Nvidia Built a Financial Machine to Keep the AI Boom Running. What Happens When It Can't?

Nvidia Built a Financial Machine to Keep the AI Boom Running. What Happens When It Can't?
The AI data center buildout runs on two things: Nvidia chips and borrowed money. It was probably inevitable that someone would start using Nvidia chips to borrow money.
Read Full Article

Phil Kunz

Author
Phil Kunz
Writer and contributor

That someone is CoreWeave, the publicly traded "neocloud" company that pioneered GPU-backed lending in 2023. The model is straightforward: take out loans using Nvidia chips as collateral, buy more Nvidia chips, repeat. The first CoreWeave loan, led by hedge fund Magnetar Capital and Blackstone, carried about 14 percent interest. More private credit firms piled in. Now the GPU-backed debt market has grown to tens of billions of dollars, and actual banks like Goldman Sachs, JPMorgan Chase, and Wells Fargo have joined the party.

But Nvidia isn't just selling chips to these companies. It's investing in them, signing contracts to rent capacity from them, and backstopping their revenue. According to Nvidia's most recent 10-Q, the company has committed $26 billion for "cloud compute services" through 2031. That's not on the balance sheet. It's tucked in a note.

Jay Goldberg, an analyst at Seaport Research Partners (and one of the few on Wall Street with a sell rating on Nvidia), thinks that number represents Nvidia's backstop agreements with neoclouds. The timing lines up: CoreWeave announced a $6.3 billion deal with Nvidia in September, and Nvidia's cloud compute commitments jumped $13 billion that same quarter.

"Nvidia is buying demand here," Goldberg told The Verge.

How the machine works

CoreWeave had revenue of $1.36 billion in Q3 2025, up 134 percent from a year earlier. Its revenue backlog hit $55.6 billion. The company went public in March at $40 per share and the stock has since climbed above $100.

The growth comes with costs. CoreWeave's Q3 SEC filing shows about $14 billion in total debt. Interest expense for the quarter was $310.6 million, more than triple what it was a year ago. The company posted a net loss of $110 million despite adjusted EBITDA margins above 60 percent.

The loans are secured by the GPUs themselves. Loan-to-value ratios vary widely, ranging from 50 percent to as high as 110 percent, according to Ryan Little, senior managing director of equipment financing at Trinity Capital. The terms depend on loan length, confidence in management teams, and other contract details.

Nvidia's role goes deeper. In 2024, Nvidia was CoreWeave's second-largest customer. The chipmaker agreed to spend $1.3 billion over four years to rent its own chips back from CoreWeave, according to The Information. Then in September came the $6.3 billion deal. CoreWeave filed an 8-K with the SEC about it.

Nvidia is less chatty. Its 10-Q mentions the cloud compute obligations but chalks them up to "R&D and DGX cloud offerings." Goldberg doesn't buy it. The numbers, he says, don't add up to just R&D spending. If those cloud commitments were included as cost-of-goods-sold, Nvidia's margins would drop from 72 percent to 68 percent and earnings per share would fall from $6.28 to $5.97.

"Something changed in the last six months where the scale got so big it's warping things," Goldberg told The Verge.

The depreciation problem

All this debt is backed by chips that lose value every year. Investors including Jim Chanos (the short-seller who called Enron) and Michael Burry (of Big Short fame) have raised questions about how fast.

CoreWeave depreciates its chips over about six years. Microsoft and Google use similar schedules. Chanos argues the real economic life is closer to three years.

"If the chips last for three years, you have to depreciate a third of what you spend," Chanos said in a Yahoo Finance interview. "That's the bet you have to make if you're a CoreWeave investor."

He's been blunt about the implications: "Business models like the neoclouds, a lot of the AI companies themselves are just loss-making enterprises right now. You've gotta hope that changes, because if it doesn't, there's going to be debt defaults on these things."

Nvidia CEO Jensen Huang hasn't helped matters. At the company's GTC conference in March, he quipped: "There are circumstances where Hopper is fine... That's the best thing I can say about Hopper." He'd previously said that once Blackwell ships in volume, "you couldn't give Hoppers away."

That's a problem if you're a neocloud sitting on data centers full of Hoppers with debt payments to make.

CoreWeave management disputes the bear case. CFO Nitin Agrawal has said the company successfully recontracts older chip capacity. CEO Michael Intrator told analysts that Ampere chips (released in 2020) remain fully booked. The company's SEC filings acknowledge that overestimating useful life is a risk, though.

What happens if it breaks

Sarah Bloom Raskin, former deputy secretary of the US Treasury, sees warning signs in the debt structures. Data centers are creating asset-backed securities. GPU debt is spawning derivatives.

"They're like the derivatives we saw with the mortgages" before 2008, she told The Verge.

Nvidia's own history offers a cautionary tale. In 2022, crypto mining collapsed. Nvidia got stuck with more than $1 billion in chip inventory. Net income dropped 55 percent. Crypto lenders repossessed so many mining rigs they started mining themselves just to unload them.

That was $4 billion in total crypto mining debt. The GPU-backed debt market is much larger, and it's connected to regular banks in ways crypto lending wasn't.

Goldberg outlined what he calls the house of cards scenario. A smaller player takes out loans to build a data center. Construction gets delayed—weather, power issues, whatever. The company can't make payments. A bigger player like CoreWeave might survive a setback. But a small company goes under, and the complexity of these deals means one collapse could cascade up until someone like Microsoft winds up eating $20 billion in debt it didn't want.

"That seems like the house of cards scenario," Goldberg said.

The competition question

Nvidia has good reasons to keep the neoclouds alive. Its revenue is concentrated: in Q2 of fiscal 2026, two unnamed customers accounted for 23 percent and 16 percent of sales. Those big customers are almost certainly hyperscalers, and they're building their own chips. Google's TPUs. Amazon's Trainium. Microsoft's Maia. Meta's MTIA.

When Google released Gemini 3 in November, trained entirely on its own TPUs rather than Nvidia hardware, the market noticed. Nvidia's stock dropped 3 percent on reports that Meta was in talks to use Google TPUs. The company posted a defensive statement on X: "We're delighted by Google's success—they've made great advances in AI and we continue to supply to Google. NVIDIA is a generation ahead of the industry—it's the only platform that runs every AI model and does it everywhere computing is done."

That's not the tone of a company feeling secure about its dominance.

Neoclouds give Nvidia an alternative customer base that won't develop competing chips. But supporting them costs money. And if competition from TPUs and custom silicon cuts into Nvidia's margins, the company may have less cash to throw at bailouts.

Goldberg thinks the current pace can't continue. "We're bumping up against the limit of what it's possible for them to support and finance," he said. "It can't go on forever. I don't know if it stops next year or the year after, but it can't go on at this pace. Something's gotta give."

What to watch

CoreWeave has about $7.5 billion in debt that needs refinancing soon. In December, the company completed a $2.25 billion convertible note offering at 1.75 percent interest, far cheaper than its earlier high-yield debt. Capital markets are still willing to fund the expansion.

But execution risk looms. Construction delays at a Texas data center pushed completion back several months. That site is supposed to serve OpenAI, which has contract terms allowing it to walk away if CoreWeave can't deliver. OpenAI-related contracts account for a huge chunk of CoreWeave's backlog.

D.A. Davidson analyst Gil Luria put some numbers on the margin pressure: CoreWeave operates at roughly 4 percent margins, less than half what it pays in interest.

"The bull case is that they'll scale into it, and that a lot of companies have low margins to start, but this is a company at scale," Luria told the Wall Street Journal.

Mark Zandi, chief economist at Moody's Analytics, sees a bigger picture. The tech sector has taken on more debt than it did during the dot-com bubble, he said. Private debt is riskier than bank debt: larger loans, higher rates, lower priority in repayment. About half of private debt borrowers also carry bank loans. When distress hits, they draw on everything at once.

"Borrowing by AI companies should be on the radar screen as a mounting potential threat to the financial system and broader economy," Zandi said.

So what's actually happening here

Nvidia has built a system where it sells chips to companies, invests in those companies, signs contracts guaranteeing their revenue, and—if Goldberg is right about the backstops—essentially subsidizes their business models. This keeps chip demand high and diversifies Nvidia's customer base away from hyperscalers building alternatives.

The question is how long it can last.

If AI monetization happens fast enough, the neoclouds generate real profits, refinance their debt, and everyone makes money. If it doesn't, and competition erodes Nvidia's ability to keep writing checks, lenders end up holding depreciated chips they can't sell. Private credit funds take losses. Those losses ripple to the pension funds, endowments, and family offices backing them. Banks with private credit exposure feel it too.

As Raskin put it: "The parallels to the financial crisis are interesting. It's rhyming in a number of ways."

Maybe the bet pays off. Maybe AI becomes so useful that demand for GPU compute keeps growing for decades. Or maybe this is vendor financing with extra steps, and we're watching a game of musical chairs where Nvidia controls the music.

The music is still playing. For now.


Never trade based on information without verifying it yourself first. To help you, we posted some sources here where you can get started with that. Note: Some of the things we say in this article may be wrong, turn out the complete opposite, or be like complete nothing-burgers.

Sources

SEC Filings

News and Analysis

Company Announcements

Background

In Better News: Renewables Surpass Coal Globally (Dec 15-22)

In Better News: Renewables Surpass Coal Globally (Dec 15-22)
Evidence of progress, once a week.
Read More

Exponential Tech vs Linear Infrastructure | Dec 15-21, 2025

Exponential Tech vs Linear Infrastructure | Dec 15-21, 2025
An in-depth reading of what scaled, shifted, or stalled in tech this past week.
Read More

Floodgates Open, But Close Just as Fast: Susie Wiles and the Vanity Fair Interview

Floodgates Open, But Close Just as Fast: Susie Wiles and the Vanity Fair Interview
Rare, extended access produced an unusually candid account of power, loyalty, and fracture inside the White House.
Read More

The Future of the IT Freelancer in the Age of AI

The Future of the IT Freelancer in the Age of AI
Your competitor works at 3 AM, never gets tired, and charges nothing per hour. Time to rethink what you're selling.
Read More

Tech Stack — Weekly Briefing (Nov 30-Dec 6, 2025)

Tech Stack — Weekly Briefing (Nov 30-Dec 6, 2025)
AWS launched Graviton5, SpaceX won approval for 76 annual launches, and OpenAI declared code red against Google.
Read More

No Gold Bar Large Enough: Trump's Gambit in Venezuela

No Gold Bar Large Enough: Trump's Gambit in Venezuela
A critical look at how U.S. militarism, transactional diplomacy, and resource ambitions are converging to push Venezuela toward a dangerous new phase of intervention
Read More
coffee.link Context for the Present Politics Tech Culture Science Cup of Coffee Tech Stack Sign up Archive Newsletter Jobs Legal Info Privacy Policy Terms and Conditions Disclaimer Contact Us Authors Privacy Policy Terms and Conditions Disclaimer Legal Info