Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The demand leg of the story is stronger than ever — Meta's capex is headed for $145B (+101%), every major platform's backlog sits at or above $496B, and the Nebius price hike holds — but this cycle the first real cracks appeared on the delivery and financing side, not demand: Oracle invoked force majeure on a gigawatt-scale campus over a power pipeline, and Huawei put dated 2027 competition on the board. The summit resolved as pure status quo, which is exactly what the capex story needed.
The demand case re-priced up again — and the falsifier didn't fire
Demand is now confirmed on four independent gauges at once: capex guidance, contracted backlog, consumption, and price. Fortune's framing of Meta is the headline: "Meta plans to spend up to $145 billion in capital expenditure in 2026, a roughly 101% jump from the $72.2 billion it spent just one year earlier" — spending that "puts Meta above every country's military budget from 2025, save for the United States, China and Russia" (Fortune, Sept 24).
The neocloud price signal — the demand-side tripwire opened last cycle — did not fire. Nebius's second hike in three months takes effect Oct 1, with the newest, scarcest silicon taking the biggest hit: B300 moves from $7.85 to $9.50 (~21%), and memory tiers jump 41% (Shattered.io, Sept 23). Notably, "this time the scarcity is showing up across the entire hardware stack at once" — not just GPUs. Bernstein's cloud audit adds consumption evidence: Google's API token flow hit ~22B tokens/minute, more than doubling in six months, with demand "consistently outpacing supply" (TechFlow, Sept 23).
What to watch: whether the Oct 1 Nebius hike holds through the quarter — the first price cut or utilization disclosure remains the cleanest demand-side falsifier.
The first crack is a delivery crack, not a demand crack — and it's about power
The binding constraint on the buildout just moved from analyst commentary into contract language. On Sept 24, Oracle sent a force majeure notice to Stack Infrastructure over Project Jupiter — its $165B, 2.45 GW AI campus in New Mexico — because the 17-mile gas pipeline feeding its Bloom Energy fuel cells has slipped from August 2026 to February 2027 after repeated right-of-way permit refusals. As the reporting put it: "A data center without electricity is a stranded building" (TradingNews, Sept 24).
The market read it as a conversion-timeline event, not a demand event. Oracle fell as low as $133.48 (-7.66% intraday) before closing down 4.9%; Bloom Energy fell 6.52% and Blue Owl — down 46% over the past year — dropped 4.39% as "AI buildout timelines stretch." Meanwhile the same quarter showed Oracle "delivered more than 300,000 GPUs to its AI cloud customers, almost triple the capacity delivered in Q4. GPU utilization stood at 97.9%" (TradingNews). That's the precise distinction the deployment-gap debate has been waiting to see priced: investors sold the delivery schedule, not the backlog. The macro layer amplified the pain — the 10-year Treasury touched 5.15%, its highest since July 2007.
What to watch: the Feb 1, 2027 pipeline completion target, and whether lenders tighten terms on Oracle-anchored projects — if financing burden migrates back onto hyperscaler balance sheets, that's a cost-of-capex event, not a demand event.
Competition now has specific quarters attached to it
The overhang that stalled Nvidia's breakout last cycle just got dates. Huawei announced it will launch the Ascend 960DT in Q1 2027 and the Ascend 960PR in Q3 2027, claiming its "Tau Scaling Law" — unveiled in June — "enables it to make high-performing chips without Western-made lithography equipment," with plans to link "supernodes" of about 100,000 processors to train frontier models (Reuters Breakingviews, Sept 24). The ecosystem is real: Huawei and peers including Alibaba's T-Head took 41% of China's AI chip market in 2025, and DeepSeek is working closely with Huawei on deployment.
On the in-house-silicon front, Google "recognized TPU system sales revenue for the first time" in Q2, with most revenue landing in 2027 — and Bernstein's structural verdict is blunt: "Google is the only hyper-scale provider truly vertically integrated with both its own frontier models and custom TPU chips" (TechFlow, Sept 23). Neither touches Nvidia's near-term numbers — but the 2027+ story now has a calendar.
What to watch: the Ascend 960DT's Q1 2027 launch — the first real test of whether China's domestic stack erodes Nvidia's position in third markets, not just the already-ceded China market.
The summit resolved as status quo — the exact outcome the capex story needed
Both tail risks around the Trump–Xi summit resolved in the buildout's favor. The state dinner — Huang, Altman, Cook, Bezos, Pichai and Musk all in the room — produced "no sign that a closed-door meeting yielded breakthroughs on thorny issues such as AI, trade, Taiwan" (Reuters, Sept 24). Trump pre-set the policy: "I want to leave it exactly where it is." Export controls hold — "Trump is likely to maintain firm export controls on US tech Xi wants loosened" — and the one deliverable was a trade truce extended to January 10.
That means no pacing pact hardened into deferred compute orders, and Nvidia's zero-China-revenue Q3 assumption stands intact — the upside option is untouched rather than extinguished. The regulatory overhang now concentrates on a single dated event: a confirmed Sept 29 meeting between Trump, Speaker Johnson, and tech CEOs on AI (Reuters) — exactly the venue where audit-gate frameworks could surface, though Trump told the UN he "rejects any attempt to construct a globalist scheme to control" AI.
What to watch: the Sept 29 meeting — any third-party audit/deployment-gate framework emerging from it is the remaining regulatory path to slower compute deployment.
What surprised us
- The first blood came from a gas pipeline, not a balance sheet. After weeks of watching for accounting cracks (DSO stretch, useful-life extensions — Microsoft quietly moved from 15 to 25 years this quarter), the first contractual stress on the buildout arrived via a New Mexico right-of-way dispute. Power and permitting, not financing, drew first blood.
- Scarcity went vertical. Nebius's hike hit memory tiers (+41%) harder than some GPU tiers — the demand story is now squeezing the entire hardware stack simultaneously, which ties Nvidia's capex story to memory supply in a way it wasn't before.
- The summit's "nothing" was the bull case. A US-China AI accord or pacing commitment was the bear scenario; "leave it exactly where it is" plus a truce extension is the outcome that preserves both the zero-China assumption and the free option.
- Huawei's R&D intensity dwarfs Nvidia's. Over a quarter of sales in H1 (~$18B) versus ~9% at Nvidia — the sanctioned challenger is out-investing proportionally.
Open threads worth a vote
- Sept 29 Trump–Johnson–Tech CEO AI Meeting: Outcome for Pacing/Audit Gates — fires on any pacing commitment, audit/deployment-gate framework, or federal data-center support.
- Huawei Ascend 960DT Launch (Q1 2027) — First Test of China's Domestic AI Stack — the dated competitive milestone now on the board.