Articles

AI Data Center Loans Sold at Discount: GPU Debt Explained

Banks including ING are shopping AI data center and GPU loans as lenders worry chip collateral will age faster than the debt. What's happening and why.

Kurumi Kurumi · · 6 min read
Rows of server racks in a dimly lit data center aisle

Lenders are trying to offload loans tied to AI data centers and GPU purchases, and some portfolios are being offered at steep discounts, according to reports from Semafor and Bloomberg published over the past week. The pullback centers on a single worry: that the chips backing many of these loans will lose their value before the debt is paid off.

Dutch bank ING is reported to be seeking buyers for about $500 million of data center loans, including a piece tied to the Stargate project. Ares Management Chief Executive Michael Arougheti said his firm would pass on GPU financing deals. Semafor reported that GPU-backed financings in particular were “getting smoked” in the secondary market.

The ING portfolio

Bloomberg reported on November 6 that ING was marketing a portfolio of 12 data center loans worth about $500 million. The package includes a $25 million loan tied to a large data center in New Mexico that is part of Stargate, the AI infrastructure venture backed by OpenAI, SoftBank Group and Oracle. The rest is linked to operators including Vantage Data Centers and CyrusOne, according to the reports.

Bids are due on November 19, with a minimum size of $10 million, according to the reports. ING has not publicly confirmed it is the seller, and no pricing has been disclosed.

Selling a loan book is routine for a large bank. What stands out is the timing and the asset class. The sale would be the first market price on construction-era AI data center debt since the November 3 midterm elections, which investors have read as raising the odds of tighter rules on data center development. A deep discount would set a benchmark that raises borrowing costs for every developer still financing a build.

Why GPU loans are getting hit hardest

Data center construction loans are secured by land, buildings, power connections and long-term leases, assets that hold their value for decades. GPU loans are secured by the chips themselves, and chips age fast.

Most hyperscalers depreciate servers over five to six years. Nvidia now ships a new accelerator generation about once a year, and each one sharply cuts the cost of a unit of compute. A loan sized against a chip’s purchase price can outlast the period in which that chip earns its expected rental income. Semafor described lenders’ central worry as loans whose lives will outlive the useful life of the technology.

Arougheti put the problem plainly on Ares’ recent earnings call. Without casting aspersions on GPU financing, he said, Ares would pass on such deals because the firm does not know how to model the depreciation curve.

Terms have already tightened. One industry analysis said lenders now advance about 55% to 65% of the cost of a new H100, down from 75% to 80% in 2024. A financing-industry guide put typical loan-to-value on newer B200 chips at 60% to 65%.

Who holds the risk

Banks are not the only lenders involved. Much of the GPU and data center debt written since 2024 sits with private credit funds, often in deals backed by chipmaker guarantees or long-term leases from hyperscalers. Reuters has described a common structure where chipmakers backstop senior debt used to buy their chips and guarantee lease payments, which moves part of the risk onto the chip supplier. Nvidia’s backstop commitments for OpenAI data centers are the largest example.

Other banks are also reported to be cutting exposure. HSBC and Barclays have reportedly looked for insurance on AI and data center credit. Industry reports say several large lenders are using significant risk transfers and discounted loan sales to cut AI-linked exposure.

Debt investors have been flagging stress for weeks. Bonds issued by GPU cloud operator CoreWeave sold off sharply in October during a broader chip and credit selloff. Investors have also been looking more closely at the off-balance-sheet structures hyperscalers use to fund AI buildouts.

The depreciation debate

The loan sales feed into a larger argument over how much AI hardware is really worth over time. Short seller Michael Burry has estimated that the gap between the useful lives companies report for their chips and their real economic lives could overstate industry earnings by about $176 billion from 2026 through 2028. Hyperscalers dispute that view and say older chips stay productive for inference after they are replaced for training.

Both sides agree the secondary market for used accelerators is still thin. Until there is a deep, liquid market for three-year-old GPUs, lenders have to guess at recovery values, and in a falling market they guess low.

How the secondary loan market works

When a bank makes a large loan, it rarely plans to hold the whole amount until it is repaid. Lenders routinely sell pieces of loans to other banks, funds and collateralized loan obligations (CLOs), which pool loans and sell slices to investors. Those sales free up the bank’s balance sheet for new lending and spread the risk.

The price a loan fetches on that market reflects buyers’ view of the chance they will be repaid in full. A loan trading at par, or 100 cents on the dollar, is considered safe. One trading in the 90s signals some doubt. Prices in the 80s or below usually mean buyers see real risk of loss. For a loan backed by GPUs, the main question is what the chips could be sold for if the borrower defaulted.

CLOs face extra limits. Most can hold only loans with a credit rating and an established arranger, and they cap how much they can hold in lower-rated debt. Industry reports describe at least one large GPU-backed loan, tied to about 36,000 Nvidia chips, that was being marketed without an underwriter or a rating, which keeps CLO buyers out. When the largest group of buyers cannot take part, prices have to fall far enough to attract funds that can.

That is why the discounts matter beyond the banks selling. The secondary price becomes the reference point for new deals. If existing GPU loans trade well below par, lenders on new deals will demand higher rates or larger equity cushions to match.

What it means

The AI buildout has relied on cheap, plentiful debt, and that debt is getting harder to find. The pullback is not a collapse, since data center construction loans backed by investment-grade leases still trade. But the market is now drawing a clear line between debt secured by long-lived infrastructure and debt secured by fast-aging chips.

Who loses. Neoclouds and smaller AI infrastructure firms that finance GPU purchases directly with chip-backed loans face higher rates, lower advance rates and fewer willing lenders. Developers still raising money for projects will feel any discount set in the ING sale. Chipmakers that have backstopped customer debt carry more contingent risk.

Who holds up. Hyperscalers that pay for chips from operating cash flow are largely insulated. Buyers with spare cash, especially large private credit firms willing to underwrite depreciation risk, can pick up loans at discounts that compensate them for it.

What to watch. The clearing price on ING’s portfolio after the November 19 bid deadline is the first hard data point. After that, watch whether other banks bring similar portfolios to market, whether the next big GPU-backed financing gets an underwriter and a credit rating, and how Nvidia’s upcoming earnings call handles questions about customer financing. If chip-backed credit stays tight, AI capacity growth will depend more on companies that can pay with their own cash.

Kurumi Kurumi · · 6 min read

Fluidstack Hits $18B Valuation on AI Data Center Boom

Fluidstack, an Oxford-founded neocloud backed by Google, has reached a roughly $18 billion valuation on the back of a ~$50B Anthropic deal and Google TPU hosting.

#Markets #AI #Data Centers
Kurumi Kurumi · · 5 min read

Vantage Data Centers Eyes $100B IPO or Sale

Vantage Data Centers is weighing an IPO at about a $100 billion valuation or an outright sale, in what would be the largest data center listing to date.

#Markets #Data Centers #IPO