“What is a GPU?” Blackstone Inc. executives asked CoreWeave Inc.’s leadership. Also: “Is AI real?”
These were entirely logical queries in the early innings of the AI boom, and at a time when CoreWeave was still a little-known AI cloud company on the hunt for funds to buy chips. Blackstone’s tactical opportunities group was convinced by the answers it received, and led a $2.3 billion private credit loan in 2023.
This marked the first major deal of its kind and became a blueprint for chip-backed debt that, three years on, is on track to become a trillion-dollar market, eclipsing even the borrowings used to build the data centers housing the chips.
It’s a testament to how intertwined the global AI expansion and credit markets have become in such a short time. Just this week, Blackstone President Jon Gray called demand for AI almost “parabolic.” But CoreWeave is also emblematic of the anxiety around the money being poured into the space — its stock and corporate bonds have been battered in recent weeks as investors question whether AI revenues can ever justify the breathtaking spend.
The market’s projected growth has drawn in virtually every major US bank and alternative-asset manager, along with backstops from AI chipmakers Broadcom Inc. and Nvidia Corp., which have a combined value of roughly $7 trillion.
Broadcom is working with Wall Street firms on a $60 billion debt transaction for a special entity that would buy chips known as tensor processing units and lease them to Anthropic PBC. Meanwhile, SpaceX is in talks to raise $40 billion for Nvidia chips. That follows Nvidia’s announcement of a $500 billion partnership with six major lenders to help customers purchase its graphics processing units, or GPUs.
The scale of the spending risks leaving debt markets dependent upon AI’s profitability. The deluge also threatens to drive up borrowing costs for AI and other sectors as tech firms compete for finite capital amid rising interest rates. JPMorgan Chase & Co. analysts predict chip spending could exceed $3 trillion through 2030, with the majority likely funded by borrowing.
“You have vendor financing being done at a scale that we’ve never seen before, and being done in a way that the risk is concentrated,” said Jason Pompeii, a senior director at Fitch Ratings. “That just increases the chance that if AI ends up being a disappointing story, there will be ripple effects throughout the supply chain.”
The strategy rests on the conviction that demand for computing power will outstrip supply for years to come.
“These financings are attached to the thing of most value in the market, which is the compute capacity produced by the chips,” said Rich Myers, global head of structured products and private capital markets at Morgan Stanley. “There’s a lot of evidence that demand for compute will continue to go up and is outpacing supply of compute.”
Yet there are risks to the build-out’s continued momentum.
Local opposition has sparked a data center moratorium on new projects in Texas, while Oracle Corp. has taken steps to put off payments on a New Mexico project mired in regulatory setbacks. Meanwhile, OpenAI and Anthropic are discussing a slower rollout over safety and existential concerns following a spate of incidents.
But much of the spending is already locked into mega data centers under construction, ensuring chip demand as those sites come online. With vendors largely sold out and severe memory shortages driving costs up, companies that make servers containing Nvidia’s AI chips informed customers in August that prices will rise by more than 15% in many cases.
“This is a new part of the economy that needs to have an orderly mechanism to finance it,” said John Watson, a senior managing director at Blackstone in the firm’s tactical opportunities group.
For Watson, there’s a parallel to the early, uncertain days of the automotive industry and Ford’s mass-market Model T car.
“If you were financing the Ford Model T, you would have asked, ‘What am I doing? Are people really going to drive cars?’” he said. “The automotive finance market is robust and well understood and is enormous in scale. It became the most bought machine in America. We have a similar vision for compute.”
Costly Chips
Fueling these capital needs is the reality that chips cost two to four times more than the data centers housing them, yet depreciate over about five to six years.
At AI cloud provider Nscale’s flagship development in West Virginia, for example, the on-site power plant is expected to cost $6.5 billion and the data center and equipment another $18 billion. The GPUs: $47 billion.
Investment-grade cloud giants Alphabet Inc.’s Google, Microsoft Corp., Amazon.com Inc. and Meta Platforms Inc. cover these hardware costs largely through cash or cheap on-balance-sheet bond sales. But OpenAI and Anthropic, burning through billions, lack those reserves and debt market clout. The same is true of neoclouds, companies focused on offering cloud computing for AI.
“The infrastructure build-out is unprecedented,” said Mike Liberatore, business finance officer at OpenAI. “There’s no playbook that shows you how you go from zero to AGI and raise money through traditional means, given it’s a very capital-intensive business,” he said, referring to artificial general intelligence.
Consequently, the industry has adapted mechanisms from project and structured finance. By creating special purpose vehicles to ring-fence each deal, issuers raise debt based on a transaction’s merits and the quality of the customer, often keeping it off corporate balance sheets.
The SPV is what borrows money to buy and own the chips. The debt is backed by the cash flow from the intended customer, which covers interest payments and fully pays down the debt over time, and the value of the chips themselves.
By aligning the debt’s maturity with the chips’ expected five-to-six-year depreciation schedule, lenders insulate themselves from risks that the hardware loses value.
Two primary versions have emerged: deals where customers rent chips from an SPV to plug into their own technology stacks, and transactions where neoclouds sell computing capacity to customers and use debt to help pay for buying the chips to build out this service.
Broadcom and Nvidia are taking this a step further by offering to backstop multibillion-dollar deals, promising to support lenders if something goes wrong. This aligns the debt’s credit rating with the chipmaker’s investment-grade status, reducing the SPV’s borrowing costs.
A representative for Broadcom declined to comment and a representative for Nvidia didn’t have a comment. Their executives have previously said they consider risks around the funding structure to be limited.
Read More: Bond Traders Fear $70 Billion Shadow AI Backstops
An investment-grade rating is vital to tapping the largest pools of capital — particularly insurance money.
The scale of digital infrastructure financing exceeds what traditional public debt markets can supply, demanding involvement from markets such as project finance, private placements, insurance and commercial real estate, said Apollo Global Management Inc.’s global head of origination Chris Edson.
“We have been focused on high-grade opportunities where risks are mitigated through collateral security, amortization, little to no residual risk or renewal risk, diversification and high-grade off-takers,” he said.
CoreWeave Blueprint
Junk-rated CoreWeave, which has made no secret of its desire to become investment grade, was founded in 2017 as a crypto miner before leveraging its GPU experience to quickly pivot to selling AI processing power. Microsoft became CoreWeave’s largest customer as the tech giant sought access to additional GPUs to support its partner OpenAI.
To pay for the chips, CoreWeave’s founders drew on past experience in commodities. They adapted a model from markets like liquefied natural gas, where customers sign contracts first, then developers use those agreements to secure debt to build facilities.
That first 2023 debt deal was backed by contracts with Microsoft, and CoreWeave had to pay up to borrow at a rate in the mid-teens.
“It was the most challenging diligence I’ve ever done in my life,” recounted CoreWeave co-founder Brannin McBee, who also described the questions Blackstone had asked the company.
As CoreWeave signed subsequent customers, it shifted from private credit deals to cheaper bank facilities. CoreWeave’s first taste of investment grade came in March, when it raised an $8.5 billion debt package backed by Meta contracts. Wall Street viewed the risk in line with Meta’s high credit rating rather than CoreWeave’s junk grade, slicing borrowing costs to roughly 6%.
Others have followed the CoreWeave playbook. Last year, when Elon Musk’s xAI needed Nvidia chips to stock its Colossus 2 data center, it raised money through an SPV that included an equity investment led by Valor Equity Partners. Apollo provided nearly $11 billion of debt for the transaction.
Apollo and Blackstone recently did a $35 billion debt deal called Big Sky to support Anthropic. This was the first transaction of a platform between Broadcom, Apollo and Blackstone called AI XPV that could finance 20 gigawatts of capacity — which Bank of America Corp. strategists estimate may require $235 billion of backstopped debt by mid-2029.
A special financing entity was created to raise the debt to buy the chips, which Anthropic would lease. Broadcom provided a backstop, meaning if Anthropic defaults, the chips are re-leased or sold to repay the remaining debt, with the chipmaker covering any shortfall.
This support led to an investment-grade rating for the vehicle’s two Broadcom-backed senior tranches. The largest of these pieces pays a 5.75% interest rate — cheaper than Anthropic could get as an unrated, private company — while the riskier $4.5 billion junior tranche priced at 8.5%.
“The sector can have many risks, but some financing structures mitigate those risks more than others,” said Ross McAdam, head of alternative debt at L&G — Asset Management, America.
In August, Nvidia assembled a powerhouse coalition of Apollo, Blackstone, BlackRock Inc., Brookfield Asset Management, KKR & Co. and Goldman Sachs Group Inc. to source financing for AI infrastructure, of which it would backstop up to 25% of an opportunity.
Broadcom has held early discussions to arrange financing to help OpenAI purchase custom chips that the pair is developing together, Bloomberg reported this month.
To reach true scale, though, the industry needs to transition from private markets into syndicated debt — where banks arrange bonds and loans to sell and trade among institutional investors. Earlier this year, CoreWeave became the first to enter the syndicated markets with chip deals: two leveraged loans that met solid investor demand.
But the ultimate goal is for these SPV structures to break into the much larger investment-grade bond market — one of the world’s deepest, cheapest pools of capital.
Bonds Tumble
CoreWeave’s corporate-level junk bonds have taken a hit recently, with some sliding to around 80 cents on the dollar and pushing yields to roughly 15% as investors worry about the company’s expanding debt. However, the market views its two chip-backed leveraged loans as separate. Both dipped slightly but still trade near par.
In a statement to Bloomberg following the selloff, CoreWeave Chief Executive Officer Mike Intrator said building a new type of technology company means there are moments where the market has to work through questions about AI adoption.
“That scrutiny is important, but what’s also important is to recognize that companies have different technologies, customer commitments, financing structures and records of execution,” he said. “We’re confident in our business model, our technology, the quality of our execution, and the relentless discipline behind our financial engine.”
Potential Risks
As capital floods in, lenders are confronting risks. Because most chip deals are private transactions, investors have no easy way to sell their positions in debt deals that have bound a small number of chipmakers, cloud computing giants and neocloud firms together. This creates heavy exposure to a small group of interconnected players.
“The big are going to get bigger,” said Dave Snyderman, managing partner and global head of alternative credit and fixed income at Magnetar. “The incumbents are likely going to win because they’re the only ones who are going to have access to capital.”
Another risk for lenders is how airtight the customer contracts really are. Document protections may prove less ironclad in practice — as they have in riskier credit markets. If AI returns disappoint, tech firms could use their clout to force lenders and neocloud firms into renegotiating deals or leases.
Everyone is assuming these contracts are tight and unbreakable, but “I look at the practical element,” said Akhil Bansal, head of asset-backed finance at Carlyle Group Inc. “If you have seven firms that are basically providing all the credit support for the GPUs, who’s got the leverage?”
The complexity is keeping some investors on the sidelines. “This is a novel structure,” said Laura Parrott, head of private fixed income at Nuveen. “We haven’t seen how that plays out if there is some sort of disruption,” she said, adding that she’s holding off on lending to big investment-grade-rated chip deals until the market develops more.
Chips Evolve
As the chip market evolves, so too will the mechanisms used to finance the hardware. Early research and real-world usage so far are showing the chips are lasting longer than current accounting treatment suggests, which would make them easier to finance.
The equivalent of the Kelley Blue Book for new and used cars could emerge for semiconductors, enabling chip trading and hedging to develop, said Karen Fang, global head of infrastructure and sustainable finance, and co-head of the global capital solutions teams, at Bank of America.
“When we get to a point where we have enough data history and track record on the chips, then you’re going to start seeing a much more robust leasing market,” she said.
And raising chip debt could get simpler for OpenAI and Anthropic once they do go public, especially if they can win investment-grade ratings and put together deals on their own.
The ultimate endgame for Nvidia, Blackstone and their Wall Street peers is to turn a unit of computing power into an asset class, where people globally can “finance compute because we understand the economic value of the compute,” Blackstone’s Watson said.
