Cycle on Sep 25, 2026

I'll run this research cycle on whether Nvidia's AI capex story is still intact. Starting by checking open threads, prior notes, and pulling the market view for NVDA.

I have the full picture of threads and prior notes. Key near-term items: the Trump–Xi state dinner was last night (Sept 24) — a one_shot watch on AI-executive outcomes; the neocloud price-cut tripwire; and the deployment-gap debate. Now let me gather fresh evidence on the summit outcome and any capex-story developments.

The search results are mostly social-media posts with conflicting details. I need proper news coverage of the summit outcome and any capex developments. Let me refine.

Nebius price-hike tripwire confirmed still intact (no cut — hike takes effect Oct 1). Now I need real news coverage of the summit outcome. Let me search authoritative sources directly.

The picture is forming: the summit was "light on substance" on AI, but a **Sept 29 Trump–tech CEOs AI meeting** is now scheduled. Let me fetch the key Reuters articles to get quotable details.

Summit outcome confirmed: no AI/chip breakthroughs, export controls maintained, and a **Sept 29 Trump–tech CEO AI meeting** now scheduled. Let me get the Huawei challenger story and check for fresh demand-side capex signals.

Major developments: Huawei's Ascend 960 challenge, Oracle's force majeure on Project Jupiter, Meta's $145B capex, and Bernstein's cloud-growth data. Let me get details on the Oracle force majeure (a potential deployment-crack signal) and the Meta capex story.

The Oracle force majeure is a significant development — the power bottleneck just produced its first contractual stress event at gigawatt scale. Let me ground two more pieces: the China/Broadcom survey (competition front) and the Bernstein cloud-quarterly data (demand side).

I have strong evidence on all fronts. One more quick fetch to confirm the Nebius pricing status (the demand-side tripwire), then I'll write the findings.

All evidence is in hand. Now I'll write the four finding updates — summit outcome, Huawei/competition, Oracle power stress, and the demand-side capex data — each grounded in what I fetched this cycle.

First note saved. The second call went out empty — let me write the remaining three notes properly.

Two more saved. Now the demand-side capex update — the strongest evidence this cycle that the story's demand leg is intact.

All four findings are saved. Now recording the watch events (two resolutions, one escalation log) and registering two new forward-looking watches.

All events and watches are recorded. The cycle is complete — here's the picture. ## Is Nvidia's AI capex story still intact? Yes — but the stress has migrated from demand to delivery and financing. **What this cycle found:** **1. The summit resolved in the story's favor** ([[nvda-china-market-access-and-trump-xi-summit]]). The Sept 24 Trump–Xi state dinner — with Huang, Altman, Cook, Bezos, Pichai, and Musk in the room — produced "no sign that a closed-door meeting yielded breakthroughs." Trump's position: AI/SI he wants to "leave it exactly where it is." Export controls hold, the trade truce extends to Jan 10, and no pacing pact emerged that could harden into deferred compute orders. Both tail risks (regulatory pacing, China escalation) resolved as status quo. **2. Demand re-priced UP again** ([[big-tech-capex-reaches-710b-in-2026]]). Meta's 2026 capex hit $145B (+101% YoY) — more than any military budget except the US, China, and Russia. Bernstein's audit shows backlogs of $678B (MSFT), $664B (ORCL), $514B (GOOGL), $496B (AWS), $104B (CoreWeave), token flow doubling in six months, and "demand consistently outpacing supply." The Nebius price hike (second in three months) takes effect Oct 1 with no cuts anywhere — the demand-side falsifier never fired. NVDA itself: $5.45T cap, 74.7% gross margin, 4/4 earnings beats. **3. The first crack is delivery, not demand** ([[nvidia-lancium-power-investment-and-energy-bottleneck]]). Oracle sent a force majeure notice on Project Jupiter — its $165B, 2.45GW Stargate campus — because the gas pipeline slipped to Feb 2027 after permit refusals. "A data center without electricity is a stranded building." ORCL -4.9%, Bloom -6.5%, Blue Owl -4.4% (down 46% YoY), all against a 5.15% 10-year. Meanwhile Oracle's GPU utilization is 97.9% — the market sold the conversion timeline, not the backlog. **4. Competition now has dates** ([[custom-silicon-and-datacenter-hardware-competition]]). Huawei's Ascend 960DT lands Q1 2027 with claims of frontier performance without Western lithography; Google booked its first TPU system revenue with most of it landing in 2027. **Watches set:** the Sept 29 Trump–Johnson–tech CEO AI meeting (the next regulatory catalyst) and Huawei's Ascend 960DT launch. The Oracle pipeline watch logged the escalation and stays open to Feb 1.

The research cycle is complete. Four findings were updated under their existing stable slugs, three watch events were recorded (two resolutions, one escalation), and two new dated watches were registered for future cycles to check.