtitle asking who loses if the bet goes wrong.
Nvidia’s Backstop Universe
Heads I Win, Tails Who Loses?
The $11T AI Buildout,
Nvidia’s Backstop Economics,
and the Limits of Nvidia’s Balance Sheet
New Constructs, a research firm, estimates the net present value of Nvidia’s off-balance-sheet liabilities in fiscal 2026 at more than $153 billion, over $100 billion higher than a year earlier. “If the demand for compute doesn’t meet these really super lofty expectations, we’re going to have a problem,” David Trainer, the firm’s founder and chief executive, told CNBC.
Twenty years in Ohio
The Ohio guarantee started out bigger. CNBC had reported that Nvidia was discussing a backstop of up to $250 billion so OpenAI could raise debt for a 10-gigawatt campus, and The Journal later reported the figure would fall below $120 billion. Talk of that deal, along with a partnership with SK Group valued at more than $500 billion, pushed Nvidia’s own five-year credit default swaps to a record 82 basis points on July 27, The Next Web reported. The signed version, disclosed Aug. 17, caps Nvidia’s payments at $105 billion.
The campus is in Pike County, at the former Portsmouth Gaseous Diffusion Plant, on federal land the Energy Department is leasing to an SB Energy affiliate. Under Nvidia’s 8-K filing, the guarantees cover leases on about 4.25 gigawatts of computing load and are triggered if OpenAI becomes insolvent and defaults or stops paying rent. Nvidia can then take over the lease, have SB Energy re-let or sell the space, let the lease end, or wait up to a year while paying certain project costs. It covers the gap between a guaranteed minimum and whatever a new lease or sale brings in. CNBC reported that Nvidia could also pay SB Energy a lump sum. OpenAI is to reimburse Nvidia for anything it pays.
The guarantees end after 20 years, if OpenAI terminates the lease or if OpenAI achieves “a satisfactory credit rating.” In return the site will run Nvidia’s full stack of hardware, with limited exceptions. Nvidia’s exposure builds as each of nine construction phases is finished, and the 10-Q says the first is expected in fiscal 2029. Nvidia also holds an option to back about 3.8 more gigawatts.
Nvidia sits on several sides of this deal. It bought $30 billion of OpenAI stock in a funding round this year. It said in August it would put $1.5 billion into SB Energy, the landlord, and Quartz reported, citing The Journal, that Nvidia has committed $3 billion as part of SB Energy’s planned share sale. It was not clear whether the two sums overlap. SB Energy, which filed to go public at the end of August, has no data centers running yet and has issued OpenAI warrants valued at about $5.5 billion. OpenAI, for its part, has put off its own I.P.O. and is raising $30 billion at about a $1.4 trillion valuation.
The customers
CoreWeave, the largest neocloud, carries about $51.6 billion of debt, including roughly $35 billion in notes, bonds and loans. Its five-year swaps rose as high as 1,194.5 basis points on Oct. 15, a level that priced a better than 60 percent chance of default within five years, Bloomberg reported, citing ICE Data Services. Yields on its bonds due in 2031 and 2032 have topped 15 percent. S&P Global Ratings rates the company B+, well into junk. Its market value, about $90 billion at its peak, is now about $31 billion.
Nvidia’s money runs all through the company. It put another $2 billion into CoreWeave in January at $87.20 a share, nearly doubling its stake, TechCrunch reported. In September 2025 it agreed to buy any CoreWeave capacity left unsold through April 13, 2032, an order worth $6.3 billion. Either side can end that contract if the other enters bankruptcy proceedings that aren’t resolved within 90 days.
Jensen Huang, Nvidia’s chief executive, has said CoreWeave doesn’t need Nvidia’s backstop. “CoreWeave’s strength is improving every day,” he told Bloomberg in Stockholm on Oct. 15. “I’m optimistic. Their cost of capital is actually decreasing.” The swaps hit their record the same day. Both things can be true. CoreWeave’s secured debt, backed by GPUs and customer contracts, has held up, and an $8.5 billion facility in March was rated investment grade and priced at about 5.9 percent, according to The Next Web. The stress is in the unsecured notes.
John Lloyd, global head of multi-sector and corporate credit at Janus Henderson, still holds the 2031 bonds. “It’s a little bit of overreacting,” he told CNBC, pointing to the lack of near-term bond maturities and blaming most of the selloff on rising Treasury yields.
The pressure has spread through Nvidia’s portfolio. Shares of CoreWeave, Nebius and IREN have each lost at least 30 percent this month, a slide that deepened after Gartner forecast on Oct. 14 that 80 percent of neocloud providers would stop operating independently by the end of the decade. Short sellers had made $6 billion in paper profits on the group in October, Reuters reported. Nvidia sells to hundreds of these clouds, and picking winners among them is now its problem as a lender too.
Firmus, the Australian company behind the largest of the three known take-or-pay deals, withdrew its A$7.1 billion listing on Oct. 9, two weeks before trading was due to begin, and said it would seek private money. Nscale, based in London, filed for a U.S. listing on Sept. 18 after losing $1.02 billion on $140.6 million of revenue in the first half. Neither Nscale nor SB Energy has updated investors on its plans in weeks. Nvidia’s own shares ended a third straight losing week on Friday, down 7.6 percent for October.
Huang’s case
Mr. Huang’s answer is that the neoclouds are short of money, not customers. Demand for them is “off the charts,” he told CNBC’s Arjun Kharpal on Oct. 22, on the sidelines of Nvidia’s GTC conference in Berlin. “What is missing in the middle is the working capital to fuel that growth,” he said. “Because there’s just so much demand across so many sectors of AI, there’s a scarcity of capital.”
Nvidia wants Wall Street to supply that capital. In August it signed memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR meant to draw more than $500 billion of outside money into A.I. infrastructure. In a blog post on Aug. 11, Mr. Huang wrote that Nvidia might cover up to 25 percent of a given financing if the hardware turns out to be worth less than expected, judged project by project. Spread across the full $500 billion, that ceiling is $125 billion. Nvidia has also seeded Brookfield’s A.I. infrastructure fund with $2 billion and put undisclosed sums into KKR’s $11 billion Helix vehicle and Valor Equity Partners’ $5.4 billion compute fund.
This is the beginning of an open capital market for AI infrastructure.
The pitch depends on GPUs holding their value. “We’re the most rentable. We’re fungible, and we’re durable,” Mr. Huang said on a podcast with the investor Brad Gerstner published last week. Rents back him for now. His blog post cited one-year H100 contracts rising from about $1.70 to about $2.35 an hour between October 2025 and March 2026, and Ornn’s index of completed H100 trades stood at $3.15 an hour on Oct. 22, up 10.5 percent in a month.
Lenders are less sure. Bankers and asset managers have questioned whether the chips can serve as long-term collateral on their own, and some want guarantees on every deal, Reuters reported on Oct. 3. Mr. Huang has argued GPUs can stay useful for up to a decade. “Banks typically underwrite GPUs over a 3-4 year depreciation schedule,” said Tony Trzcinka of Impax Asset Management. The precedents lenders like rest on someone else’s credit: CoreWeave’s $8.5 billion facility relied on payments from Meta, and Broadcom backstopped more than 80 percent of a $35 billion structure financing compute for Anthropic.
The Lucent precedent
The comparison investors reach for is Lucent Technologies. In the late 1990s Lucent lent money to young telecom carriers so they could buy its gear, and its commitments reached $8.1 billion, equal to 24 percent of its $33.6 billion in revenue, the venture capitalist Tomasz Tunguz calculated. Nortel extended $3.1 billion and Cisco promised $2.4 billion. When the carriers failed, Lucent set aside $2.2 billion for bad debts in 2001 and $1.3 billion in 2002. By 2002 its revenue was 69 percent below its 1999 peak.
Mr. Tunguz ran the same numbers on Nvidia in October 2025, when its investment pledges, led by a $100 billion letter of intent with OpenAI, came to $110 billion, or 67 percent of trailing revenue. He still concluded the cases differ, since Nvidia’s biggest buyers were cash-rich and its sales were limited by supply, and wrote that “this merry-go-round has paying riders.” The OpenAI pledge was later cut to the $30 billion stake. He laid out the comparison in a thread on X.
Nvidia just committed $110B in vendor financing—85% of annual revenue. Lucent’s playbook in 2000? 20%. Is this the telecom bubble all over again? I analyzed the numbers. Thread 🧵
Nvidia drew the same line itself. In a seven-page memo to analysts in November 2025, it said that “unlike Lucent, NVIDIA does not rely on vendor financing arrangements to grow revenue” and that its customers pay within 53 days, Yahoo Finance reported. Jim Chanos, the short seller known for calling Enron’s collapse, wasn’t persuaded. “They’re [Nvidia is] putting money into money-losing companies in order for those companies to order their chips,” he told Yahoo Finance.
On the memo’s terms, Nvidia is right. It hasn’t lent customers money to buy chips, and almost none of its exposure is a loan. But the take-or-pay floors and lease guarantees do the job Lucent’s loans did: they let buyers that couldn’t borrow cheaply on their own credit buy the equipment, with the seller’s balance sheet behind them. And several of the buyers carried this way burn cash, as Lucent’s carriers did. The difference is the seller. Lucent’s financing was a quarter of its sales. Nvidia’s maximum guarantee exposure of $108.5 billion is less than a year of its free cash flow at the first-half pace.
Weighing the risk
The bull case starts with that cash. Revenue in the quarter ended July 26 was $96.2 billion, more than double a year earlier, and Nvidia guided to $108 billion for the current quarter. “They’ll do $700 billion in revenue next year,” Gene Munster, founder of Deepwater Asset Management, told CNBC in September. “That affords them the ability to finance some of their customers.” The Ohio exposure doesn’t begin until the first buildings are finished.
Risk reward is asymmetric in Nvidia’s favor as long as its balance sheet remains strong.
The same note named the catch. Nvidia loses, its authors wrote, only if the backstopped neoclouds fail to meet their commitments at the same time Nvidia’s own cash generation slows, and those two risks are correlated. Lenders call that wrong-way risk: the guarantor is weakest when the guarantee is called. Morgan Stanley’s credit analysts, Lindsay Tyler and Nishant Satyam, started covering Nvidia’s debt on Aug. 24 with a neutral rating and estimated its total exposure could reach about $200 billion by the end of 2028, about $170 billion of it off the balance sheet, BigGo Finance reported. Nvidia’s swaps were then around 83 basis points, more than double where they traded for most of their history and still the price of a highly rated borrower.
Moody’s said guarantees are the fastest-growing kind of off-balance-sheet financing in 2026. The agency doesn’t count them as debt when it rates corporate buyers. David Gonzales, a vice president at Moody’s, told CNBC the chips will likely keep “values in excess of contractual residual value amounts” in the near future, which would limit the risk in many of Nvidia’s structures. How much exposure investment-grade A.I. companies can carry, he said, is “the question we’re evaluating at all times every day.”
These things are advanced structures that are unique, and so there is no blanket real approach you can take
Nvidia has been looking for insurers to take on some of the risk that GPUs lose value. “Insurers are not yet a complete panacea to this problem,” said Tom Hoad, executive director of product innovation at the insurance group Howden. Warren Hosseinion, a co-founder of Compute Labs, which tokenizes GPUs to raise money for neocloud operators, said Nvidia doesn’t want to be an insurer, “but they’re willing to play that role while the market and the industry catches up.”
Measured against the record, the competing claims sort out this way. Mr. Huang’s point about CoreWeave’s falling cost of capital holds for its secured debt and fails for its unsecured notes, which yield more than 15 percent. His claim that GPUs are durable is supported by rents, which are rising, and doubted by lenders, who still price the chips over three to four years. Mr. Trainer’s warning is conditional, and Nvidia has disclosed no payment under any of its guarantees. The weakest part of the structure sits in the fine print. The take-or-pay floors are a one-way bet sold to the neoclouds, worth little to them in a boom and a great deal in a bust, and the bust is when Nvidia’s own orders would be under pressure.
What comes next
CoreWeave is expected to report third-quarter results in the first half of November. Analysts expect negative free cash flow of $10.1 billion, the most since it began reporting quarterly, according to FactSet data cited by CNBC. Bond investors told CNBC that new deals from lower-rated issuers are largely on hold until after the Nov. 3 midterm elections.
Nvidia hasn’t said when it will report the quarter that ends this month. Last year it announced on Oct. 29 that it would report on Nov. 19. The 10-Q that follows will show whether the A.I. cloud agreements and guarantees grew again.
The first $6 billion of those A.I. cloud commitments falls due in Nvidia’s fiscal 2028, which begins in early 2027. The first of the nine Ohio phases is expected the year after that.