Imagine a row of hulking server halls lit like airplane hangars, their power bills measured in the millions and long-term contracts inked like mortgages. That image — of companies racing to lock in compute and electricity for artificial intelligence — helps explain a quiet but important reshaping of corporate balance sheets.
Lenders and bond investors are waking up to two linked realities. First, tech companies are borrowing and committing capital at unprecedented scale to build AI-optimized data centers. Second, much of that borrowing no longer shows up the way it used to: about $120 billion of AI data-center obligations have been pushed into leases, joint ventures and project-finance vehicles, keeping headline leverage lower while leaving real, enforceable claims on cash flows.
New plumbing for old problems
Rather than issuing traditional unsecured corporate debt, firms are packaging site-level obligations into special-purpose vehicles (SPVs) tied to power purchase agreements (PPAs), take-or-pay capacity contracts, or long-term leases. That approach preserves ratings room and frees up borrowing capacity for other strategic moves. It also lets companies share construction and operational risk with utilities, landlords and specialist operators.
But these structures are not magic. Take-or-pay contracts still obligate a company to pay for capacity whether it uses it or not; equipment refresh cycles, cooling and power costs still eat margins; and the protections in project documents determine who eats the losses if utilisation falls. For bond investors accustomed to judging leverage by debt on the balance sheet, the risk is that headline metrics understate the true strain on future cash flows.
Why investors are getting nervous
Debt markets have noticed. The recent flood of borrowing tied to AI — from corporate bonds to structured financings — has helped push global tech debt issuance to record territory. Investors are tightening their scrutiny: they want full transparency on lease terms, step-in rights, residual-value assumptions and how these deals sit relative to corporate credit lines.
The mechanics matter in ways that show up in pricing. Asset-backed and secured structures can trade differently from vanilla unsecured bonds; hybrids and subordinated tranches carry different default dynamics. And when calendar congestion is intense, issuers pay higher new-issue concessions to get deals away, which can widen spreads and create selective entry points — if you know where to look.
Supply, demand and the calendar squeeze
Strategists increasingly expect heavy investment-grade issuance as companies refinance cheap pandemic-era paper and raise fresh capital for AI expansion. That supply can push spreads wider and make long tenors more costly. In practice, issuers are likely to lean into secured or asset-linked formats — which shifts risk onto investors who must read the fine print.
For portfolio managers, the advice is practical: insist on transparency, stress-test cash-flow models against slower AI adoption and higher power costs, and watch the ratings mix. A crowded issuance calendar can deliver yield opportunities, but those gains often come with complexity.
The energy angle — not just chips and racks
AI’s appetite for electricity has turned power contracts into the new covenants. Projects frequently tie financing to grid connections and long-term PPAs, and regions with constrained supply face potential delays that can reverberate through returns. That’s why some companies are exploring unconventional sites and concepts to secure capacity — from brownfield grid upgrades to far-out ideas like orbital data centers.
Even if that sounds futuristic, initiatives to rethink where and how compute sits are relevant. For investors, the message is: grid risk is credit risk. A project delayed by permitting or hit by a spike in energy prices can compress expected cash returns and stress sponsors.
What to look for in documentation
Not all off-balance-sheet debt is the same. Helpful questions to ask before buying into a structured deal include:
- Who ultimately guarantees the SPV and under what conditions?
- Are revenue streams truly ring-fenced, and how senior are the claims?
- What termination, step-in and residual-value protections exist for sponsors?
- How sensitive are cash flows to utilisation, energy prices and equipment refresh cycles?
Investors should also prefer transparent issuers with demonstrable asset-backed cash flows — utilities, regulated operators, and landlords with long contracts tend to offer more predictability than speculative build‑and‑hold projects.
Where AI demand could go next
The infrastructure story is broader than a single rack or campus. Firms are rethinking latency, regional footprints and redundancy — all of which influence financing strategies. Some of the biggest demand drivers are the same features that made generative AI so pervasive: models integrated into productivity tools and search that chew through storage and compute. That trend helps explain why companies are locking in long-term capacity across cloud, colocation and even experimental deployments like space-based concepts such as Google's Project Suncatcher.
Meanwhile, AI features that splice into everyday workflows — exemplified by advances in model-driven search and document grounding — are amplifying the need for always-on compute and storage, which in turn feeds these financing cycles. See how enterprise AI tooling is changing expectations in applications like email and drive search in coverage of Gemini Deep Research.
Debt investors are right to be cautious. The math behind many of these deals is sound only if utilisation, power and upgrade costs behave. When they don’t, the contractual commitments remain.
Some parts of the market will benefit — specialist lenders, project financiers and issuers that price risk transparently. Others will find surprises in footnotes. The coming year promises more paper, more structured formats, and a steeper learning curve for investors who thought AI was just about models. In this market, covenants and contracts may matter more than marketing slogans, and a clause buried in an SPV agreement could determine who wins and who takes the loss.