Everything About AI
NextFin News - Artificial intelligence's builders are staring at a refinancing wall: roughly $1 trillion of debt tied to the AI build-out is projected to be outstanding by 2027, and the cheap credit that funded the boom is no longer on offer. Goldman Sachs estimates AI-related debt will approach $1.2 trillion in 2027, up from an expected $750 billion by the end of 2026, while credit spreads across AI-linked debt have widened materially as investors demand a higher premium for the risk. Hyperscaler investment-grade bond issuance is on track to reach about $250 billion this year and climb to $400 billion in 2027. The question is no longer whether AI will transform the economy - it is whether the financing structure built to pay for it can survive the bill coming due.
This is not a story about AI demand failing. It is a story about maturity structures, leverage opacity, and the moment when a growth narrative collides with a debt schedule. The companies that locked in long-duration, fixed-rate funding early will refinance comfortably. The ones that borrowed short against assets that do not yet generate cash face a different 2027.
The Debt Wall Takes Shape
The scale of the borrowing is unprecedented for a single technology theme. In 2025, the five largest hyperscalers - Alphabet, Amazon, Meta, Microsoft and Oracle - funded roughly 26% of their capital expenditure with investment-grade bonds, according to Goldman Sachs. That share is projected to rise to 33% this year and 35% in 2027, when $400 billion of new investment-grade issuance is expected against an estimated $1.14 trillion in capex. Those figures exclude another roughly $412 billion of AI-linked debt issued in 2026 by companies outside the hyperscaler group, spanning software, semiconductors and data-center financing.
The acceleration is visible in the raw pace of issuance. Barclays strategist Venu Krishna calculates that total capital raised by hyperscalers, data-center special-purpose vehicles and neo-clouds - across investment-grade debt, high-yield debt and equity - more than doubled from $172 billion in 2025 to $346 billion so far in 2026. Through early June, the five hyperscalers alone had raised about $159 billion in bonds, a 47% increase over everything they issued across all of 2025. Alphabet became the first technology company in decades to tap the century bond market, selling $31.51 billion across a global offering in February, including the tech industry's first 100-year bond since Motorola's 1997 issuance.
Moody's Ratings puts the broader infrastructure bill at at least $3 trillion over the next five years, spread across servers, computing equipment, data-center facilities and new power capacity. The ratings firm also flagged a shadow layer of financing: roughly $1.2 trillion of off-balance-sheet arrangements, of which more than $820 billion is tied to data centers still under construction. A large share of the AI build-out is being paid for with debt that does not sit on the balance sheets of the companies whose names are on the data centers.
The structure of that debt matters as much as the size. Much of it is raised through special-purpose vehicles that sign triple-net leases with hyperscaler tenants - agreements in which the tenant covers rent, taxes, insurance and maintenance. The SPV then issues bonds or takes out construction loans secured by the lease. On paper, the arrangement looks bankable: a creditworthy tenant, a long lease, steady rent. In practice, the risk migrates to three pressure points - the tenant's right to terminate if the facility is not delivered on time, the SPV's ability to refinance when the construction loan matures, and the residual value of a building designed around a specific generation of AI chips.
Why Borrowing Costs Are Rising Now
Three forces are converging to make refinancing more expensive. First, supply: an unprecedented volume of credit is hitting the market at once. Second, investor fatigue: after absorbing the flood, buyers are demanding compensation. Third, the macro backdrop: Bank of America expects the Federal Reserve to raise rates twice more - in December 2026 and March 2027 - holding at a terminal range of 4.25% to 4.5% through the end of 2027.
The fatigue is already showing up in prices. S&P Global counts $225 billion of bonds issued by hyperscalers and related entities such as Nvidia so far this year, putting them on pace for a record 2026 haul - but warns that hyperscalers are now paying a higher premium relative to risk-free yields. Krishna put it directly: "Over the past two quarters, with unprecedented credit supply hitting the market, credit spreads have widened materially for investment grade and high yield bonds, as well as convertible bonds, across hyperscalers, data centers, and neo-clouds."
The stress is most acute at the lower-rated edge of the market. CoreWeave's junk bonds saw yields surge to about 15% in October, a level that typically signals distress, after the data-center operator's co-founder said the company was exploring investment-grade debt markets. Oracle's Project Jupiter - an $18 billion loan package tied to its New Mexico data-center campus - has been quoted at 89 to 91 cents on the dollar as arrangers including Santander and Jefferies struggle to place the debt with a broader pool of investors. Oracle sits one notch above junk following a July downgrade by S&P.
The Oracle case illustrates how project-specific risks compound financing risk. The Green Chile natural-gas pipeline meant to power Project Jupiter was denied a permit for a second time by the New Mexico State Land Office, pushing its in-service date to February 1, 2027. Oracle raised $43 billion of debt and $5 billion of equity in fiscal 2026 and expects roughly $40 billion more in fiscal 2027, while capital expenditure jumped to $55.7 billion from $21.2 billion a year earlier - producing negative free cash flow of $23.7 billion. S&P has noted that roughly half of Oracle's $638 billion in remaining performance obligations comes from a single customer: OpenAI. Concentration of that magnitude is precisely what lenders scrutinize when they price refinancing risk.
Power is the second constraint feeding back into credit. Morgan Stanley warned in August that "capital alone no longer clears a site," arguing that community resistance has moved from a public-relations challenge to a material development risk. The bank projects a 38-gigawatt U.S. data-center power shortfall between 2026 and 2028, with some regions facing grid-interconnection waits of five to seven years. A data center without power is not an income-producing asset - and debt underwritten against its future rent has no cash flow to service it.
The Mechanism: Why This Cycle Is Different
Every capital-expenditure boom ends in a hangover; what matters is the transmission channel. The AI build-out differs from prior technology cycles in three structural ways.
Duration mismatch. Data centers are 15-to-30-year assets, but much of the debt funding them carries far shorter maturities and floating or near-term reset rates. Construction loans typically mature in two to three years, just as the facility is being completed - the exact moment when the SPV must refinance into the permanent market. When that paper rolls over in 2027 and beyond, it reprices into a higher-rate world. The asset pays back over decades; the liability demands new terms in years.
Off-balance-sheet leverage. Special-purpose vehicles and lease structures keep a large share of the debt off the hyperscalers' own balance sheets. That preserved investment-grade ratings during the build phase, but it also means the true leverage is opaque to investors pricing the equity. When SPV-level cash flows tighten - because a data center is delayed, a tenant renegotiates, or power costs rise - the losses do not stay neatly quarantined. Either the hyperscaler absorbs the obligation to keep the capacity, or the lenders take an asset that may be worth far less than the loan.
Revenue is promised before capacity exists. A significant portion of the financing is underwritten against future tenant contracts and backlog that assume AI demand continues compounding at the current pace. If utilization lags, the debt still comes due. Oracle's situation - where lenders are scrutinizing how reliable and timely customer cash inflows will be - is the canary.
History offers a cautionary analog. The telecom and fiber-optic build-out of the late 1990s was financed on the same premise: demand for bandwidth would grow exponentially forever, so any capacity built would be absorbed. When the financing structure proved more fragile than the demand thesis, the correction was brutal - hundreds of billions of dollars of fiber debt defaulted, and much of the infrastructure changed hands for cents on the dollar. The technology survived and ultimately thrived; most of the original financiers and builders did not.
None of this means the AI build-out is wrong. It means the financing was priced for a world of abundant cheap capital and uninterrupted demand growth. Both assumptions are being tested at once.
The Counter-Thesis: Hyperscalers Are Not 2000-Era Dotcoms
The strongest argument against the rate-shock narrative is that the borrowers are fundamentally different from the speculative borrowers of past credit cycles. The five hyperscalers carry investment-grade ratings, generate enormous operating cash flow, and are borrowing to fund capacity with contracted backlog rather than speculative dreams. Alphabet's century bond drew heavy demand. Demand for AI compute continues to outstrip supply in most major markets, with CBRE reporting North American data-center vacancy at just 0.9% in its first-quarter 2026 global trends report, and the four largest North American markets absorbing 2.2 gigawatts of capacity over the prior year, a 34% increase. In this view, 2027 refinancing is a manageable rollover for the strongest balance sheets in corporate history, not a crisis.
There is also a circular-financing dynamic that mutes the immediate risk. Hyperscalers borrow to build data centers; chipmakers and equipment vendors take the revenue and park it in short-term instruments; some of that liquidity flows back into the credit instruments funding the build-out. As long as the circle turns, refinancing is available. JPMorgan credit strategists have noted that bond portfolios, which historically moved more closely with interest rates and bank performance, could now become more correlated with the operating fortunes of large technology companies as AI debt issuance expands.
That argument holds for the top tier - but not for the long tail. The debt binge has pulled in smaller infrastructure players, neo-clouds, bitcoin miners pivoting to AI hosting, and SPVs with single-tenant exposure. These borrowers do not have the hyperscalers' ratings, cash flow, or pricing power. A two-tier market is emerging: investment-grade giants refinancing at a higher but manageable cost, and sub-investment-grade developers facing 15% yields, distressed loan prices, or an inability to place debt at all.
The falsifying signal for the bearish view is specific: if the five hyperscalers refinance their 2027 maturities at spreads that do not widen materially beyond current levels, and if data-center utilization rates hold above roughly 90% through 2027, then the rate shock is contained to the fringe and the core AI thesis remains intact. Conversely, if investment-grade spreads widen another 50 basis points or more into 2027 rollovers, or if utilization slips as new supply delivers, the stress migrates up the credit ladder.
What Comes Next: Winners, Losers, and the Watch List
Short term (into 2027): volatility is likely to concentrate in sub-investment-grade AI credit and in the equities of developers carrying the most near-term maturities. Investment-grade hyperscaler bonds may see wider spreads but should remain liquid. The companies best positioned are those that locked in long-duration, fixed-rate funding early - and those with contracted backlog that covers debt service. Convertible bonds, which surged as a funding tool for AI developers, are particularly sensitive to a simultaneous drop in the underlying equity and a widening of credit spreads - the worst combination for that instrument class.
Medium term (2027-2028): expect a shakeout among smaller data-center developers and neo-clouds. Projects without firm power, firm tenants, or firm financing will stall or change hands. Larger players with intact balance sheets may acquire distressed assets at a discount - consolidation is the likely terminal state of a credit-driven build-out. The Oracle loan package, still held by its arrangers at distressed prices, is an early test of how much loss the syndicate is willing to absorb before forcing a restructuring.
Long term (structural): the AI infrastructure build-out itself is a structural shift, not a cycle. Data centers, power, and compute capacity are durable assets with multi-decade utility. The financing structure, however, is cyclical - and it is the structure, not the technology, that is at risk. The build-out will continue; the roster of who owns the assets will change.
Three signals deserve monitoring. First, the spread between hyperscaler investment-grade bonds and comparable Treasuries into each major 2027 refinancing - a widening beyond 50 basis points would signal that even the top tier is paying for the cycle's excesses. Second, data-center vacancy and utilization rates reported by commercial real-estate brokers such as CBRE and JLL; a rise from the current sub-1% North American vacancy would be the first hard evidence that supply is catching demand. Third, Federal Reserve policy: Bank of America's forecast of a 4.25%-4.5% terminal rate through end-2027 sets the refinancing floor, and any shift from that path would reprice the entire debt wall.
"Any kind of large capital expenditure cycle that we have seen over history at some point leads to the risk of overinvestment," said Mohit Mittal, chief investment officer of core strategies at Pimco, which manages about $2.3 trillion in assets.
The AI revolution may be real. But revolutions are financed, and the bill for this one comes due in 2027 - at rates nobody locked in when the borrowing began.
Explore more exclusive insights at nextfin.ai.
