Infra funds offer long-term GPU debt

AI data centers are starting to look less like tech projects and more like toll roads or power plants. That is the real story here. The hardware changes fast, but the campuses, substations, cooling systems, and power contracts do not. So the money behind them is changing too — away from ordinary bank loans and toward long-duration infrastructure capital that can sit there for 15 or 20 years while the site throws off contracted cash flow. (cnbc.com) ### Why are infra funds showing up here? Because AI campuses need infrastructure money. A modern hyperscale site is not just a warehouse full of servers — it is land, transmission access, backup power, water, fiber, and a very expensive shell built around fast-obsolescing compute. Brookfield has been explicit that it wants to provide AI infrastructure and even “GPU infrastructure as a service” under long-term contracts, while Digi(cnbc.com)d at data centers, fiber, and cloud assets. (datacenterdynamics.com) ### What problem are they solving? The mismatch between asset life and lender comfort. A data center shell can last decades. A utility interconnection can take years to secure. But a GPU generation may feel old in 5 to 7 years, which makes traditional lenders nervous about underwriting very long debt against AI-heavy facilities. That gap is exactly where infra investors think they can earn a premium — by financing the durable parts of the stack and leaning on long-term customer commitments to protect the downside. (cnbc.com) ### What do these deals actually look like? Usually not like a plain corporate bond. The common structure now is a project vehicle or joint venture that owns or develops the campus, raises private debt, and signs long-term leases or capacity offtake deals with a hyperscaler. The cloud company may keep the debt off its own balance sheet while still anchoring the economics through lease payments, reserved capacity, and sometimes guarantees. The BIS basically describes this as debt-like exposure wearing an infrastructure costume. (bis.org) ### Why does the lease matter so much? Because the lease is the bridge between fast tech and slow capital. Look at Applied Digital’s April 23 deal: a new U.S. investment-grade hyperscaler signed a 15-year lease worth about $7.5 billion for 300 MW at Delta Forge 1. That kind of contract gives lenders something they can model. It turns “maybe this AI demand sticks” into a cash-flow stream with dates, counterparties, and renewal options. (ir.applieddigital.com)lied-digital-announces-new-u-s-based-high)) ### Why not just let hyperscalers fund it themselves? They still fund a lot themselves — and aggressively. But the spending curve is getting absurd. Big hyperscalers are on track for roughly $650 billion to $700 billion of combined capex in 2026, much of it tied to AI infrastructure. Even for companies with huge balance sheets, offloading some of that into private vehicles preserves flexibility and spreads risk across debt investors, infrastructure funds, and insurers. (commercialsearch.com) ### Where does power fit in? At the center. The scarce asset now is often not the building — it is electricity. JLL’s 2026 outlook says power, not location or cost, is becoming the primary site-selection constraint, with multiyear waits for grid connections. That is why these financings increasingly bundle real estate with utility commitments, behind-the-meter generation, and long-term energy arrangements. You are not really financing a data hall. You are financing permission to consume huge amounts of power for years. (jll.com) ### What is the catch? The catch is that some of this risk is being pushed into corners of finance that are less transparent. BIS warns these structures can amount to “shadow borrowing” — economically debt, but sitting in project vehicles and private-credit structures rather than on hyperscaler balance sheets. If AI demand disappoints, GPUs age faster than expected, or refinancing windows shut, the stress lands on private lenders, insurers, and the banks backstopping those vehicles. (bis.org) ### Bottom line? This is what happens when AI stops being just a software story and becomes a civil-engineering story. Brookfield, DigitalBridge, and similar firms are trying to become the bond market for the physical AI stack — long-dated, asset-backed, and tied to power. If that model sticks, the next phase of the AI boom will be financed less like venture capital and more like infrastructure. (datacenterdynamics.com)a-centers/))