Nvidia Capex Digest — The First Pullback, and the Price of a Signature
TL;DR
The record run ended the only way it could: not with a Nvidia problem, but with a demand-base problem — OpenAI disclosed a ~$50B run-rate versus the ~$68B investors had been told, and the whole AI complex sold off 3–8% in a session. Meanwhile the financing layer quantified its central asymmetry: Broadcom guarantees 85% of its custom-chip debt while Nvidia caps residual support at 25% per project — and the first public-equity exit for AI infrastructure (Firmus) just failed to price.
The pullback finally came, and it was aimed at the demand base, not the seller
The equity story's first real scare arrived via an OpenAI revenue disclosure, not anything Nvidia said or did. On Oct 8, the FT reported — CNBC confirmed — that OpenAI told investors its annualized revenue was roughly $50B at end-September, well below the $68B widely reported in late September; NVDA closed at $230.48, down 2.96%, with Oracle -6%, CoreWeave -8%, AMD -5%, and Broadcom -5% in CNBC's tally (nvidia-record-q1-results-and-trillion-dollar-order-book+1).
"A person familiar with the matter said the $68 billion figure included gross revenue from OpenAI's partners, which helps investors make a more direct comparison with its chief rival, Anthropic." — CNBC, Oct 8
The accounting nuance is real — investors "grossed up" OpenAI's partner-mediated revenue to make it Anthropic-comparable, and OpenAI countered with 77% total run-rate growth and 107% enterprise growth in Q3 (openai-revenue-trajectory). But the market's read was simpler: this was the first hard crack in the demand anchor, and the stock took its first genuine pullback — $5.78T to $5.73T, now 88.2% of its 52-week range, still +14.3% above the 200-day (nvidia-record-q1-results-and-trillion-dollar-order-book
+1). Nvidia's own P&L is untouched — TTM revenue $302.97B (+105.9%), 74.7% gross margin — which is exactly why the repricing matters: it's a bet on the durability of the buyers, not the seller.
What to watch: the audited IPO documents — OpenAI confidentially filed in June for a 2027 debut, Anthropic's S-1 expected ~Nov 15 — which will settle which run-rate was real (openai-revenue-trajectory).
The financing layer quantified the guarantee asymmetry — and it's Broadcom's problem, not Nvidia's
Last cycle's digest framed the fault line as collateral quality; this cycle the market put a spread on it. On the one deal with public terms — June's $35.3B AI XPV stack for Anthropic — Broadcom guaranteed $30B of $35.3B (85%), while Nvidia's stated ceiling for residual-value support is 25% per project (off-balance-sheet-ai-infrastructure-financing-models+1).
"financing has become part of the product... A custom chip that comes with funding is harder to walk away from." — TIKR, Oct 8
The unguaranteed B notes pay 8.5% versus 5.75% on the guaranteed A2 — a 275bp gap that is, in effect, the market's price for Broadcom's signature. Broadcom's three financing packages since June now exceed $145B, roughly 1.26x its ~$115B fiscal-2027 AI revenue outlook, with the correlation risk stated plainly: a frontier lab stops paying when demand for its models disappoints, which is precisely when a chip built for those models is worth least (off-balance-sheet-ai-infrastructure-financing-models+1). This sharpens last cycle's seller/guarantor split — Nvidia's ≤25% cap now looks conservative by comparison, and the OpenAI run-rate disclosure landed two days after the debt trio put that same lab at the center of the debt-funded buildout.
What to watch: the first $500B-platform term sheet with disclosed guarantee terms — that's the benchmark Nvidia's 25% cap will be judged against.
The public-equity exit slammed shut while private credit stayed bid
The financing story's other half is the exit, and it failed this cycle. Firmus Grid, the Nvidia-backed Australian data center operator whose clients include Meta and OpenAI, postponed its planned A$7B IPO — Australia's second-largest ever — after institutional bookbuilding fell short at the A$11/share price targeting ~A$30B; backer Maas Group fell 30% on the news (off-balance-sheet-ai-infrastructure-financing-models+1).
"The AI build-out started on cash... It's increasingly running on credit, and credit changes the risk profile entirely." — Nigel Green, deVere Group, via TIKR
The divergence is the signal: private credit is still bid for AI infrastructure debt, but public equity buyers balked at the asking price — in the same week the 10-year Treasury touched 5.326%, a 24-year high, against a 3.75% fed funds rate (off-balance-sheet-ai-infrastructure-financing-models+1). A buildout that can only refinance through private credit at these rates is a buildout whose marginal cost of capital just moved against it.
What to watch: whether Firmus's contemplated private round prices near its A$30B target — the cleanest test of whether private capital genuinely disagrees with the public market's valuation.
What surprised us
- The pullback was triggered by accounting, not demand. The $18–20B shortfall is substantially a gross-up artifact — investors inflated OpenAI's partner-mediated revenue to force an Anthropic comparison — yet it still knocked 3–8% off the complex. The market was primed to sell on any demand-base disappointment, however labeled (openai-revenue-trajectory
).
- OpenAI is raising again immediately. Early-stage talks for a ~$30B round, after a $122B round in March — with CFO Sarah Friar insisting the company is still "very well capitalized." Post-scare round pricing will be the private-market's vote on the demand base (openai-revenue-trajectory
- Anthropic's hard numbers are brutal and the IPO is still coming. $4.6B of 2025 revenue against a $42B net loss, New Constructs valuing it at $150B versus a $2T ask — "the most ridiculous IPO of 2026" — and the S-1 lands ~Nov 15 regardless (openai-revenue-trajectory
- Hock Tan's financing book is concentrated in exactly the two customers he named. Only two of Broadcom's six major AI customers need financing — Anthropic and OpenAI, the "two geniuses in the middle of Mongolia" — meaning the vendor-financing flywheel and the customer-quality risk sit on the same two names (off-balance-sheet-ai-infrastructure-financing-models
+1).
Open threads worth a vote
- [watch] OpenAI ~$30B Funding Round Confirms, Sizes, or Falls Apart — post-scare round pricing is the private-market vote on the demand base's funding capacity, the input that keeps the Broadcom >$50B Nexus and SpaceX $40B debt-funded chip orders serviceable.