Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The capex story is intact at every observable gauge — NVDA pushed to $5.77T, Anthropic's S-1 showed the demand base doubling sequentially — but the tension underneath hasn't resolved; it's sharpened into a question about collateral quality. The credit market is now differentiating not just by guarantee tier but by whose silicon sits in the data center, and that's the new fault line for Nvidia's financing flywheel.
The equity layer is pricing a milestone, not a debate
The stock is behaving like the capex story is already won. NVDA closed Tuesday Oct 6 at $239.24, a fresh all-time high at $5.77T market cap, dragging the S&P 500 and Nasdaq to records with it — Nvidia is now 13% of the Nasdaq and 8% of the S&P 500 (CNBC). The options market has turned the $6T crossing into a dated event:
"about a 50% chance Nvidia will reach a $6 trillion valuation between now and the end of this month, and a 13% chance it happens this week... The odds it happens between now and Dec. 18 is about 67%" — CNBC
Morgan Stanley reinstated NVDA as its top semiconductor pick on Oct 6, and the fundamentals under the record are real: TTM revenue of $302.97B (+105.9% YoY), 74.7% gross margin, four straight EPS beats (nvidia-record-q1-results-and-trillion-dollar-order-book+1). The caveat from last cycle still stands — RSI at 84 leaves no cushion — but the equity layer keeps voting "intact" with momentum the credit layer keeps defying.
What to watch: the $6T crossing itself — options have it at even odds by end-October, and the stock needs only $248.
The credit market split into two records in the same week — one for price, one for scale
The financing layer bifurcated. At the weak tier, JPMorgan is pitching ~11% all-in on Volta's $5B leveraged loan — 6.25–6.5pp over benchmark — "one of the highest borrowing costs seen in the market for risky debt that's become popular for financing the AI boom" (Bloomberg via Yahoo). The deal is unrated, which "typically prevents collateralized loan obligations — the biggest buyers of this type of debt — from purchasing the loans." At the strong tier, BofA, Citi and Morgan Stanley sent syndication letters for the $42B Broadcom-guaranteed senior tranche of the Anthropic package, with Blackstone leading the $18B junior tranche (TechTimes).
Rothschild's Robert Leitao put the concentration risk plainly:
"It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened." — TechTimes
Rationing has started at the margins: PGIM anchored the first CLO with an explicit AI-collateral cap (15% of pool), and BofA estimates CLO chip/data-center exposure could reach $100B (off-balance-sheet-ai-infrastructure-financing-models+1). The flywheel is still funded — but every weak-tier print sets a new record yield.
What to watch: Volta commitments due Oct 14 — closing at marketed terms confirms 11% as the clearing level for unrated GPU-collateralized neocloud debt.
The new fault line: what the collateral is made of
Last cycle the question was guarantee tiers; this cycle it's chip type. TechTimes flags that TPU collateral is structurally less fungible than GPU collateral — a TPU "cannot be repurposed for general-purpose compute the way a GPU can," leaving "essentially one buyer class: frontier AI developers" — while existing credit frameworks were "developed largely around Nvidia GPU collateral" (off-balance-sheet-ai-infrastructure-financing-models+1). The Volta 11% print is what unrated GPU paper clears at; what ASIC paper would clear at unguaranteed is masked by Broadcom's A-minus guarantee. If GPU-backed paper systematically clears tighter than ASIC-backed paper at equal credit tiers, Nvidia's ecosystem holds a structural cost-of-capital advantage its competitors have to finance around.
What to watch: the first unguaranteed ASIC-collateral term sheet — it would unmask the spread the Broadcom guarantee is hiding.
The demand base answered the deceleration bears with an S-1
The bear chain's key claim — that Anthropic's ARR had plateaued since June — ran into the S-1's own preliminary numbers: Q2 2026 revenue of $11.6B, up from $4.73B in Q1, overtaking OpenAI's $6.7B quarter for the first time (anthropic-ipo-compute-obligations). The commitment arithmetic moved too: the $518B of compute obligations ($161.2B Broadcom / $111.1B Google / $110B Amazon, roughly 80% non-cancelable) now works out to ~11:1 against the ~$46B annualized run-rate, versus ~$113 committed per $1 earned against 2025 revenue.
"Anthropic just overtook OpenAI in quarterly revenue for the first time... The company that markets itself on caution just out-earned the company that markets itself on speed." — Yahoo Finance
The counter-read persists — disruptionbanking calls it "the Most Alarming Risk Disclosure in IPO History. Nothing in It Is Binding" — and the figures are unaudited and preliminary. But the direction matters: the demand base is growing into its obligations faster than the 2025 snapshot implied, and the Broadcom leg ($161.2B) is exactly what the $60B syndication is funded against.
What to watch: whether Q3 shows the sequential doubling was structural or "a one-quarter spike driven by enterprise adoption cycles."
The rental-price tripwire still hasn't fired
CoreWeave CEO Mike Intrator, at the company's inaugural Fully Connected conference: "There has absolutely not been any reduction in orders... The market is looking for more infrastructure... and it's just a question of where you put it" (BigGo Finance). On the rate overhang — the most sensitive point for a debt-funded neocloud — "The margins that we are earning are expanding faster than the rates are going up," a trend he projected through 2027 into early 2028. Nebius's 17–21% pay-as-you-go hikes took effect Oct 1, the second hike in three months (gpu-deployment-gap-overbuild-debate).
The telling datum is CRWV itself: near $87, up ~22% YTD but still ~53% below its June 2025 record, even as sell-side bulks bullish (Truist at $165, roughly double spot). Public equity investors are pricing financing-cost and dilution risk, not demand weakness — the bear chain still lacks its confirming disclosure.
What to watch: the first neocloud price cut or write-down — until one appears, the shortage reading of high rental yields keeps winning.
What surprised us
- The demand base's own S-1 became the bulls' best exhibit. The deceleration case leaned on Anthropic's ARR plateau; the S-1's preliminary Q2 showed sequential doubling and a first quarter of positive adjusted operating profit. The bear's primary source flipped.
- Collateral quality, not credit quality, is emerging as the differentiator. The market spent last cycle pricing guarantee tiers; now the question is whether a GPU is better collateral than a TPU — a question that directly advantages Nvidia's ecosystem and that nobody was asking six weeks ago.
- OpenAI lost the revenue lead and the profitability race simultaneously. A $12.3B quarterly operating loss against Anthropic's first positive adjusted quarter — the two flagship demand stories are diverging, not just competing.
- CoreWeave's margin claim inverts the rate-hike bear argument. "Margins expanding faster than the rates are going up" means rising financing costs are being passed through — the neoclouds are currently the price-setters, not the price-takers.
Open threads worth a vote
- [watch] CLO AI-Exposure Caps Spread Beyond PGIM's 15% Cap — a second CLO manager adopting an AI concentration cap would be the credit channel's first visible rationing signal.
- [question] Does the credit market price Nvidia-GPU collateral at a premium to custom-ASIC (TPU) collateral? — the answer determines whether the 11% Volta print is about AI risk generally or about collateral quality specifically.