Cycle on Sep 27, 2026

I'll start this research cycle by reviewing the open threads, prior notes, and pulling the market data for NVDA and the macro backdrop.

I have the full picture now. Key open items this cycle: 1. **Verify the reported 25% Rubin production cut** (2M → 1.5M units on HBM4 validation delays) — the top open question thread. 2. **Sept 29 Trump–Tech CEO AI meeting** fires in two days — check for pre-meeting positioning. 3. Any fresh demand-side or financing-side developments since Sept 25. Market data already pulled: NVDA at $225.07 (P/E 28.45, 74.7% GM, 4/4 beats, TTM rev +105.9%), 10-yr at 5.18%. Now let me verify the Rubin claim and scan the last few days of news.

The Rubin-cut claim isn't corroborating so far — instead I see counter-evidence (Supermicro shipping NVL72 racks, Micron volume HBM4 shipments, UBS *raising* its 2027 GPU forecast to 8.8M units). Let me fetch the two most substantive recent articles and search directly for the cut claim.

The Rubin-cut claim is taking shape: it traces to **KeyBanc analysts** — and the Chosun article citing it is dated **April 9, 2026**, suggesting the Chinese report may be recycling stale news. Meanwhile Rubin is now shipping (Supermicro NVL72 racks Sept 23, Micron volume HBM4, production shipments began August). Let me pin down the KeyBanc claim's date and check the Sept 29 meeting / Nebius pricing threads.

The Rubin-cut claim is now pinned: **KeyBanc's 2M→1.5M estimate dates to April 8–10, 2026** (MLQ, Chosun, The Register) — the recent Chinese report is recycling stale news. Since then Rubin has ramped (Supermicro shipping NVL72, Micron volume HBM4, UBS *raising* 2027 to 8.8M units). The Sept 29 meeting is confirmed for Tuesday; the Nebius Oct 1 price hike stands, with a new Rothschild Redburn sell case as the articulate bear counter. Let me grab the neocloud bear-case article and sweep for any hyperscaler capex developments I might be missing.

The neocloud bear case is now well-sourced (Rothschild Redburn, Sept 21). One more fetch to properly cite the Rubin ramp counter-evidence (Supermicro shipping), then I'll write the findings.

I have the Rubin verification complete. Before writing the demand-side note, let me fetch one solid source for the Nebius Oct 1 price-hike specifics (currently I only have search snippets).

I have all the evidence verified and gathered. Now writing the three findings — updating the existing notes for each tracked subject (memory/HBM bottleneck, demand side, overbuild debate) — then resolving the Rubin thread.

The call failed — I omitted `topic_id` and `slug`. Retrying with all required parameters.

Note 1 written. Now the demand-side update and the overbuild-debate update, in parallel.

All three findings are written. Now resolving the Rubin production-cut question thread, which this cycle's verification answers.

Thread resolved. No new watches or threads are needed — the existing watches (Sept 29 meeting, neocloud price cuts, hyperscaler capex cuts, GPU write-downs, FOMC, HBM pricing) already cover every forward-looking item this cycle surfaced. Ending the cycle.

Cycle complete. Summary of what this cycle found on **"Is Nvidia's AI capex story still intact?"**: **Answer: Yes — and the constraint has fully migrated from demand to delivery.** **1. The Rubin "production cut" claim is stale news — resolved the open thread.** The Chinese report of a 25% cut (2M→1.5M units) on HBM4 validation delays traces verbatim to **April 8–10, 2026** KeyBanc/TrendForce reporting. Since then: Rubin production shipments began in August, Supermicro has been shipping Vera Rubin NVL72 racks since Sept 23, Rubin is expected at ~20% of Q3 data-center revenue, and UBS *raised* its 2027 forecast to 8.8M GPUs. Kress still expects supply to be the bottleneck "at least through the end of fiscal year 2028" — supply rationing revenue is not demand destruction. **2. The demand side keeps re-pricing up.** Nebius's Oct 1 hike (GPU rates +17–21%, memory +41%, second hike in three months — 2022-era H100s getting *more* expensive) means the demand-falsifier watch (first rental price cut) remains unfired. The bear case has migrated: Rothschild Redburn's new Sell ratings don't require demand to collapse, only for scarcity pricing to normalize as supply broadens — a later-dated, different falsifier. **3. The overbuild debate got its first hard number — at the vendor.** Nvidia's $0.4B charge for excess H200 inventory is immaterial at $303B TTM revenue, but it's the first actual write-down disclosure: generational churn stranding the previous chip at Nvidia's own balance sheet while older silicon still appreciates in the rental market. **Next catalysts already on watch:** the Sept 29 Trump–Johnson–tech CEO AI meeting (fires in two days — next cycle checks for pacing pacts, audit gates, or federal data-center support), and Nvidia's Q3 FY2027 print on Nov 17 ($108B guide, ~74% GM).