Stay informed with free updates
Simply sign up to the Big Tech myFT Digest — delivered directly to your inbox.
Tech companies are increasingly turning to loans backed by the chips on which their large language models are trained as they hunt for ways to fund their massive AI investments.
Such loans, which are secured against graphics processing units and backed by leases to the tech groups, are popular with a sector burning hundreds of billions of dollars a year in an AI arms race on chips that can quickly become obsolete.
Investors have been attracted by yields in the high single digits to mid-teens, which are typically higher than those on debt issued by the tech companies themselves.
“Investors are very excited,” said David Ridenour, a partner specialising in finance and restructuring at law firm King & Spalding. “People are willing to dive into [GPU deals] on a take-it-or-leave-it basis.”
Pioneered by cloud computing provider CoreWeave in late 2023, GPU-backed debt has grown in popularity as demand for advanced chips skyrockets and prices soar. Citigroup estimates that GPUs and associated servers can account for 30 to 40 per cent of total project costs for data centres.
The loans are typically taken out by special-purpose vehicles formed by tech companies and investment firms for the purpose of acquiring a cache of high-performing chips, which would then be leased to the tech businesses to train their AI models.
This arrangement allows Big Tech groups, whose use of debt markets is growing rapidly, to shift the loans off their corporate balance sheets.
Last month, Apollo announced a $3.5bn financing package for a digital infrastructure fund managed by Valor Equity Partners, which would buy Nvidia’s GB200 hardware, known as AI superchips, and lease them to Elon Musk’s xAI.
IREN Limited, an AI cloud service provider, also secured $3.6bn in loan commitments from Goldman Sachs and JPMorgan earlier this month to buy chips for its AI contracts with Microsoft.
Lenders often have to act fast and write big cheques in these transactions, according to a lawyer familiar with GPU financing.
“A big player would basically ask, ‘would you like to participate in a deal that closes in two weeks and throw in a couple hundred million?’” the lawyer said.
The rising popularity of such loans highlights investors’ clamour for asset-backed finance, where banks and private credit funds seek esoteric debt secured by stable cash flows.
Deals usually come with “hell or high water” clauses that prevent tech companies from terminating the leases early, helping mitigate the risk that these GPUs become obsolete as AI technology quickly evolves.
“It is a very new space and a lot of people are grappling with the question of GPU lifespan,” said Dorina Yessios, US co-head of energy, infrastructure and natural resources at A&O Shearman. “That has to be factored into underwriting, just like any other equipment financing.”
Moody’s, which has begun to rate GPU-backed debt, said it withdrew credit ratings once underlying leases came to an end.
“Typically, deals we rate repay everything within the first lease term, so you don’t have to think about the extent of useful life,” said John Medina, senior vice-president in Moody’s global project and infrastructure finance team.
Nevertheless, some investors remain concerned that the economic life of GPUs could end up being shorter than expected, while the market value of older AI chips is often questionable due to a lack of price history in the nascent industry. Current valuations could also be inflated by short-term chip supply shortages, investors said.
“We really want to ensure the GPUs’ useful life well exceeds the amortised period of our investment,” said Jen Marques, head of strategy and structuring for Oaktree’s structured credit strategy.
“Those things won’t make it three years before they are antiquated. It’s a huge gamble,” said an investor who has turned down multiple GPU financing pitches.
“The idea of reselling GPUs from a few years ago [after a default] is like beating a dead horse.”
