Listen to the latest episode of Credit Exchange with Lisa Lee
Published in London & New York
10 Queen Street Place, London
1345 Avenue of the Americas, New York
Creditflux is an
company
© Creditflux Ltd 2026. All rights reserved. Available by subscription only.
Analysis AI financing
AI financing creates a new asset class
by Lisa Fu and Kathryn Gaw
Financing for chips that power AI data centres has led to innovative structures and multibillion-dollar deals as capital comes in from every corner of the financial markets
The debt capital markets may have just birthed a new asset class. Over the past 18 months, providing financing for chips such as GPUs and TPUs – vital equipment that sits inside data centres – has become a massive effort, bringing new players into the sector and leading to complex deals worth tens of billions of dollars.
“It is unique in the pure quantum of capital required,” Eric Wittleder, chief investment officer for Brookfield’s infrastructure credit strategy, says. “This is a once in a generation capital opportunity.”
In fact, according to an Apollo analysis, hyperscaler capex spending from the likes of Google, Meta and Amazon is likely to exceed USD 1tn next year. Goldman Sachs has predicted that in order to fully finance the AI buildout, an estimated USD 7tn will be required between 2026 and 2031 across compute, power and data centres.
At this scale, the capital must come from every corner of the financial markets, incorporating equity, sovereign capital and public and private debt as well as – in the words of Goldman Sachs’s co-head of global banking and markets Dan Dees – “new joint venture structures, some of which have yet to be invented.”
As AI continues to consume compute, data centres get larger and these chip financing deals that help put chips into those shells will also get larger. Theoretically, a single chip financing deal could get to USD 100bn, Jake Streit, the managing director for TMT at SMBC, tells Creditflux.
Only getting bigger
“I would say what’s driving the size in the market is the cost,” Streit says. “So, if you put things in perspective, chips cost around three to four times the cost of the warm shell.” A one-GW data centre can cost USD 10bn, but the cost of the chips for that data centre cost roughly USD 35-45bn, he adds. Chips financed to date by SMBC’s estimation have been about USD 128bn to USD 130bn, Streit says. Streit estimates that the number will approach half a trillion dollars within the next three to four years.
As deals grow ever larger, the market has moved away from earlier chip financing deals that relied on bank balance sheets or smaller equipment finance specialists. Now, massive deals are taken down by a variety of parties all seeking different types of debt profiles.
We should go as fast as we can, irrespective of anybody else
Jensen Huang
CEO of Nvidia
On CBS News, 18 September
Uncertainty around the value and life of the chips, serving as collateral, was previously a barrier to bringing in a wider group of investors and creating larger deals. But over the last year or two, the market has come up with some novel structures to reduce the risk associated with potential chip resales and gotten better at matching risk-return to the wide variety of players willing to take down this debt.
“As it evolves, the chip financing space is [borrowing] more and more financing technology from other asset classes, including infrastructure,” Wittleder tells Creditflux. “That’s opening up opportunities for more traditional institutional investors, private placement investors, and others.”
The only control or “guardrails” that AI needs is a STRONG AND SMART (High IQ) PRESIDENT, and the USA has that, in spades!
US president Donald Trump
On Truth Social, 14 September
One example of this is the tranching out of different risk profiles, a technique that has long been used in the telecoms and infrastructure sectors that is now being applied to chip financing, Wittleder says. Similarly, a lot of the chip financing deals have evolved from bilateral private credit deals to bank clubs, then syndicated term loans, then 144As and, most recently, tranched issuances.
The gamechanger, however, has been the addition of guarantees and backstop agreements from creditworthy hyperscalers and chip makers, market sources say.
Look no further than the USD 8.5bn GPU financing deal involving CoreWeave and Meta that closed in March, and the USD 35bn TPU financing deal involving Anthropic and Broadcom that closed in June. Both of those deals managed to secure an investment-grade rating for a portion of the debt by relying on some kind of backstop or guarantee agreement.
In the first, the deal was backed by the GPU infrastructure as well as a tight master services agreement with Meta, an IG-rated company. In the second, the chips were put up as collateral and Broadcom – another IG-rated company – agreed to provide a residual guarantee. That means Broadcom would make the lenders whole if the chips did not retain their value.
AI is certainly powerful enough to drive events that, you know, cause a billion deaths
Bill Gates
On NBC News’s Meet the Press, 27 September
A new asset class
Many in the market say these newer chip financing deals are becoming their own asset class, and the vast majority agree these deals need to be distinguished from data centre deals all together.
“We look at the chip financings and the data centre financings as having different characteristics,” Rob Horn, senior managing director and the global head of infrastructure and asset-based credit at Blackstone, tells Creditflux. For both, the growth in AI is driving the spending, but the attributes are very different. Notably, a data centre has a 20- to 30-year useful life, while the chips seem have a much shorter life and chip technology is evolving rapidly, Horn says.
The term of the debt is also kept short to lower risks, with most chip financings structured to fully amortise with the contracts or within the assumed life of the chips, market participants add.
“The whole feature of it was designed to get this level of financing to an end user that is not independently financeable at that level … it’s going to be sort of the standard for which other deals need to be,” says Matt Eagan, head of the full discretion team at Loomis Sayles.
New technologies must remain at the service of the human person, rather than becoming yet another instrument of domination and injustice
Pope Leo XIV
During a UNESCO address, 25 September
Indeed, many of the actual chip users are still young companies and may not have the credit rating necessary to secure cheaper IG debt. Others may want to keep their financials more private. Creditflux previously reported that Anthropic was unwilling to show its private financials to lenders underwriting a chip financing deal ahead of a much-anticipated IPO.
“Your first line of defence is the value of the chips,” Eagan says. But in the end, the corporate guarantee backs up the deal and so the corporate rating of the chip maker matters, he added.
There are deals where residual guarantees are provided by the chip makers. Then there are other deals that involve credit support or guarantees from a neocloud parent entity that is not the end user of the chips, says Kailash Gupta, a partner at law firm Davis Polk. Lastly, there are deals where the chips are tied to long-term customer contracts, and the lenders may feel more comfortable lending against the contracts rather than the uncertain value of the chips. In those deals, the credit quality of the customer in the contract is very important, he adds.
These newer deals include elements of both corporate credit risk as well as asset-backed finance risk. That means the diverse participants in the space may categorise and underwrite these transactions differently from one another. Market players have made comparisons to existing sub-asset classes such as equipment finance, vendor finance, infrastructure credit and even aviation financing. Others say this is emerging as a distinct asset class of its own – dubbed ‘digital infrastructure’ by some analysts.
“The chip financings are a mix of asset-backed, structured credit and project finance. They’re all of the above,” Blackstone’s Horn says. “So you need both industry expertise and you need structuring expertise.” At Blackstone, that means many of the investment professionals who cover digital chip financing as a new asset class have skills in the sector and in structured credit.
I do believe we can continue to be the world leader in frontier research and expect reasonable guardrails … And if we get this wrong and it genuinely gets away from us and begins designing things to hurt people, it will be the last mistake we’ll ever make
Chris Coons
US senator from Delaware
On Bloomberg This Weekend, 13 September
S&P Global has begun defining these deals as “digital infrastructure”, and the ratings agency says it has been working on a new methodology to underwrite the risk, drawing in elements of other deal structures such as infrastructure and aviation financing.
KBRA applies the term AI/digital infrastructure most commonly, although the ratings agency told Creditflux that it does not have a standard term as such. Fitch applies a framework for sectors where dedicated criteria have not yet been developed, with the methodology focused on the linkage to the sponsor, originator, or operator, and the transaction’s cashflow coverage.
“What lenders are underwriting is the investment grade contracted cash flows from the hyperscaler, That’s one of the reasons why it fits within the definition of infrastructure finance, rather than equipment finance,” says Wittleder.
Others do not place much emphasis on categorisation at all. Instead, they may evaluate the deal on its own then later look to place pieces of the deal across a variety of existing investment vehicles with distinct strategies and risk-return targets, market sources say. That means investors can choose to participate in the IG debt tranches or take down sub-IG syndicated debt or take on something more akin to private ABF risk.
AI compute/chip capex estimates rise
Source: Goldman Sachs baseline estimates (USDbn)
New risks
But new risks emerge whenever a space has rapid growth, starts introducing new deal structures and attracts newer participants.
“You have this dynamic appearing, that these chip companies are entering the market to unlock the market,” says Raj Joshi, senior vice-president for corporate finance at Moody’s Ratings. Some see this level of involvement as a red flag, others argue that the demand is real and the customers are solid companies, regardless of whether the chip maker is supporting the customer in various ways.
These new deals hinge on an IG-rated chip maker essentially acting as a guarantor for a less creditworthy chip user. The danger is that some believe these guarantors may have overcommitted themselves across too many deals and they may end up unable to fulfil all their obligations. This would leave lenders with a bunch of chips that may or may not have kept their value. “There’s just a risk there that we’re not able to quantify any more, even for a public company,” says Riley Thompson, VP and head of direct sales at Mitsubishi HC Capital America, a major non-bank commercial equipment finance company. It is near impossible for a lender to have enough staff with true legal expertise to keep track of and review all the guarantee and backstop agreements being made, he says.
To complicate things further, sometimes even the user of the chips in these deals can be unclear. The source of the repayment should be coming from the user, and the backstop is meant to be insurance, Thompson says. The lenders need to know whether the users can even afford the chips at this cost. This becomes difficult to determine if the chip users on the deal end up selling some of their compute to third parties.
That said, participants generally agree that the overwhelming demand for chips – even if their value long term is unclear – provides some comfort. After all, as long as this AI craze continues, these chips will probably stay in short supply. “A lot of this depends upon the commercial success and technical leadership of frontier [AI] labs,” Joshi says. “If you look at them, their aggregate build up demand is high [but] anytime you have a high concentration, there’s a risk in the market.”
Chips with everything
Chips are evolving faster than the power capacity of data centres
In the blink of an eye, artificial intelligence has gone from making memes to potentially ending all of humanity, and the chips that are enabling these AI models have been evolving at a similarly fast pace. But for lenders, this rapid growth is causing a major headache.
“If the product does not have a history, that means the onus is on the structure to lay out its useful life,” Joshi from Moody’s says. “That’s challenge number one. Challenge number two is where else on the planet do you see an example where a product is improving 50 times, sometimes, every six to 12 months?”
These computer microchips are specially designed to serve as infrastructure to support AI.
Chip makers have continued to introduce newer, more efficient versions of these chips to better power AI. This means that the profile of these chips is unlike other types of equipment or infrastructure assets. For example, data centres are also a key AI infrastructure component, but modelling the life of that real estate or structure has been more straightforward.
Similarly, assets such as aircraft or consumer credit cards have enough data and history to allow lenders to model for certain risks. These assets are now regularly seen in the ABS market as a result.
Managers would want to see a very fungible and deep user base for the chips before this market starts producing deals that can rely on just the value of the asset, Blackstone’s Rob Horn says. In contrast, an aircraft can be sold all over the world and there is more forward-looking clarity around the value of those assets.
Assumptions about the store of value for the chips and the useful life for these assets are actively being challenged.
Previously, the market expected these AI chips to have a life of around three to five years. Now, many companies are using chips past this five-year mark, and some chips have even been sold on the secondary market at a premium.
“Certain generations of chips have slowed their economic depreciation, and even increased in value,” says Bernie Margulies, the CEO of American Compute, a San Francisco–based insurance and structured finance company that provides residual value insurance for Nvidia GPU hardware. “This is primarily due to the newer chip generations being much more power- and cooling-intensive. This makes older chips the only viable option for many.”
Margulies adds that AI players have two options – buy older chips that can work inside older data centres without high power requirements, or buy newer chips but struggle to place them anywhere because the higher-power data centres are still under construction.
Other market participants have noted that the supply/demand dynamics of the chip market are also impacting on the value of used chips, causing their prices to shoot up sporadically whenever new chip stocks are low.
That being said, credit investors tend to err on the side of caution. Market sources tell Creditflux that lenders prefer to use more conservative assumptions and keep debt tenors short to match.
“Largely, the chip financings are structured such that they fully amortise with contracts. That is how we finance them,” Horn says. “But we do believe the underlying compute has value and that that compute has retained its value better than the market has expected.”
