AI capex debt surge breaks hyperscalers' investor pact
Get the Finance newsletter
Daily finance — markets, central banks, M&A, the prints that move money. Free.
- Amazon, Meta, and Alphabet all unveiled sizable increases in full-year capex plans during earnings season, with UBS data projecting aggregated AI hyperscaler capex could top $770 billion in 2026 — 23% higher than previously expected.
- UBS credit strategists said in a Feb. 18 note that the capex increases imply a $40 billion to $50 billion ramp-up in hyperscaler borrowing, pushing public-market debt issuance to $230–$240 billion this year.
- Oracle tapped the bond market for roughly $18 billion last September, and Alphabet recently issued around $20 billion in debt — including a rare 100-year sterling-denominated bond — as tech debt issuance accelerates.
- Mirabaud's Al Cattermole said the shift breaks the 'unspoken contract' that kept AI capex equity- or cash-funded, forcing creditworthiness questions on firms long treated as 'essentially cash-plus,' and flagged Alphabet's planned capex of almost 50% of revenue as 'unheard-of.'
- BlackRock warned mega-cap tech is using the credit 'bonanza' to bridge the gap between current investment and future revenues while bond markets already struggle to digest large public deficits, and said it prefers high-yield and European bonds.
- Vanguard's Shaan Raithatha acknowledged hyperscalers start from strong positions but flagged hidden risks building via special purpose vehicles, asset leasing, and off-balance-sheet activity.
- Cattermole warned debtholders face obsolescence risk: if Nvidia chips are outstripped by a competitor mid-loan, the data centers backing 5–8 year debt could lose their value.
Why it matters: Bondholders who treated these AA-/A-rated tech giants as 'cash-plus' must now evaluate creditworthiness on $230–$240 billion in expected 2026 debt issuance, with UBS projecting $40–$50 billion of that coming directly from hyperscalers. That gives fixed-income investors a new lever to price the AI trade — sorting winners from losers — but also exposes them to data-center obsolescence if chip architectures shift mid-loan.
