Hyperscaler AI Capex Has Crossed Into Debt Financing. Here Is What That Structural Shift Means.

On 10 February 2026, Alphabet entered the sterling bond market with a £1 billion 100-year note maturing in February 2126. It was the first century bond issued by a technology company in nearly 30 years. The offering attracted £9.5 billion in bids — nearly ten times the £1 billion sought. Across the full multi-currency offering, orders exceeded $100 billion against a raise of approximately $32 billion, more than five times oversubscribed.

The detail worth examining is not the investor appetite — it is the mismatch at the core of the transaction. Alphabet was issuing 100-year paper, priced at spreads appropriate for a company with tens of billions in annual free cash flow and a balance sheet most sovereigns would envy, to fund infrastructure centred on graphics processing units (GPUs) with useful lives of four to six years. Credit markets, on this evidence, were willing to price that mismatch as negligible risk. Whether they are right is the question that defines the hyperscaler capex cycle of 2026.

The Hyperscaler Capex Threshold

The five largest US hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — are projected to spend between $700 billion and $750 billion in capital expenditure in 2026, up approximately 67% year-over-year. That follows a 73% increase in 2025. Aggregate capex has roughly tripled in two years: the group spent approximately $256 billion in 2024 and approximately $443 billion in 2025.

The headline number has been widely reported. What has received less systematic attention is the funding mechanism shift. For the better part of a decade, these companies financed infrastructure internally: capex consumed roughly 40% of operating cash flow on a historical ten-year average, leaving substantial free cash flow for buybacks, dividends, and reserves. That ratio has broken decisively. Aggregate capex now consumes approximately 90% of the group’s operating cash flows, according to analysis from MUFG Americas’ AI Supercycle financing report. After accounting for buybacks and dividends, aggregate capex exceeds projected operating cash flows — a structural breach of the internally-funded model.

Amazon is the sharpest illustration. Its operating cash flow for the trailing twelve months ended Q4 2025 was $139.5 billion — up 20% year-over-year. Capital spending over the same period was $128.3 billion, up 65%. With 2026 capex guidance of approximately $200 billion, the gap cannot be closed by operating cash flow alone. Morgan Stanley projects Amazon will generate negative free cash flow of approximately $17 billion in 2026; Bank of America sets the deficit at $28 billion.

What the Money Is Buying

Approximately 75% of aggregate 2026 hyperscaler capex — roughly $525–$563 billion at the projected $700–$750 billion range — is directed at AI infrastructure: GPU clusters, the networking to interconnect them at scale, and the data center construction to house both. The remaining quarter covers traditional cloud infrastructure maintenance and end-of-life equipment replacement.

Microsoft’s Q1 FY2026 capex of $34.9 billion — a single quarter’s spend that exceeded full-year capex from several years prior — illustrates the composition: approximately half went to short-lived assets, primarily GPUs and CPUs supporting Azure AI demand; the other half funded long-lived assets including $11.1 billion in data center finance leases intended to support monetisation over 15-plus years. Microsoft expects AI capacity to increase by more than 80% through fiscal year 2026 and to roughly double its data center footprint over two years.

Alphabet guided to $175–185 billion for 2026, nearly doubling the $91.4 billion spent in 2025. In Q4 2025, approximately 60% of capex went to servers and 40% to data centers and networking — a mix reflecting the GPU-dense inference workloads driving demand growth. Google Cloud revenue reached $17.7 billion in Q4 2025, up 48% year-over-year, but even at that growth rate, revenue conversion of the infrastructure investment is measured in years, not quarters.

The Credit Market Response

The debt financing shift has reoriented a substantial portion of corporate bond market activity toward technology. Big Tech (Alphabet, Amazon, Meta, Oracle) raised approximately $93 billion through investment-grade bonds in 2025 — roughly 6% of total US investment-grade (IG) issuance. Morgan Stanley expects $250–300 billion in hyperscaler bond issuance in 2026 from hyperscalers and related joint ventures. JPMorgan projects approximately $300 billion annually across all AI/data-center deals for the next five years, with the total multi-year figure for the technology sector reaching $1.5 trillion.

Credit market appetite has so far been strong. Alphabet’s February offering attracted more than $100 billion in orders. Bank of America’s Matt McQueen, a senior debt capital markets banker, described the broader numbers as “like nothing any of us who have been in this business for 25 years have seen.” JPMorgan’s Tarek Hamid noted a structural consequence: bond portfolios that historically “traded much more correlated with rates and banks’ performance are now going to be correlated with technology companies’ performance.”

The Off-Balance-Sheet Problem

The visible debt issuance may understate actual financial commitments. Moody’s published analysis in February 2026 flagging $662 billion in uncommenced data center lease commitments held by five hyperscalers (Amazon, Meta, Alphabet, Microsoft, Oracle) that sit entirely off balance sheet under GAAP accounting. The $662 billion represents leases contractually signed but for which services have not yet been received. Total undiscounted future lease commitments for the group reached $969 billion at end-2025.

The $662 billion equals 113% of these five companies’ adjusted on-balance-sheet debt. Moody’s has indicated it may apply non-standard adjustments to reflect these exposures in credit analysis. The agency has characterised these off-balance-sheet structures as limiting financial flexibility if conditions change rapidly.

The Duration Mismatch

The Bank for International Settlements (BIS) has published the most direct statement of the structural concern: the AI infrastructure buildout is creating an asset-liability duration mismatch. Long-term bonds — including, at the extreme end, Alphabet’s century note — are being issued to fund assets with four-to-six-year useful lives. If AI adoption or revenue growth is slower than projected, the leverage embedded in these structures amplifies any financial shock.

The risk is more immediate in the secondary financing ecosystem. CoreWeave’s $7.5 billion GPU-collateralised debt facility carried variable rates averaging approximately 11% and began repayments in January 2026 — as the market value of the underlying GPU collateral was declining. The gap between asset duration and financing duration is manageable for AAA-rated hyperscalers. It is a live risk for the neoclouds, data center real estate investment trusts (REITs), and GPU lessors that have leveraged their balance sheets against hyperscaler demand without equivalent credit depth.

Oracle sits at the exposed end of the hyperscaler group. Barclays downgraded Oracle’s debt to underweight in early 2026, warning it could fall to BBB-minus — the lowest investment-grade rating before junk status, a signal that at least one major bank considers Oracle’s debt under meaningful pressure.

Enterprise Implications

For cloud buyers — chief financial officers (CFOs) and chief information officers (CIOs) managing multi-year infrastructure commitments — the debt financing shift has two near-term implications.

The first is pricing. Hyperscalers have not formally announced price increases tied to capex intensity, but the structural direction is clear. When capex is funded at 4–4.5% investment-grade coupon rates and operating cash flows are consumed at 90%, the timeline for monetisation shortens. Enterprise contracts with annual renewal clauses are the first exposure point; reserved-instance agreements locked before 2026 provide one to three years of price protection.

The second implication is negotiating leverage. Hyperscalers need committed cloud revenue to justify the capex cycle to credit markets. Long-term, high-volume committed-use contracts become more valuable to the seller than they were when infrastructure was funded internally. Enterprise buyers who can commit at scale have more negotiating room on total contract value, tiered pricing, and egress fee structures than the current pricing environment might suggest.

What to Watch

Three signals are worth monitoring in the second half of 2026.

Revenue-to-capex conversion ratio. AI-related cloud revenue from the hyperscalers was approximately $25 billion in 2025 against roughly $450 billion in AI-specific capex that year. The ratio needs to improve materially for the debt service math to work comfortably. Q2 and Q3 2026 earnings calls will be the first read on whether inference workload revenue is scaling in proportion to infrastructure investment.

Secondary market credit spreads. Watch for divergence: tightening spreads on primary hyperscaler paper concurrent with widening in the secondary ecosystem (data center REITs, GPU lessors, neoclouds) would signal that credit markets are beginning to price the layered leverage rather than just the hyperscaler name.

Moody’s adjusted debt treatment. Moody’s has indicated it may apply non-standard adjustments to reflect the $662 billion off-balance-sheet lease pool. A formal methodology update from a major ratings agency would represent a meaningful repricing signal for the entire sector’s debt outstanding — and would change the leverage ratios in every credit model currently built on GAAP figures.

The infrastructure buildout is real, the demand signals from enterprise AI adoption are genuine, and the companies executing it have balance sheets that can absorb a multi-year payback period. The shift that has occurred is structural: AI infrastructure is now a credit market instrument, not just an operational one. That changes the risk equation for every participant in the ecosystem — the companies building, the investors financing, and the enterprises buying.

Related posts:

$300 Billion in AI Infrastructure Spending: Where the Hyperscaler Capex Is Actually Going AI Consumption Pricing: What the Per-Token Shift Means for Enterprise Budgets Three Fusion Milestones in Six Months Signal a Shift in Commercial Timelines Grid-Scale Battery Storage Crossed 1 TWh Cumulative Capacity. The Business Model Is Still Unsettled.