AI & Credit: The Absorption Test

The AI buildout, which involves the largest debt-funded capex cycle since the construction of the railroads in the 1800s, is testing the limits of the US credit market – the deepest in the world – and the market seems to be doing its job. Because so much AI debt flows through the liquid bond channel, underwriting discipline arrived early in the form of wider spreads.

KEY POINTS

• Bottom line: We are living through a generational buildout that already rivals some of the biggest investment cycles in modern history – and its pace keeps accelerating. The capital needs are extremely high and the capital markets have been adjusting rapidly. For large and well-capitalised companies, such episodes are heartburn, not a heart attack. With power, chips and now the cost of debt all rationing the buildout, owners of powered and funded compute hold the advantage.
• Cycle: At 8.0% of US GDP, cumulative AI investment already exceeds that seen for shale, cloud, and the nuclear grid. With $7tn of investment projected over 2026-30, AI becomes the third largest investment cycle in 150 years, nearly matching the internet, and behind only the US railroads. AI debt reaches 5.3% of US GDP by 2030; counting the off-balance sheet layer, it goes to 7.7%, making it the most significant credit cycle since US rail.
• Funding: 60% of the $7tn comes from internal cash flow ($4.2tn), 26% from debt ($1.8tn), and 14% from equity ($1tn). Debt carries 40% of funding in 2026 and this year’s free cash flow after capex runs near zero. We expect the picture to improve with internal cash recovering to 68% of the mix by 2030.
• Debt: AI bonds now account for 8.1% of the US Investment Grade market and 4.2% of US High Yield, having held near 5.5% of IG for two years before breaking out 2.6pts in nine months. Our base case sees AI bonds taking 15%+ of the IG market and 10%+ of HY by 2030. Issuance is currently running at more than 3x last year’s pace, with $417bn printed year-to-date.
• Equity: Equity has surprised to the upside: $250bn raised year-to-date, roughly half of external funding. Still, the IPO window has narrowed of late and may be temporarily closed: two recent big-ticket IPOs, SPCX and the SK Hynix US ADR, have traded poorly post-debut.
• Off-balance sheet: $600bn+ already stands off-balance sheet (OBS): $475bn of capitalised hyperscaler leases plus $125bn+ of SPV and JV non-recourse debt. This channel lives outside our $7tn AI investment base case because the capex occurs at developer level. Most SPV debt is designed to amortise fully.
• Chip financing: A parallel market in chip-collateralised debt is forming. Announced programmes total $1.3tn+. Only $58bn has funded across six closed deals, all private or 144A, with zero index bonds. Another $32bn is in market at the time of writing, with talk on the largest deal moving from UST+425 to a 9-10% yield. If this debt keeps printing off-index, roughly $500bn of our projected debt leg lands outside the index.
• Scenarios: The first 25-75bp of hyperscaler widening changes how the buildout gets funded, but the buildout proceeds. At +100bp the paths diverge: hyperscaler supply halves and capex drops 20%, while neocloud supply falls two thirds and capex is cut 40%. At +150bp the cycle breaks: debt shrinks from $1.8tn to $475bn and the neoclouds get frozen out.
• Feedback: The bond market gives the AI cycle a feedback mechanism that was not present in previous super-cycles. Issuers get a live mark on their cost of capital and adjust fast: Amazon and Meta deferred their plans, tenors shortened, the riskiest deals were pulled. That responsiveness argues for a longer cycle.
• Demand: Capex-generated revenue is arriving. Cloud revenues grew by an average of 54% yoy on a $260bn base across Amazon, Microsoft and Google. Backlogs stand at a combined $1.5tn. Amazon says “the lion’s share” of its 2027 capacity is already reserved, with a meaningful share booked into 2028. The item to watch is unit pricing once capacity comes online; the same turn that hit telecoms, shale and rail.