As lenders line up to finance the ongoing data centre boom, one cornerstone of artificial intelligence development is being overlooked.

Fitting-out a data centre with the latest graphics processing units (GPUs) – the hardware essential for training large language models – can now cost more than constructing the building itself.

That’s leaving some GPU buyers in a bind. While the likes of Alphabet and Meta can draw on vast balance sheets to acquire the latest Nvidia chips, not all are so lucky.

“There’s a massive gap in the market for structured credit for GPUs,” said David Lindström, partner at Optio Investment Partners.

The Stockholm-based private credit firm is looking to sign strategic partnerships with neoclouds – firms that rent out GPU capacity to AI developers – to provide asset-backed financing. Ticket sizes would range between USD 50m-USD 400m with terms ranging from three years to five years and be secured against the GPUs and underlying rental contracts.

It’s an area where banks play a relatively limited role, in part due to the hefty risk-weights they face in GPU lending.

Data centre tenants, rather than the data centre owner, are typically on the hook for buying and installing GPUs. Building one gigawatt of AI data centre capacity can cost around USD 35bn, with GPUs comprising around 39% of total capex, according to broker Bernstein.

Hyperscalers like Microsoft and Meta can tap billions in operating income and free cashflow to buy GPUs, and lock-in purchases through multi-year commitments. Larger neoclouds, meanwhile, often turn to GPU-collateralised debt. Last year, CoreWeave turned to a syndicate of banks to finance a USD 2.6bn delayed draw term facility, in part to acquire the latest Nvidia chips.

Smaller neoclouds, meanwhile, often lack the track record, name recognition, and even the credit rating to secure long-term financing for GPUs. Most tend to finance the eye-watering capital expenditure required to purchase the hardware using equity.

It’s a practice Lindstrom views as unsustainable, given the data centre GPU market is on course to grow to USD 202bn by 2032, up from USD 40bn in 2024, according to SNS Insider – a fourfold increase over a decade.

“Unless we find a way to structure a product that works long term, [the capital market] is going to run out of equity to finance the GPUs that go in them,” Lindström said. “With proper structured products, such as with securitisation, you can allocate appropriate risks to appropriate investors at appropriate attachment points.”

Optio started life providing billions in off-balance sheet structured financing to Volvo to fund part of the car manufacturer’s fleet. The private credit platform is also exploring lending against battery storage, heat pumps and solar panels.

One hurdle when applying securitisation to GPU financing is gauging what will happen once the neocloud’s initial contract with its end-customer ends. Most neoclouds tend to sign offtake contracts with AI developers between 12-18 months, but rating agencies tend to apply steep haircuts to on-demand revenue forecasts to be generated after the contractual cashflow ends, which can make structuring a rated transaction tricky.

Still, neoclouds’ business of renting out GPUs by the hour is a profitable one. GPUs are in short supply, with lead times for the Nvidia’s H100 and AMD MI250 of up to a year, according to AI service provider Clarifai. While the cost of renting GPU capacity has fallen from its 2024 highs, it has bounced back this year on the back of demand from AI developers.

But while GPU prices are sky-high now, they can fall sharply once a new edition is released and may quickly become outdated. But Lindström argues that the current debate around obsolescence risk misses the range of ways that GPUs can be used after they are no longer suitable for training large language models.

While the latest GPUs tend to be used to train LLMs for around 18 months, they can be reused for other purposes afterwards. Many are used for inference – the day-to-day operations of an AI model – or sometimes in place of CPU capacity, he said.

Like any fast-growing sector, the GPU financing market endures plenty of teething pains. While an aluminium smelter hedges its electricity cost through the futures, neoclouds and their end-users cannot do the same with GPU capacity.

“GPU financing is where oil was before we had a futures market for oil,” Lindström said, predicting that a futures market for GPU capacity could be where the market is heading next.