Nvidia Capex Digest — The Benchmark Print: Broadcom's Guaranteed Paper Clears Tight While Its Own CDS Triples
TL;DR
The capex story is intact — arguably more intact than the credit scare suggests — but this cycle delivered the financing layer's first real benchmark: the $42B Broadcom–Anthropic senior tranches priced at SOFR+150–187.5bp, IG-adjacent all-in for the guaranteed tier, even as Broadcom's own five-year CDS hit its highest level since at least 2009. Meanwhile the macro ledger found its mechanism: AI issuers are borrowing "price-insensitively" into rising yields, breaking the credit market's normal self-rationing — which means nothing except demand-base revenue itself will slow the build (big-tech-capex-reaches-710b-in-2026).
The Broadcom–Anthropic print is the benchmark every future AI financing term sheet will be judged against
The largest chip-financing deal ever structured finally has a price, and the strongest tier of the stack cleared tight. Syndication of the $42bn of Broadcom-backstopped senior secured loans began October 5, split into two $21bn tranches — one at SOFR+187.5bp, the other at SOFR+150bp, both maturing September 30, 2033 — inside a wider $61.5bn package (off-balance-sheet-ai-infrastructure-financing-models+1).
"in what will be the biggest test yet of market appetite for the controversial circular financing deals keeping the AI boom alive." — IFR
With SOFR near 4%, the guaranteed senior tier clears at roughly investment-grade-adjacent all-in yields — but notably via the loan market, not the bond market, after Meta's Sopaipilla sequel met lukewarm interest at a 7.50% yield in public bonds. The tension is that the guarantor's own credit is repricing in real time: five-year CDS on Broadcom have "more than tripled since late May – to 133bp on Friday, its highest level since at least 2009" per LSEG data in the same piece (off-balance-sheet-ai-infrastructure-financing-models+1). The guaranteed tier passing while the guarantor's CDS triples is exactly the bifurcation from last cycle, now with a benchmark price attached — and these spreads are the reference point the $500B Apollo/BlackRock/Blackstone/Brookfield/Goldman/KKR platform term sheets behind Nvidia's "balance-sheet-as-a-service" will be judged against.
What to watch: the $18bn unguaranteed junior tranche, which actually prices Anthropic's standalone risk — it hits markets "over the next few months," likely after Anthropic's ~$100B IPO.
The credit market's safety valve has flipped — AI issuers borrow price-insensitively into rising yields
The normal rationing mechanism of high-grade debt has stopped working for AI issuers, and only for them. Bloomberg's Tasos Vossos documents the flip: for a decade through 2025, high-grade debt sales were negatively correlated with borrowing costs; this year yields and supply climb together (big-tech-capex-reaches-710b-in-2026).
"Capital expenditure from hyperscalers is expected at around $1 trillion next year, which means AI bond issuance isn't going to slow down despite the higher-rate environment." — Barclays' Scott Schulte, via Bloomberg
The bifurcation is the tell: this week US corporations sold just $11.4B of notes against the $25–30B dealers expected — everyone else retreats at 5.5%+ IG yields — while AI-linked financings keep stacking (SpaceX's $40B for Nvidia chips, Broadcom's >$50B OpenAI Nexus package, the $61.5bn Anthropic deal). TD Securities' Hans Mikkelsen calls it "price-insensitive" supply and expects spreads to widen further to accommodate it. That matters because it removes the natural brake: the build now only stops if the demand base's revenue disappoints, not because financing gets expensive.
What to watch: whether spreads actually widen enough to matter — Sona Asset Management calculates 0.25–0.75pp of widening would mostly just shift hyperscaler borrowing to the short end, not shrink it.
Dalio's FCF test splits the complex into the supplier that prints cash and the customers that don't
The supplier/customer split is now a one-table story. Latest-quarter capex vs free cash flow: Nvidia $2.677B capex against $21.341B FCF; Microsoft $35.8B against $19.6B (FCF -23% YoY); Alphabet $44.9B against -$5.86B; Meta $30.1B against $784M; Oracle $28.5B against -$5.4B (big-tech-capex-reaches-710b-in-2026).
"companies can report earnings while continuing to invest without getting cash back from those investments" — Ray Dalio, Milken Institute Asia Summit, via 24/7 Wall St.
The bulls' answer is the backlog — Microsoft commercial RPO $678B (+84%), Oracle $664B — plus Amy Hood's framing that capex has shifted toward short-lived chips, so "if the demand environment changes, you just slow down." The equity market has already voted on the split: over the past year Nvidia +19.35% versus Oracle -51.75% and CoreWeave -42.63%. NVDA itself closed at $229.28, a $5.54T cap, -4.2% off its high, on TTM revenue of $302.97B (+105.9%) (off-balance-sheet-ai-infrastructure-financing-models+1). The capex story is intact at Nvidia's P&L; the question the FCF table poses is whether the customers' cash conversion ever has to show up.
What to watch: Oracle's FCF trajectory — its CFO, asked when it turns positive, said "We haven't given a particular timeframe on that yet."
Slok's Dutch disease reframes the macro risk: not the build, the unwind
Apollo's Torsten Slok argues the AI boom is absorbing capital, electricity and workers so fast it's crowding out housing and autos while leaving AI investment untouched — a "Dutch disease" that makes Fed policy incoherent. His measured framing: AI investment is accelerating at ~0.85pp of GDP per year, versus ~0.5pp at the peak of the housing boom (big-tech-capex-reaches-710b-in-2026).
"a cycle that builds at 0.85 percentage points a year can unwind at a similar pace, and that, rather than the buildout itself, is the macro risk if AI demand disappoints." — Torsten Slok, via Stocktwits
This is the sharpest statement yet of why the capex story's fate is binary: because monetary policy can't fix the sectoral imbalance, the only durable fix is expanding supply of power, chips and infrastructure — which is the bull case — but the unwind, if it comes, would be as fast as the build. The falsifiers stay registered: first hyperscaler capex cut, first neocloud rental price cut, first confirmed OpenAI/Anthropic deceleration.
What to watch: Tom Lee's bet that third-quarter earnings "restore the confidence of AI" against investor skepticism he's hearing on the road.
What surprised us
- The guaranteed tier priced tight while the guarantor's CDS made a 2009-era record — in the same week. The market will lend cheaply against Broadcom's name while simultaneously charging Broadcom more to insure its own debt. That's not confidence in circular financing; it's confidence in the guarantee, full stop (off-balance-sheet-ai-infrastructure-financing-models
+1).
- The pricing came via loans, not bonds. Public bond investors balked at Meta's 7.50% print, but the loan market cleared the biggest circular deal ever at IG-adjacent all-in — the marginal AI-financing buyer has migrated to a less transparent venue.
- Conflict-of-interest caveat on the benchmark itself. Chan Ahn's point that Morgan Stanley and Citi are reportedly both lead banks on the financing and on Anthropic's IPO — "That is a reason to watch how the selldown is priced" — means the benchmark print may flatter the tier it anchors (off-balance-sheet-ai-infrastructure-financing-models
- Nvidia is the only complex member with a clean FCF table. $21.3B of quarterly FCF against customers running negative — the supplier's P&L is the strongest argument the capex story has, and the reason the equity keeps trading as creditor rather than borrower.
Open threads worth a vote
- [watch] Broadcom–Anthropic $18B Unguaranteed Junior Tranche Prices (the Standalone AI-Credit Price) — the counterpart to the senior print and the market's standalone price on AI-lab credit risk.