How contracts, utilization, software, and power shape AI infrastructure financing

This RAISE Summit 2026 panel is about who pays for AI infrastructure. IREN and Crusoe build data centres. Modular makes software that runs across different chips. Ornn is building a market for GPU capacity. Alva Energy upgrades existing nuclear plants to produce more power.

AI infrastructure costs billions upfront. The panel explains what makes lenders say yes: customers who commit to buy, hardware that can be reused, and enough power to keep the machines running.

Jeremie Eliahou Ontiveros says neoclouds are deploying billions while power procurement raises capital needs on the utility side. He asks which businesses can self-fund and which need outside capital or large hyperscaler contracts.

Tim Davis describes Modular as software for running AI models across different chips. He says portability could make more accelerators liquid enough for debt financing.

Kent Draper says IREN owns its GPUs and liquid-cooled data centres. Its five-year Microsoft GB300 contract uses GPU-specific debt and a Microsoft prepayment. IREN later disclosed that those sources fund 96% of GPU capital expenditure. Its 3.31% all-in cost includes the zero-cost prepayment; the debt alone costs 6%.

Erwan Menard says Crusoe can sell long-term megawatt leases, GPU hours, or managed inference tokens from the same infrastructure. Wayne Nelms says Ornn aggregates spare compute and wants diversified demand to become financeable. James Krellenstein says Alva Energy proposes project-financed uprates at existing nuclear plants.

Crusoe and IREN use vertical integration to control power, construction, GPUs, delivery, costs, and margins. They finance each asset against its own life.

IREN's Microsoft GPU contract runs for five years. Draper expects its data centres to remain useful for more than 20 years and says lenders are beginning to value the uncontracted period. Nelms says forward prices for GPU capacity and residual hardware value could help finance later years.

Davis says CUDA software lock-in keeps operators tied to NVIDIA hardware. A portable layer could let them choose accelerators on price and move workloads across CPUs and GPUs.

He expects neoclouds to move from selling raw compute toward selling higher-margin model tokens. Another panelist says short contracts and sub-investment-grade customers have made that model difficult to finance, but reports growing lender interest in smaller AI labs, enterprises, and AI-native companies.

Krellenstein says nuclear demand is strong but supply is limited. He cites Amazon–Talen at Susquehanna, Microsoft's contract supporting the restart of Three Mile Island Unit 1, Meta-backed Vistra uprates, and Google's planned restart of Duane Arnold.

Alva proposes adding output at existing licensed reactors. Krellenstein says the sites can reuse transmission and avoid parts of the process required for a new plant. Fixed-price construction contracts would shift cost-overrun risk and support project finance. He gives uprates a three-to-five-year horizon and new builds a five-to-ten-year horizon.

Panelists reject treating every supplier-backed deal as circular. Their test is whether customers use the capacity. A chipmaker can support initial credit while the owner serves shorter contracts or on-demand users.

Public structures include credit support, residual-capacity purchase obligations, revenue sharing, and equity. The panel says utilization can keep support from being called. Its final question—whether a compute or token index can make short-term demand financeable—remains unanswered.

Power and data-centre projects are planned five or ten years ahead. Customers want compute now, and few can commit today to buy tokens in 2032 or 2033.

Nelms compares the gap with commodity markets, where producers sell long-term supply to finance infrastructure while demand stays closer to delivery.

Davis says alternative accelerators are difficult to sell when each vendor requires a different software stack. NVIDIA has CUDA, AMD has ROCm, Google has XLA and its TPU stack, and AWS has Neuron.

He says Modular can move workloads across chips. Better portability could increase utilization and reduce the investment risk of non-NVIDIA hardware. Qualcomm announced an agreement to acquire Modular in June 2026, with closing expected in the second half of the year.