The AI Supply Machine

THE BOND BRO | DEEP DIVE The AI Supply Machine How the hyperscaler bond binge reprices the Treasury long end — and the hidden leverage the tape doesn't show.

The one-line thesis The largest, best-rated companies on earth have started borrowing at historic scale to fund AI infrastructure. The credit market absorbs it at firm spreads — so the pressure doesn't show in OAS. It shows in the Treasury long end, as duration the market must digest and term premium it won't give back. And the public bond tape is only half the leverage.

1. The Ledger — From $28B to a $300B Run-Rate For years the hyperscalers funded growth from cash flow. That ended in H2 2025. The five largest — Amazon, Alphabet, Meta, Microsoft, Oracle — averaged ~$28B/yr of US bond issuance from 2020–2024. In 2025 they issued ~$121B (BofA), ~$90B of it in Q4 alone. Through five months of 2026 they'd already sold ~$159B, +47% y/y.

The landmark deals: Oracle $18B (Sept '25); Meta $30B (Oct '25), the largest non-M&A high-grade deal on record; Alphabet $17.5B + Amazon $15B (Nov '25). 2026 didn't slow: Oracle $25B/8 tranches (Feb); Alphabet ~$32B multi-currency including a rare 100-year sterling bond; Meta another $25B (Apr); Amazon ~$54B YTD, including this week's $25B deal that drew ~$41B of orders. AI-tied debt hit ~$1.2T by late 2025 — ~14% of the high-grade market, past banks as the largest sector.

2. The Mechanism — Capex Eats the Cash Flow The reason they're borrowing is arithmetic. The Big Four guided to ~$700B of 2026 capex, roughly double 2025 (UBS: ~$770B all-in). These are cash machines — OCF for the five hit ~$577B in 2025 (from ~$378B in 2023). But capex now consumes close to 100% of operating cash flow in 2026, vs a ~40% ten-year average (UBS). The gap between capex and cash flow is the bond issuance. That's the whole engine. Balance sheets stay strong (debt-to-cash actually falling 0.94→0.75) — this is optimization at four of the five. The exception is Oracle: Baa2/BBB, negative free cash flow until ~2029. It borrows because it has to.

3. Why Spreads Stay Tight but Rates Don't Record supply hasn't blown out spreads. IG demand is deep — ~$0.5T into taxable bond funds in 2025, foreign buyers ~$304B of US corporates (Breckinridge). That pool clears jumbo deals at firm OAS, often sub-5% coupons. When demand absorbs supply at the index level, the pressure exits through the risk-free curve: every jumbo deal is duration the market must hold, repricing as higher real yields and term premium, not wider credit. That's why the long end stays cheap on soft data, on risk-off, on a crude collapse — and won't rally. The single-name tell is Oracle: index spreads calm, but its 5-year CDS has more than tripled since the September deal (MUFG). The must-borrow name carries the stress the index doesn't.

4. The Hidden Pipe — What the Bond Tape Doesn't Show Public issuance understates the true AI debt load, because much of the financing never touches the bond tape — it sits in off-balance-sheet structures: infrastructure JVs and private credit. Microsoft has largely stayed out of public debt markets, funding through fund-level vehicles. The BlackRock/GIP/MGX AI Infrastructure Partnership executed the ~$40B Aligned Data Centers acquisition (late 2025) — largest data-center deal on record — with ~70% leverage at the fund level, not on any hyperscaler's balance sheet. The visible IG supply reprices Treasuries; the private-credit/JV channel is a second, less-visible pipe of AI leverage. When you hear "fortress balance sheets," remember a chunk of the risk has been routed off those balance sheets entirely.

5. The Tail, and the Trade The tail: if AI capex disappoints, three things bite at once — jumbo deals don't refi cleanly, the ~70%-levered off-balance-sheet channel strains, and because this debt sits with institutions not retail, the impact concentrates in leveraged holders and data-center financing. That's the scenario a reported Treasury working paper is said to have flagged: fewer retail holders than the dotcom era means a drawdown lands where financial plumbing matters.

The trade: own the steepener. The front end reflects policy; the long end carries the supply. Term premium stays bid and long-end cheapness is structural — the Fed sets the front, but can't cut the issuance away. Tactically, the jumbo new-issue calendar is where concession shows up — favor tenors and names where the supply premium overpays for the credit. Lower-rated issuers (Oracle) and ultra-long tenors (40–100y) move first.

Sidebar: the club is widening — NVIDIA and SpaceX have run ~$25B deals; Alphabet's 100-year sterling bond (~+120 over gilts) shows how far out issuers reach when demand is there. More issuers, longer tenors, same duration-absorption story.

The bottom line The AI buildout has a financing footprint, and it runs straight through the Treasury long end. Record IG supply that clears at firm spreads doesn't relieve the curve — it reprices it. For a rates book: a persistently cheap long end and a steeper curve for as long as the machine runs. For a credit book: a barbell — fortress names absorbing supply at tight spreads, and a short list of must-borrow and off-balance-sheet exposures where the real risk has been quietly parked.

Full PDF -> https://tinyurl.com/Dispatch070826

Disclosure: Educational/informational only; not investment, tax, or legal advice, nor an offer or recommendation to buy or sell any security. Levels as observed and subject to change. The Pressure Gauge is a proprietary qualitative framework, not a model output. Any specific issuer or CUSIP is used to illustrate structure, not as a recommendation. TEY assumes a 40% marginal rate; tax treatment is entity-specific — consult your own advisor. Past performance and historical spread relationships do not guarantee future results.