Ask a CFO at any of the five big hyperscalers how they’re paying for the largest capital spending programme in corporate history, and you’ll increasingly hear a version of the same reassuring answer: most of it isn’t actually on our books. Meta’s $27 billion Hyperion data centre in Louisiana sits inside a special purpose vehicle in which Meta owns just 20 percent. Oracle, Amazon, Alphabet and Microsoft have all leaned on similar structures, and the industry now discloses roughly $662 billion in data-centre lease commitments that haven’t even started yet, alongside the debt already issued. It sounds prudent. It is, in a narrow accounting sense, true. It is also the wrong thing for a CEO or board member to find reassuring, because the debt hasn’t disappeared. It has simply moved somewhere harder to see, and the people now holding it didn’t necessarily choose to.
That distinction, not the size of the AI capex number itself, is the story every executive should be tracking this year, whether or not their company has anything to do with AI infrastructure.
The scale, briefly, because it matters
The big five hyperscalers spent roughly $380 billion on capex in 2025 and are guiding to something closer to $700-785 billion for 2026, with a run rate approaching $1 trillion by 2027. Corporate bond issuance from this group alone hit $121 billion in 2025, against a five-year average before that of $28 billion. By the halfway mark of 2026, global AI-related debt issuance was already running at roughly four times the 2025 pace. Morgan Stanley’s research puts cumulative AI capex at $2.9 trillion through 2028, against operating cash flow of about $1.4 trillion, leaving a financing gap of roughly $1.5 trillion that has to come from somewhere other than free cash flow or public equity.
Private credit is filling a large share of that gap: outstanding private credit exposure to AI-linked firms had already passed $200 billion by late 2025, up from origination volumes in the low single-digit billions a decade earlier, and is projected to supply roughly $800 billion of the financing gap through 2028. Layer on GPU-collateralised loans (Nvidia’s $6.3 billion facility with CoreWeave, an $8.5 billion investment-grade GPU-backed securitisation in 2026), asset-backed and commercial mortgage securitisation of data centres, and sovereign vehicles like Saudi Arabia’s PIF committing over $100 billion to compute capacity, and you have a financing stack for the AI economy that barely resembles the one that funded the last two decades of enterprise technology.
Why the structure, not just the size, is the risk
None of this is illegal, and most of it isn’t even unusual by the standards of infrastructure finance, which has used project-level SPVs for pipelines and toll roads for decades. What’s different here is who ends up holding the paper, and how confident anyone can be in the assumptions underneath it.
The Bank for International Settlements said as much in a note earlier this year from economists Egemen Eren, Ingomar Krohn and Karamfil Todorov, warning that these off-balance-sheet structures are “strengthening links between hyperscalers and non-bank investors.” Translated out of central-bank language: the entities now carrying AI infrastructure risk are increasingly insurers, pension funds and private credit vehicles that ultimately draw on retirement savings and insurance premiums, not the regulated banks that carry capital buffers and face stress tests for exactly this kind of concentration. A Chicago Booth working paper by Stefan Hepp models what happens if AI valuations get re-rated on disappointing returns: a $10-14 trillion hit to AI-linked equity value, translating into $60-140 billion of realised credit losses, absorbed “largely by fund investors and recognised gradually” rather than in one visible event. That gradualism is a feature of the structure, not a comfort. It means the losses show up in fund marks and insurer solvency ratios quietly, long after the decisions that created the exposure were made, and long after the executives who arranged the financing have moved on to the next deal.
Then there’s the circularity, which any competition lawyer or securities litigator will tell you is the part that should worry boards most. Nvidia takes equity stakes in AI labs and extends credit to cloud providers; those cloud providers commit hundreds of billions in future spend to buy Nvidia chips (OpenAI’s roughly $300 billion commitment to Oracle and separate $38 billion deal with Amazon are the two most-cited examples); the resulting revenue commitments are then used to underwrite the credit ratings on the debt that finances the data centres those chips sit in. Quinn Emanuel’s litigation practice, in a client note on the sector, points out that credit ratings on under-construction AI facilities increasingly rest on the creditworthiness of a small number of tenants rather than the economics of the project itself — a dynamic with an uncomfortable resemblance to the tenant-concentration and rating assumptions that unravelled in mortgage securitisation markets after 2008. Add that AI industry revenue in 2025 was estimated at roughly $60 billion against capex of around $400 billion, and the gap between the growth story being financed and the cash actually being generated looks wide enough that a modest disappointment in AI monetisation, not a catastrophic one, is enough to test these assumptions.
What this means if you’re not building data centres
The instinctive response for a CEO outside the AI infrastructure business is to treat this as someone else’s balance sheet problem. That’s a mistake, for two reasons. First, private credit has become a mainstream allocation for corporate pension plans and insurance-linked treasuries precisely because it has been marketed as a stable, low-volatility source of yield; boards approving those allocations rarely ask how concentrated the underlying loan book has become in a single sector’s infrastructure bet. Second, and more strategically: when repricing eventually comes, it won’t arrive as a uniform shock. It will arrive as a sudden divergence in who can still raise capital on reasonable terms and who can’t — and that divergence will map onto genuine differences in AI unit economics, not just balance sheet cleverness. The companies whose AI investments are backed by demonstrable revenue, not projected revenue, will keep raising capital. Everyone else will discover that off-balance-sheet financing was cheap only because nobody was pricing the risk correctly, and that the bill, when it comes, arrives with less warning than a bank loan would have given.
India’s version of this story hasn’t been written yet
India’s own data-centre boom offers an instructive contrast, mostly because it hasn’t yet developed the plumbing to replicate the American model. Reliance’s roughly $20-30 billion Jamnagar AI campus and the Adani-Google partnership committing around $15 billion through 2030 are being funded largely through strategic equity, sovereign-adjacent capital and corporate balance sheets, not securitised private credit — India’s domestic private credit market, at roughly $3.5 billion, is a rounding error next to the $200-billion-plus pool now financing American AI infrastructure. That is not a permanent state of affairs. Global private credit managers are actively expanding into India as regulatory tightening pushes some lending away from NBFCs, and GIFT City is being built, in part, to host exactly this kind of fund structure onshore. India has a genuine, narrow window to write rules for tenant-concentration disclosure, collateral valuation and non-bank exposure limits before its own infrastructure financing gap gets filled by instruments nobody in New Delhi or Mumbai has fully stress-tested yet — a chance the US financial system, moving at the pace this cycle has demanded, has already lost.
The uncomfortable question for any executive to sit with isn’t whether AI infrastructure spending is justified. It’s who actually holds the risk if the revenue doesn’t arrive on schedule, and whether that answer would change how confidently they’re building on top of it.
