The Financing Problem Behind the AI Infrastructure Boom: What Lenders and Sponsors Need to Watch

AI infrastructure is accelerating at a historic pace with hyperscalers, neoclouds, and a fast-growing group of specialized compute providers racing to add data center capacity. For example, JPMorgan recently estimated capex for the five largest US hyperscalers will reach $697 billion in 2026, up $173 billion since the year began.¹ The size of the buildout is astounding, but the real story is how the build is being financed and the risks developing.

Companies can no longer self-fund this pace of expansion, increasingly turning to debt, structured financing, and GPU-backed collateral to bridge the gap. Even if the most bullish assumptions about AI demand prove true, a brewing timing gap between when capital is deployed and when cash flow is being generated is creating real risk. For lenders, sponsors, and operators across the ecosystem, both risks, not just the demand debate, are important in underwriting new obligations.

The shift to debt

The five largest hyperscalers issued roughly $121 billion in US corporate bonds in 2025 alone, more than four times their 2020-2024 annual average of $28 billion, and issuance in the first half of 2026 has already exceeded the full prior year. An example of this is revealed in a recent article that stated how Bank of America expects the group to borrow around $140 billion annually over the next three years; JPMorgan projects roughly $300 billion a year in AI and data center related financing across the next five.²

Layered beneath the investment-grade names is a fast-growing shadow structure of special-purpose vehicles, private credit, and GPU-collateralized loans funding data centers the hyperscalers lease but don't own outright. These obligations show up as future lease commitments rather than on-balance-sheet debt, and some analysts estimate they now rival the hyperscalers' own reported debt loads in scale.³

This shift does not make the buildout unsound; however, it does raise the stakes for underwriting. As more capital is supplied by lenders and investors that depend on predictable project economics, the consequences of delayed energization, slower tenant ramp-up, or declining equipment values become more significant.

Even the bull case has a timing problem

Assume, for a moment, that every bullish premise about AI holds. Many industry leaders believe that we're still early in the cycle: Jevons' paradox keeps demand climbing as compute gets cheaper and more capable, frontier labs are closing in on more general capability, and robotics, autonomous vehicles, and "world model" applications eventually require an order of magnitude more compute than today's chatbots and copilots.⁴

Even under that scenario, financing risk doesn't disappear and could amplify. Robotics and physical-world AI are still pre-revenue categories being built on multi-year timelines, while business and consumer adoption of even today's models is still early. Capital providers should look beyond AI's expanding market potential and focus more on near-term cash generation capability.

Many current financing structures appear to underwrite long-term demand as though it were already dependable near-term cash flow, which is a mismatch that lenders, sponsors, and operators need to address.

Where the gap comes from

Several structural bottlenecks can delay the point at which capital outlay becomes revenue and cash flow, regardless of how the demand debate resolves:

• Grid interconnection. US interconnection queues held more than 2,600 gigawatts of proposed generation and storage as of early 2026, with wait times commonly running 24 to 72 months and stretching to 5-7 years in constrained markets like ERCOT and PJM. A fully financed and built data center that can't get energized cannot serve tenants, generate revenue, or operating cash flow.⁵
• Power and equipment scarcity beyond the grid connection itself. Transformer supply is a specific choke point, with China controlling roughly 60% of global manufacturing capacity for the specialized steel involved.⁶ Memory and server components are tightening too, with reports of sharp AI-driven price increases for server DRAM and NAND as chipmakers prioritize AI customers over other buyers.⁷ These constraints can increase project costs, extend deployment schedules, and pressure expected returns before a facility is operational.
• The "good enough" substitution risk. In theory, as open-source model capabilities improve and model sizes shrink, companies can be just as effective across many use cases running these models on-premise or on modest self-hosted infrastructure, rather than paying for API calls to frontier labs. As that gap continues to close, a meaningful slice of anticipated frontier-model demand, and the compute built to serve it, may migrate away from the large, expensive clusters the buildout assumes will stay full.
• Permitting, regulation, and citizen backlash. This has moved from a niche concern to an active drag on deployment timelines. Recent polling has found a strong majority of Americans oppose a data center being sited near them, in some surveys exceeding opposition to a nearby nuclear plant.⁸ Local and state pushback delayed or blocked at least 75 projects worth roughly $130 billion in the first quarter of 2026 alone,⁹ and both Texas and New York have since imposed statewide restrictions on new development.¹⁰ For a financed project, every month of delay can mean additional carrying costs without without matching cash flow.
• Leapfrog risk from orbital compute. Google, Nvidia-backed Starcloud, and others are actively developing space-based data centers that could someday sidestep terrestrial land, power, and permitting constraints entirely. Starcloud already flew a prototype satellite carrying an Nvidia H100 GPU in late 2025, and Axiom Space put the first dedicated orbital data center nodes into orbit in January 2026. This is earlier-stage than the terrestrial buildout, but it's no longer purely theoretical. Google's own Project Suncatcher prototypes are slated to launch in early 2027.¹¹ Taken together, it's a real argument for caution around the terminal value embedded in today's terrestrial buildout assumptions.

The bear case for AI demand doesn't have to be right for the AI investment bubble to burst, gaps between investment outlay and cash flow generation or a shifting AI landscape could be the catalyst.

A caution on collateral

For lenders and investors, the sharpest version of this risk sits in the hardware itself. GPUs are increasingly being used as loan collateral. For example, CoreWeave pioneered the model and has since raised tens of billions in GPU-backed debt and equity, but the asset doesn't behave like the real estate or utility-grade equipment private credit is used to financing. ¹²

The issue is that collateral with a short competitive life, uncertain residual value, varied views on depreciation accounting treatment, and pricing influenced by constrained supply should be underwritten differently from commercial real estate. That may mean higher yields, shorter maturities, more conservative advance rates, and greater scrutiny of utilization assumptions and customer commitments. More broadly, lenders and investors should scrutinize any buildout underwriting that leans on today's extraordinary compute margins persisting for the life of the loan.

Watch the second-order exposure

The most visible risk sits with the data centers themselves, but a wider circle of "pick and shovel" businesses have quietly built significant exposure to the same cycle. Global construction capacity for data centers is stretched thin, with over 70% of markets reporting contractor capacity as tightening or overstretched and specialist trades in particularly short supply.¹³

The large, diversified public general contractors remain genuinely bullish, with strong backlogs to show for it. The more concentrated exposure may sit one or two tiers down: smaller subcontractors, specialty engineering firms with limited data center experience, and equipment rental and leasing companies that have concentrated capital or headcount around a single sector's cycle. If the buildout timeline stretches or its financing tightens, these businesses have far less balance sheet cushion to absorb the disruption than the hyperscalers underwriting the projects themselves.

The AI infrastructure buildout is not necessarily overextended, and the underlying compute demand may be there. But its financing should account for a wider range of timing, execution, utilization, and collateral-value outcomes than the market may be pricing today.

Related Reads:

• How Cybersecurity and Compliance Build Customer Trust in the Age of AI
• Enterprise AI Tools Are Now a Malware Distribution Channel. Here's What You Should Do About It.
• AI in Internal Audit and SOX: Applying Methods that Work