Nvidia Capex Digest — The AI-Credit Curve Completes Itself While the Capital-Cycle Question Gets a Number
TL;DR
The two threads that were open last cycle both resolved, and both resolved in the same direction: the unguaranteed junior tranche of the Broadcom–Anthropic package found its anchor at a 10%+ coupon — roughly 300bps over what similarly rated borrowers pay in public markets — and BofA put a historical frame on the equity selloff, comparing $1T of consensus hyperscaler capex to 3% of US GDP, above the 1870s railroad boom. The demand story is untouched; what moved this cycle is the price of the debt that carries it, and the first serious sell-side argument that the technology can win while the investors lose.
The unguaranteed tier of AI credit now has a disclosed price — and the borrower itself priced the conflict
The financing ladder that has been building for weeks finally completed its top rung. The FT reported, with Reuters wire confirmation, that Anthropic offered Blackstone a coupon above 10% to anchor the junior tier of the $60B Broadcom-arranged "Project North Star" package — the largest chip-financing arrangement to date — with Blackstone-managed funds committing $9B of their own capital, half of the $18B Class B tranche, and syndicating the rest (off-balance-sheet-ai-infrastructure-financing-models+1).
"Blackstone has been offered a 10% coupon, with total returns exceeding that once issuance discounts and fees are counted, versus roughly 7% that similarly rated borrowers pay in public bond markets." — AI Weekly, summarizing the FT
Two details make this more than a pricing print. First, the instruments can convert into Anthropic shares — lenders demanding equity upside on top of a double-digit coupon is textbook late-cycle credit behavior. Second, Anthropic's own IPO prospectus flags the vendor-financing conflict, warning that triggering certain payment or performance defaults could accelerate a material portion of its lease obligations while cutting off access to the $42B credit line — the cross-default architecture of the circular model, stated in the borrower's own filing (off-balance-sheet-ai-infrastructure-financing-models+1). Last cycle's digest ended with the question of whether this unguaranteed tier would price at all; it did, at ~300bps over the public equivalent, which extends rather than answers the worry.
The systemic context hardened the same week: Goldman Sachs put AI-exposed US high-yield risk premiums at 360bps over Treasuries after a 25bps two-day widening, Gartner warned that more than 80% of neocloud providers could cease independent operations by 2030, and Nebius — with $575M of AI cloud revenue growing 514% YoY — still fell 12% on CoreWeave debt contagion while carrying $5.7B of quarterly capex against $582.3M of group revenue (off-balance-sheet-ai-infrastructure-financing-models+1). Nebius is the cautionary exhibit: demand is real, but the funding gap between EBITDA and buildout is the entire story.
What to watch: the final close of Blackstone's $9B anchor and syndication of the remaining junior paper — and whether the equity-conversion feature becomes standard across AI-lab deals.
BofA turned the selloff from a valuation squabble into a capital-cycle argument — and named the falsifier
The equity drawdown extended, but the framing changed in a way that matters more than the tape. The Philadelphia Semiconductor Index fell an eighth consecutive session on Oct 16 — its longest losing streak since 2008 — with SOXX closing at $493.39, roughly 25% below its 52-week high, driven by macro rather than credit: WTI at $101.91 and the 10-year at 5.10% (nvidia-record-q1-results-and-trillion-dollar-order-book+1). The new element is BofA strategist Sebastian Raedler's note, which anchors the bear case in history rather than hype: consensus hyperscaler capex expectations have tripled from $330B in early 2025 to $1T — about 3% of US GDP, versus 2.2% at the peak of the 1870s railroad boom (nvidia-record-q1-results-and-trillion-dollar-order-book
+1).
"Most major capital cycles of the past have ended in disappointment, even though the underlying technology often lived up to expectations." — Sebastian Raedler, BofA, via 24/7 Wall St.
The monetization evidence Raedler cites is the sharpest part: average token prices have fallen roughly 50% since the summer, and Ramp's business-AI spending data show outlays fading since August even as token volumes keep rising — price falling faster than volume grows. BofA's structural point cuts at Nvidia specifically: at these multiples its revenue depends directly on hyperscaler budgets holding, making it far more exposed than Microsoft to any capex retreat (nvidia-record-q1-results-and-trillion-dollar-order-book+1). Yet NVDA was the survivor of the complex, sliding only 1.43% to $214.13 while AMD — despite a strong quarter and Anthropic's commitment to deploy up to 2GW of MI450 GPUs from H1 2027 — fell 18.58% over the week. Alternative silicon is levered to the same capex cycle, and the market is pricing that.
What to watch: Microsoft's late-October print — Azure growth against its ~45% constant-currency guide is BofA's own trigger for whether the token squeeze has reached a major seller of AI capacity.
What surprised us
- The two open threads resolved in one day, both upward on the credit curve. The $42B senior tranche had cleared at SOFR+150/187.5 last cycle; the $18B junior now has Blackstone at 10%+ with conversion rights. Every rung of the AI-credit curve stepped higher in the same week the equity layer got its capital-cycle warning — the two repricings are converging on the same conclusion from opposite sides (off-balance-sheet-ai-infrastructure-financing-models
+1).
- AMD's good news didn't save it. A ~$13B Q3 revenue guide and Anthropic's 2GW MI450 commitment still left the stock down 18.58% on the week at 199x trailing earnings — the market is telling you that second-source silicon is not a hedge against the capex cycle, it's a lever on it (nvidia-record-q1-results-and-trillion-dollar-order-book
+1).
- The macro driver is being underweighted in the AI narrative. The eight-day SOX streak was proximately driven by $101.91 oil and a 5.10% 10-year — inflation that raises the odds of another Fed hike — not by any AI-specific default. If rates fall, the credit story may look less apocalyptic without anything about the order book changing.
Open threads worth a vote
- [watch] Microsoft FQ1 Earnings — Azure Growth vs ~45% cc Guide and Capex Path — this is now the single most important falsifier short of an outright capex cut: a miss or capex trim would be the first hard confirmation that token-price deflation is reaching a major seller of AI capacity, with Amazon reporting Oct 29 and Alphabet and Meta the same week (nvidia-record-q1-results-and-trillion-dollar-order-book