Nvidia Capex Digest
TL;DR
The capex story is intact on substance — demand re-accelerated again this week, with SemiAnalysis reporting "GPU supply has gone to zero" and xAI on track to roughly double its chip count by year-end (nvidia-ai-capex-cycle-sustainability). But the market's public tests are slipping, not failing: Nscale's IPO is still unpriced and Anthropic's S-1 still isn't on EDGAR (anthropic-ipo-public-market-test
). The new headline constraint is physical, not financial — Oracle invoked force majeure on a 2.45 GW Stargate campus because it cannot get gas (hyperscaler-capex-surge-2026
).
The binding constraint moved from chips to power — and it's now a contractual event
The buildout's bottleneck just graduated from background risk to legal paperwork.
"Oracle has sent a force majeure notice to the developer of Project Jupiter, a Stargate data center campus in New Mexico… the notice would allow the company to delay payments should the facility miss its 2028 target to come online." — TechCrunch
The cause is energy supply: the Bloom Energy fuel-cell campus is waiting on an Energy Transfer gas pipeline delayed nearly six months to February 1, 2027 after regulators repeatedly denied permits, and New Mexico's environment department faces a November 23 deadline on the air-quality permit (hyperscaler-capex-surge-2026). The contrast is the whole story — Oracle delivered 300,000+ GPUs in its latest quarter, roughly triple the prior one, while its own flagship campus waits on a 17-mile pipeline. Both parties are minimizing it ("Project Jupiter remains on our planned schedule"), and Oracle shares fell more than 3% on the news (hyperscaler-capex-surge-2026
).
What to watch: The Nov 23 New Mexico air-permit decision — the first hard test of whether force majeure is an outlier or a template.
Demand is so extreme that even broken infrastructure sells
The physical demand signal didn't just hold — it re-accelerated, and the strongest evidence yet landed against the overcapacity thesis.
"slavering investors have just about run out of pockets to stuff checks into. GPU supply has gone to zero." — SemiAnalysis, ClusterMAX 3.0
ClusterMAX 3.0 rated 77 GPU-cloud providers from hands-on testing and 200+ customer interviews, finding "ridiculous backwardation" — customers paying steep premiums for compute available within weeks — and only 19 of 77 earning a Medallion rating (nvidia-ai-capex-cycle-sustainability). Musk says Colossus 2 already runs 110,000 GB200s and 440,000 GB300s, with another 220k GB300 "fully operational next week" — implying ~1.44M GPUs by year-end (nvidia-ai-capex-cycle-sustainability
). This matters because it directly rebuts the depreciation argument: four-year H100 contracts are being signed now, and a backwardated rental curve is the opposite of the glut that thesis requires (gpu-depreciation-debate-hyperscaler-earnings
).
What to watch: The Ornn/Blackwell rental rate (< $3.00/hr) — the live tripwire for the depreciation debate.
The financing web got itemized — and Nvidia is rationing it
The circular-financing question finally has a line-by-line answer, and the answer includes a behavioral shift.
"Nvidia in its 10-Q disclosed $530B of gross off-balance sheet guarantees across six different line items – a jump from $184B of obligations disclosed in the prior quarter." — SemiAnalysis, Backstop Universe
The itemization resolves a standing question: the $279B of supply/capacity commitments is "primarily memory" — an input-side commitment, not customer circularity; the demand-side backstops are separate lines, including $108.5B guaranteeing SB Energy's 4.25 GW campus leased to OpenAI for twenty years (nvidia-infrastructure-leasing-circular-financing). The key change: Nvidia has paused new AICP revenue-floor deals and pivoted to residual-value guarantees costing $9.4B/GW versus $59B/GW — repricing its backstop business, not abandoning it (nvidia-infrastructure-leasing-circular-financing
). Meanwhile Nscale raised $3.36B pre-IPO from Third Point, Apollo, Citadel, and Nvidia's $1B convertible — blue-chip capital underwriting the riskiest neocloud leg even as ClusterMAX lists Nscale "Not Recommended – Unavailable."
What to watch: The late-November 10-Q — whether total commitments cross the $350B tripwire.
The public tests slip to November — delay, not distress
Last cycle's dated catalysts haven't resolved; they've been postponed, which keeps the price-discovery question open.
"As of September 26, the SEC EDGAR database contains no public Anthropic filing. The world's largest IPO – a $2 trillion target that would exceed every listing in stock market history – is running behind schedule." — Yahoo Finance
Anthropic's revenue is accelerating — from a $14B annualized run rate in February to $47B by mid-May, targeting $100B annualized for 2026 — so the slip looks like timing, not demand (anthropic-ipo-public-market-test). But the delay is attributed partly to the collapse of the AI-safety coordination thesis, including a Sherman Act cartel suit filed Sept 18 and a narrowing capital-markets window. The equity still refuses to re-rate: NVDA closed at $225.07 ($5.42T cap, 28.5x trailing) — the market pays a mid-20s multiple for 106% growth and 74.7% gross margins, still pricing margin erosion and financing risk (nvidia-ai-capex-cycle-sustainability
).
What to watch: EDGAR for the Anthropic S-1, then pricing — final valuation vs the ~$2T target and whether Nvidia's reported $10B anchor materializes.
What surprised us
- SemiAnalysis conceded the depreciation debate's untestable point. "No one will take on the risk of assessing a residual value on 4, 5, or 6 year old GB300's today" — realized rentals favor long GPU lives, but the terminal-value market for old Blackwell racks literally doesn't exist yet, which is exactly why Nvidia's guarantees matter (gpu-depreciation-debate-hyperscaler-earnings
).
- The supply-commitment panic was mostly a misread. The $279B headline number that looked like customer circularity is mostly memory purchases — the debate's most-cited figure was the wrong line item (nvidia-infrastructure-leasing-circular-financing
).
- Polymarket puts an AI-industry downturn by Dec 31 at about 10% — prediction markets are far calmer than the equity's de-rate implies (gpu-depreciation-debate-hyperscaler-earnings
- OpenAI and Anthropic are diverging on going public. Anthropic's S-1 is late but coming; Altman says "right now would be an ill-advised moment to go public" — no 2026 IPO (anthropic-ipo-public-market-test
).
Open threads worth a vote
- [watch] New Mexico air-quality permit decision on Project Jupiter fuel cells (Nov 23 deadline) — a denial would validate force majeure as a template for power-constrained buildouts; approval keeps the 2028 target plausible.