AI’s GPU boom is meeting a more expensive credit market

AI infrastructure companies are borrowing billions for GPUs just as rates and investor scepticism are making that debt harder to place. The practical twist is that many large borrowers have hedged much of their floating-rate exposure, so the immediate danger is less a sudden rate shock than the rising cost and uncertainty of financing the next wave of compute.

Watch Desk analysis

What happened

The Information reports that CoreWeave, Iren and Nscale are among the companies using delayed draw term loans to fund GPU purchases as hardware is bought and deployed. The structure lets borrowers secure a large pool of capital while drawing it over time, with interest generally beginning on the sums actually taken down.

Many of these facilities use floating rates linked to SOFR, plus a spread that varies with the borrower’s creditworthiness. The report says recent large GPU loans have carried spreads of roughly 2.25 to 5.5 percentage points above SOFR. Some borrowers have reduced their exposure through interest-rate swaps: CoreWeave hedged at least 75% of anticipated floating-rate borrowing under one $2.6bn facility, while its newer $8.5bn facility requires hedging of at least 95% of anticipated borrowing. Nscale’s disclosed facilities require hedges covering between 90% and 110% of projected floating-rate borrowing.

The Information also reports that Blackstone’s AI-focused investment chief, Jas Khaira, said the market still has not mapped out who will buy all the debt needed for the infrastructure buildout. He pointed to public investment-grade markets and private credit, but said financing would probably clear at a higher price as more supply arrives.

Why it matters

Hedging can soften the impact of higher benchmark rates, but it does not make AI infrastructure cheap. Borrowers still face loan spreads, hedging costs, equipment commitments and the question of whether customers will pay enough for the resulting compute. New facilities may also be more expensive if lenders demand greater compensation for the risk that AI demand, chip prices or data-centre schedules fail to match the forecasts.

The financing model matters because GPU capacity is being bought ahead of, or alongside, future revenue. CoreWeave has told investors that its delayed draw loans are structured to align debt payments with contractual customer payments, but that protection depends on those contracts and customers holding up. A server hall is not a cash machine merely because it contains very expensive silicon. It still needs utilisation.

Our read

This is a useful correction to the familiar AI infrastructure story, which often treats chips, power and data centres as the only constraints. Credit is another bottleneck, and the bill is becoming more visible as lenders face a flood of new issuance and unanswered questions about the ultimate returns on AI investment.

The immediate takeaway is not that GPU financing is about to collapse. Large borrowers have built in meaningful protection against rate rises, and delayed draw structures can match borrowing to deployment. But hedges only manage one risk. They do not settle who ultimately carries the debt, whether customer revenue will cover it or how much investors will charge for believing the forecasts.

What to watch

• Whether new GPU facilities carry wider spreads or tougher hedge requirements.
• How much borrowing is supported by signed customer contracts rather than projected demand.
• Whether private credit, bonds, vendor finance or equity absorb the next wave of AI infrastructure spending.
• CoreWeave’s and Nscale’s disclosures on drawdowns, customer concentration and debt service. The next phase of the AI race may be decided as much by lenders’ appetite as by accelerator performance.

Discussion spark: Should investors treat extensive interest-rate hedging as evidence that AI infrastructure debt is well managed, or as a distraction from the harder question of whether customers can pay for all that compute?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.