The debt behind AI: off-balance-sheet SPVs, bonds…

The AI debate has so far played out on two registers: stock-market valuations and the circular revenue looping between a handful of players. A third, quieter register actually decides how sound the structure is: how the build-out gets paid for. As long as the tech giants funded their data centers out of their own cash, the risk stayed contained on their balance sheets. That is no longer the case. In 2026, hyperscaler capex swallows almost all of their operating cash flow, and the shortfall goes looking for funding in the debt markets. Morgan Stanley expects close to $570 billion of AI-related issuance this year; by late 2025 this debt was already the single largest slice of the investment-grade bond market. This piece follows the plumbing of that debt-financed boom: the off-balance-sheet Meta-Blue Owl structures, a bond market beginning to choke, private credit and insurers at the end of the chain. And the counter-argument, which is not a weak one.

When capex outgrows cash flow

For a decade, the big cloud players funded their data centers the way any highly profitable company funds its growth: out of its own earnings. That is what long defused the comparison with the 2000 telecom bubble, where the infrastructure was built on credit. That dam has just broken. According to Morgan Stanley, hyperscaler investment is on track in 2026 to consume close to 100% of their operating cash flow, against a ten-year average of 40%. The rest has to be borrowed.

The mechanism is arithmetic before it is speculative: when capital spending exceeds what operations generate, the difference is financed by issuing debt or equity. The hyperscalers have chosen debt, and at scale. Again per Morgan Stanley, global AI-related debt issuance should reach close to $570 billion in 2026, more than double 2025. By late October 2025, the outstanding stock of this debt had already passed $1.2 trillion, becoming the largest segment of the investment-grade market and overtaking US banks as the biggest sector in the JPMorgan US Liquid index. A regime change: AI is no longer only a story of expensive stocks, it has become a story of credit.

Off-balance-sheet: the lesson of the Meta-Blue Owl deal

The most discreet form of this debt shows up on no tech balance sheet. On 21 October 2025, Meta announced a joint venture with funds managed by Blue Owl Capital to build its Hyperion megacampus in Louisiana, for a development cost of roughly $27 billion. The split is the heart of the design: Blue Owl’s funds own 80% of the structure, Meta only 20%. Meta contributes the land and construction-in-progress, collects a one-time distribution of about $3 billion, keeps operational control through a lease, but leaves the debt outside its own accounts.

That debt is enormous. According to trade press reporting, PIMCO anchored a bond tranche of roughly $26 billion, rated at the top of the quality spectrum and amortising over a very long maturity, described as the largest private-credit deal ever closed. All of it through a special-purpose vehicle that carries the leverage in Meta’s place. What Meta keeps, on the other hand, is a residual-value guarantee for the first sixteen years: if the lease is not renewed, Meta commits to a capped cash payment. In other words, the economic risk has not entirely left the house; it has merely changed accounting line. As the real-estate press put it, these structures let companies fund tens of billions of infrastructure with debt that will never officially appear on the balance sheet.

The appeal for the issuer is twofold: preserve the group’s credit rating and deconsolidate a colossal amount of borrowing. The drawback for the observer is symmetric: it makes the sector’s true leverage harder to measure, exactly the kind of opacity that runs through the wider migration of credit risk beyond the regulatory gaze.

The bond market starts to choke

Alongside the off-balance-sheet channel there is the public market, and it is running flat out. In the first seven months of 2026 alone, AI-related bond issuance topped $250 billion, of which $218 billion in investment-grade debt and $31.9 billion in high yield, the latter almost entirely ($27.9 billion) earmarked for data centers. The acceleration is brutal: across the whole of 2025, AI-linked high yield weighed only $14.1 billion. Amazon placed $25 billion in July, Oracle turned the bond market into its main funding lever, with fiscal-year capex above $55 billion and borrowing plans that sent its share price sliding once investors took their measure.

The point of stress shows up at the bottom of the credit stack. CoreWeave, a cloud operator specialised in AI compute, issued six-year notes in June 2026 at around 9.6%, a cost that betrays how the risk is perceived, and its paper then traded below par, around 96.5. The word used by credit analysts is telling: buyside “indigestion”. This is not yet a market freeze, but the first sign that investor appetite, however voracious, has a limit. That is the crux: investment-grade AI debt places without difficulty, but the most fragile segment, backed by fast-depreciating equipment, is already testing the edges of demand.

Private credit and insurers at the end of the chain

Following the debt all the way through leads to whoever holds it. A growing share sits not in liquid bond funds but in private credit, whose corporate lending assets are expected to exceed $2 trillion in 2026 according to Moody’s. This is the compartment that absorbed the Meta deal, and the one funding a share of the riskiest build-out. The problem the agency flags is not the volume but the legibility: light covenant documentation, interest paid in kind (PIK) rather than in cash, loans secured against fund net asset value, layered leverage at the vehicle level. All structures that obscure real leverage rather than remove it.

Behind private credit there is often the life insurer, the big final buyer of these long-dated assets. And that is exactly where Moody’s locates one of the contagion channels in a shock. In a January 2026 analysis, the agency maps what would happen if AI-related valuations fell 40%: private-credit managers forced to renegotiate to avoid defaults, exposed insurers, a wealth effect on consumption through falling equities. On top of that sits a risk specific to data centers: the performance of securitisations backed by these assets depends on tenant demand for compute capacity. If AI adoption plateaus or shifts direction, over-investment is paid for in half-leased campuses. The full chain, from the debt-financed chip to the insurance policy, is longer and more opaque than the equity surge alone suggests.

Vendor financing closes the loop

One last link connects this debt-financed boom to the other big debate of the moment. Part of the demand that makes this debt sustainable is itself manufactured by the suppliers. Nvidia has committed more than $40 billion of equity stakes in 2026, including around $30 billion in OpenAI, a customer that uses those funds to buy, directly or not, Nvidia chips. This vendor financing echoes the Nortel-Lucent episode of the telecom bubble, when equipment makers lent to their own customers to prop up sales. We laid out this mechanism in the anatomy of AI circular financing and through the BIS warning on the boom’s financial fragility. The point that matters here is the joint: if a fraction of demand is recycled capital rather than organic need, then the debt issued to serve that demand rests on a narrower base than it appears.

The counter-argument

The concern is real, but the catastrophist reading runs into several solid facts. First, most of this debt is investment grade. Of the $250 billion issued in 2026, $218 billion is investment grade, carried by companies whose cash flows are, themselves, very real and among the highest in the world. Reaching straight for the 2000 telecom comparison ignores that Meta, Microsoft, Amazon or Alphabet generate profits the over-leveraged operators of that era never came close to.

Second, borrowing while it is cheap is a rational balance-sheet decision, not necessarily a headlong rush. Diversifying funding sources, issuing long-dated fixed-rate debt, deconsolidating infrastructure through a long-term partner: these are humdrum financial-engineering practices for firms of this size. The Meta-Blue Owl deal, precisely, rests on amortising, well-rated debt backed by a top-tier tenant, a long way from fragile credit.

Third, the dividing line is sharp. It separates debt backed by established cash flows, which places without difficulty, from bottom-of-the-range debt backed by fast-depreciating equipment and still-hypothetical revenue. It is this second compartment, a minority by volume, that shows the first signs of strain. Systemic risk does not arise from debt financing as such, but from its concentration on a handful of issuers and from the difficulty of measuring real leverage once it migrates off balance sheet and into private credit.

What to watch

Three signals beat one forecast. The first is the behaviour of high yield backed by data centers and the level of credit spreads: a widening would say that investor demand, voracious today, is closing up. The second is the performance of data-center securitisations, a thermometer of real demand for compute capacity. The third is the marking of positions in private credit, the least liquid and most opaque link. For the reading tools, our guides on private credit, CLOs and leveraged loans and credit spreads give the grid. The real question is not whether AI will deliver on its technological promises, but whether the debt building it will hold long enough for it to.

Sources

Morgan Stanley, forecast for AI-related debt issuance (June 2026), via Yahoo Finance: capex ≈ 100% of operating cash flow, ≈ $570bn of issuance in 2026, $1.2trn outstanding by late 2025.
AI-related bond issuance in 2026, Yahoo Finance: investment-grade / high-yield split, CoreWeave, buyside “indigestion”.
Meta Platforms, press release on the Hyperion joint venture with Blue Owl Capital (21 October 2025).
Data Center Dynamics: the debt tranche anchored by PIMCO.
Bisnow: the off-balance-sheet structure.
Yahoo Finance / Reuters: Oracle’s capex and borrowing plans.
Moody’s, 2026 private credit outlook and data-center credit risk: assets, opaque structures, contagion channels.
Bloomberg: cross-commitments between Nvidia, OpenAI and cloud providers (circular financing).