Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The demand tape printed hot again — Nebius's second hike took effect Oct 1, Goldman raised its 2027 hyperscaler capex estimate to $1.2T, and Barclays' "napkin math" now sees $401B of Nvidia hyperscaler revenue in 2027 (big-tech-capex-reaches-710b-in-2026). But the fight has moved to the financing layer: Michael Burry published a direct attack on the residual-value math behind GPU-collateral lending, and the first real-world stress test — Project Jupiter's force majeure — already knocked an $18B loan to ~90 cents on the dollar. Meanwhile the 10-year closed September at 5.29%, above the 2007 peak it matched last cycle.
The demand gauge printed hot — but the market stopped paying up for it
The pricing tape didn't just hold; it steepened. Nebius's on-demand B300 rate climbed 21% on Oct 1, from $7.85 to $9.50 per hour — the second hike in three months, leaving the same chips roughly 56% more expensive than in early May (big-tech-capex-reaches-710b-in-2026).
"[W]e sold out of capacity because, as fast as we bring capacity online, we can sell it." — Nebius CFO Dado Alonso, via The Motley Fool
On top of that, Goldman now sees the five big US hyperscalers spending ~$800B in 2026 and ~$1.2T in 2027 — above consensus — with Barclays' Tom O'Malley estimating Nvidia hyperscaler revenue of $401B in 2027 and keeping Overweight with a $275 target (Invezz). The tension: AI infrastructure stocks have de-rated to ~22x forward earnings from 32x in April even as every fundamental gauge tightens — the market is pricing execution and financing risk, not demand weakness. Both tracked falsifiers (a neocloud price cut, a hyperscaler capex cut) remain unfired.
What to watch: whether Nebius's first capacity-auction economics ($/MW contracts at $40–50M) become the new quoted baseline — that's where scarcity gets priced next.
The financing flywheel hit its first public resistance
GPU-collateral lending crossed two thresholds in the same week: formalization and pushback. GMI Cloud closed $668M on Sept 30 — APAC's first GPU-backed syndicated loan, with Nvidia itself in the equity and CTBC leading the debt — and in the waterfall, Nvidia sits "effectively in a quasi-super-senior position" ahead of the banks financing its own chips (AI Weekly). Days later, Reuters reported lenders questioning whether Nvidia's chips hold value long enough to serve as collateral, with Nvidia capping residual-value guarantees at no more than 25% (Invezz/Reuters).
"Wall Street is much more conservative." — Tony Trzcinka, Impex Asset Management, via Invezz/Reuters
This matters because the financing layer is where demand becomes leverage. The supplier of the collateral getting paid before the lenders financing it is an extraordinary negotiating position — and exactly the structure lenders are now pushing back on. The FT-brokered GPU-collateral insurance talks remain open with no deal, terms, or collapse announced (off-balance-sheet-ai-infrastructure-financing-models+1).
What to watch: whether the first $500B-platform transaction closes with residual-value guarantees attached — that's the tripwire for whether the flywheel extends or stalls.
Burry made the bear case an accounting argument — and bought the puts to match
The overbuild debate got its most pointed articulation yet. In "Don't Believe Your Lyin' Eyes," Burry attacks the Nvidia investor-deck slide bulls celebrated, arguing its residual-value chart "does not reflect actual resale prices for second-hand equipment" but a DCF model built on an assumed eight-year chip life — and that rising A100/H100 rental rates show "scarcity due to memory and power shortages, not GPU die shortages and not durability" (Stocktwits). His 1968 analogy: leasing companies using eight-year depreciation against IBM's four were right about hardware longevity and still went insolvent when System/370 launched.
He then covered his common-stock shorts and replaced the NVDA short with September 2027 puts struck in the mid-$100s — roughly 45% below the $230.86 close (gpu-deployment-gap-overbuild-debate). The essay is an argument about assumptions, not an event — no write-down or depreciation reversal has fired — but its target, the residual-value evidence underwriting the entire collateral lending stack, is the load-bearing assumption of the financing layer above.
What to watch: any GPU inventory write-down or depreciation-schedule change in an upcoming 10-Q — the first hard datapoint that would move this from essay to event.
Power, not capital, is the constraint — and the credit market just priced it
Project Jupiter's force majeure reached the loan market. The ~$18B construction syndicate (Blue Owl-developed, Oracle-leased, Stargate buildout) is quoted at 89–91 cents on the dollar after Oracle invoked force majeure over a power delay — a 17-mile gas pipeline slipped six months to February 1, 2027, after repeated New Mexico right-of-way denials (Global Data Center Hub).
"A late-power notice in New Mexico hit a loan banks already struggled to sell." — Global Data Center Hub
The "power is the binding constraint" thesis now has a price: a tenant protecting itself from a power delay knocked roughly 10 points off an $18B loan. Bloom Energy whipsawed 9% down and 8.5% back up on Oracle's reaffirmation (Investing.com). A new thread opened on the November air-quality permit decision for the 2.46-GW microgrid — the next binary.
What to watch: the November permit decision; if it slips, the 90-cent print becomes the template for how AI-infrastructure credit reprices.
Rates broke the 2007 peak — and NVDA is absorbing it anyway
The 10-year closed September at 5.29%, above the 5.26% 2007-peak level flagged last cycle, with fed funds at 3.75% after the September hike (first-rate-hike-since-2023). NVDA sits at $230.86, 1.8% below its 52-week high at 29x trailing earnings — the market is still paying up, but Goldman's count shows the AI complex compressing regardless (22x forward vs 32x in April). The Fed signaled one more 2026 hike (median 4.1%); a second would push the discount rate on every debt-funded layer of this buildout higher still.
What to watch: the late-October FOMC decision — a hold would relieve the overhang; a second consecutive hike would compound the financing strain already visible in the Jupiter loan and Oracle's negative free cash flow.
What surprised us
- Nvidia's position in the GMI Cloud waterfall. The collateral supplier being quasi-super-senior to the banks financing the collateral is the kind of structure that only gets agreed at the top of a financing cycle — and it's now the explicit thing lenders are renegotiating.
- Burry's strongest demand evidence cuts both ways. His reframe of rising H100 rental rates as memory/power scarcity rather than GPU durability is genuinely uncomfortable: it means the pricing tape we track as a demand signal could invert on component-supply news, not on an AI demand shock.
- The credit market found the constraint before the equity market. Power delays showed up as a 90-cent loan quote, not an NVDA drawdown — the financing layer is now the most price-sensitive part of the complex.
- The de-rating happened on a hot print. AI infrastructure at 22x forward from 32x in April, in the same week Goldman raised its 2027 capex estimate above consensus — the market is pre-paying for the financing risk before it materializes.
Open threads worth a vote
- [watch] Project Jupiter air-quality permit decision (November 2026 deadline) — granted, the power path clears ahead of the Feb 1, 2027 pipeline; denied or slipped, the force majeure hardens and the 89–91 cent print becomes the template.