Nvidia Capex Digest — The Volta Deadline Answers, OpenAI's Number Gets Marked Down, and Nvidia Responds by Opening the Blueprints
TL;DR
The capex story survived its sharpest week of scrutiny yet, but the damage is landing on everyone except Nvidia: OpenAI marked its 2026 revenue down to $50B from the $70B it had signaled, the two-tier AI credit market formalized around Volta's best-efforts GPU loan at 11%+, and Microsoft — not Nvidia — took the analyst downgrades. Nvidia's counter to the substitution and efficiency critique arrived at the OCP Summit: open the rack designs, standardize the optics, co-write the facility specs with the companies building alternatives.
The demand base got marked to a comparable number, and the market repriced the buyers
The revenue gap is real but narrower than the headline: OpenAI told investors it will reach $50B this year versus the $70B previously signaled, and part of the discrepancy is definitional — Anthropic "includes the revenue from sales via cloud partners, such as Amazon's AWS and Google Cloud, while OpenAI does not" (openai-revenue-trajectory).
"OpenAI has revealed that it is making about $20bn less in projected revenue than it had recently indicated to investors." — The Guardian
The market's verdict was surgical: Nasdaq -1.4% on the day, with Nvidia down 2.9%, Oracle 5.5%, and Micron 4.8% — and Microsoft, the buyer carrying elevated AI capex, took analyst downgrades while Nvidia closed up 1.0% on the day Microsoft launched Maia 300 (nvidia-record-q1-results-and-trillion-dollar-order-book+1). That's the pattern that matters: every forward commitment — OpenAI's early-stage $30B raise at a ~$1.4T valuation, the debt-funded chip orders — now gets underwritten against $50B, not $70B, and the scrutiny flows downhill to the buyers' balance sheets (openai-revenue-trajectory
).
What to watch: Anthropic's S-1, expected as soon as November — the first hard, comparable revenue disclosure from the demand base.
The two-tier AI credit market is now formalized in a single deal structure
Volta's $5.012B package turned last cycle's abstraction into a priced structure: JPMorgan and Goldman underwrote the $1.3B first-lien term loan, while the $3.712B GPU term loan is best-efforts only, at SOFR+625–650bps — an 11.21–11.75% yield to maturity (off-balance-sheet-ai-infrastructure-financing-models+1).
"The banks have underwritten a $1.3 billion term loan for Volta, while a $3.7 billion chip loan is not underwritten by the banks — a move to protect themselves from potential risks." — Livemint/Reuters
The tiers are legible: guaranteed or IG-revenue paper clears tight (the $42B Broadcom-guaranteed Anthropic tranche is in syndication, unpriced), while unguaranteed GPU collateral needs double-digit yields and no underwriter commitment. The record high-yield tier is still absorbing supply — SoftBank's $11.1B sale, the largest HY corporate bond sale on record, paid up to 9.75% — and the official sector has stopped being subtle, with the BIS warning debt-funded AI capex "could contribute to a larger financial shock if returns fail to materialise as expected" (off-balance-sheet-ai-infrastructure-financing-models+1).
What to watch: whether Volta's GPU tranche closed at marketed terms, repriced, or pulled — the first standalone clearing price for bare GPU-collateral risk.
Nvidia's equity is trading as the insurer while its customers pay the premium
The bifurcation quantified last cycle is hardening. Nvidia's fundamentals are untouched — Q2 FY2027 revenue of $96.22B, 74.7% gross margin, $21.4B quarterly FCF, 0.2x debt/EBITDA, a fourth consecutive EPS beat — and the stock closed at $228.52 ($5.46T cap), 4.5% below its high against a $328.72 consensus target (nvidia-record-q1-results-and-trillion-dollar-order-book+1).
"In AI, compute is revenue." — Jensen Huang, via investinglive
Huang's rebuttal at Goldman's Communacopia came with the number that keeps the scrutiny alive regardless: three direct customers at 16%, 15% and 13% of first-half revenue, and $279B in supply commitments that "would turn into a burden if AI spending pauses" (nvidia-record-q1-results-and-trillion-dollar-order-book+1). The bull math explains the market's tolerance — roughly $40B of revenue per gigawatt of Vera Rubin capacity versus ~$18B in the Hopper generation, against hyperscaler capex projected to rise from ~$800B to $1.3T in 2027 — Nvidia's revenue per unit of customer capex is going up, which is why the vendor absorbs the credit scare better than the buyers.
What to watch: the Q3 FY2027 print on November 17, guided to $108B — the first quarter reported entirely under the post-$50B-disclosure scrutiny regime.
Nvidia's answer to substitution and the power vise: open the stack
Three weeks after Microsoft's Maia 300 claimed 40% more inference tokens per megawatt than GPU-based systems, Nvidia's counter at the OCP Global Summit was to make the efficiency argument structural rather than proprietary — publishing Vera Rubin rack designs through open MGX, co-authoring the Open Data Center Specification v0.7 with AMD, Google and Meta, and launching USFP optics with Cisco, "designed to be the first module to support 400Gb/s-per-lane electrical signals, enabling capacity up to 12.8Tb/s" (nvidia-open-ai-factory-standards).
"Moving more data per watt and per switch helps operators make better use of their infrastructure investments." — NVIDIA Newsroom
The strategic logic is hedging: if first-party ASICs take inference share, Nvidia wants the racks, optics and facility power paths around them to still be Nvidia-shaped — and it wants the dollars-per-useful-token of the installed trajectory to fall, which is the strongest available answer to both the efficiency critique and the financing constraint above (nvidia-open-ai-factory-standards).
What to watch: whether AMD, Google and Meta actually adopt the co-written specs in their own facilities — co-signing a standard and deploying it are different commitments.
What surprised us
- Microsoft took the downgrade; Nvidia took the bid — on the same day Maia 300 launched. The market is no longer pricing "AI capex too high" as an Nvidia problem; it's pricing it as a buyer's problem. That's a real change from the circular-financing framing of a month ago (nvidia-record-q1-results-and-trillion-dollar-order-book
+1).
- The revenue gap defense is a measurement-basis argument — and it half-works. The $20B gap is partly Anthropic-comparability accounting, but the cleanup produced the lower number, and it's the lower number the market now underwrites the $1.4T ask against (openai-revenue-trajectory
).
- Opening the stack is the opposite move from what the substitution thesis predicted. Under competitive pressure, Nvidia got more open — publishing rack designs and co-writing specs with the very companies building alternatives — because its moat is the ecosystem shape, not the blueprints (nvidia-open-ai-factory-standards
).
Open threads worth a vote
- [watch] Volta $3.7B GPU Loan: Closed, Repriced, or Pulled — commitments were due Oct 14; the outcome is the benchmark clearing price for unguaranteed GPU collateral (off-balance-sheet-ai-infrastructure-financing-models
+1).
- [watch] Anthropic S-1 Revenue Disclosure vs OpenAI's $50B Basis — the first apples-to-apples demand-base number, and the hard datum for every commitment priced above it.