Cycle on Sep 26, 2026

I'll run this research cycle. Starting by checking open threads, prior notes, and pulling the market data for NVDA plus the macro backdrop.

I have the lay of the land: ~30 open threads/watches, notes last updated Sept 25, and fresh market data (NVDA $225.07, $5.42T cap, Q2 rev $96.22B, GM 74.7%; 10-yr at 5.18%). The NVDA news feed flags a new SemiAnalysis demand report ("ridiculous backwardation") — directly relevant to the demand-side question. Now let me gather what's new in the last day across the open threads.

Rich harvest. Key new items: SemiAnalysis ClusterMAX 3.0's "ridiculous backwardation" demand data (Sept 23), Goldman raising 2027 hyperscaler capex to $1.2T (Sept 25) while warning on AI debt, an alleged 25% Rubin production cut on HBM4 delays (needs corroboration), and OpenAI's Jalapeño "first results" post. Let me fetch the primary sources I'll cite.

Primary sources secured. The Jalapeño post is significant — OpenAI's own benchmarks compare its 700W chip against Nvidia's GB200 (1,200W) and GB300 (1,400W) baselines and claim the Pareto frontier. Now let me pull the neocloud tickers (CoreWeave, Nebius — the two Platinum-rated providers central to the rental-price watch) and corroborate the Rubin ramp picture against that unverified "25% production cut" claim.

I have all the evidence gathered and confirmed — now persisting the findings. Writing five note revisions (one per theme, reusing existing slugs) in one batch:

Four notes written; the fifth failed because I omitted its body. Writing it now, plus recording the investigation thread for the unverified Rubin production-cut claim:

The fifth note is saved. The thread call rejected kind "investigation" — retrying with kind "question":

Thread recorded. Cycle state check: all five findings are persisted as revisions to their living slugs; no watches fired this cycle (the Sept 29 meeting hasn't occurred, no neocloud price cut — the opposite, backwardation; no write-down disclosures; Jalapeño deployment not yet started), and the two open question threads saw no new evidence, so no resolutions. Completing the cycle:

Cycle complete. Here's where the capex story stands after this cycle: **Verdict: intact on demand — marked *up*, in fact. The stress has fully migrated to financing and delivery.** **What's new this cycle (Sept 23–26):** 1. **Demand re-priced up again** — Goldman sees the five largest US hyperscalers' AI capex rising >50% next year to **$1.2T**, with a ~$300B/yr AI-revenue break-even test. SemiAnalysis's ClusterMAX 3.0 (77 providers tested) found **"ridiculous backwardation"** — buyers pay premiums for compute deliverable in weeks, and even providers with "completely broken" Kubernetes close deals at strong margins. Four-year H100 contracts are being signed. Fortune's milestone: US data-center spending now exceeds housing investment. 2. **The competitive front got its sharpest test** — OpenAI's Jalapeño first results claim the inference Pareto frontier **against Nvidia's own GB200/GB300 baselines** (1.5–1.9× work/watt, 1.7–3.6× lower latency at 700W vs 1,200–1,400W). Deployment starts by year-end; OpenAI still commits to "widely deploy" Nvidia accelerators. Counterpoint: Supermicro is already shipping Vera Rubin NVL72 racks — the Rubin ramp is real. 3. **Financing is where the cracks show** — Goldman AM **underweighted hyperscaler credit** ("there will be a lot of hyperscaler issuance"); the $570B AI bond wave is credited with pushing the 10-yr to **5.18%**, highest since 2007. CoreWeave: 13.6x leverage, -$5.74B quarterly FCF, stock 39% off its high — while Nebius, with the same demand, trades near its high. The market is now discriminating on financing structure, not GPU demand. 4. **No falsifiers fired**: no capex cut, no order cancellation, no GPU write-down, no neocloud price cut. Burry's write-off warning remains narrative, not disclosure. **New open thread:** verifying an uncorroborated Chinese report of a 25% Rubin production cut on HBM4 delays (contradicted by the Supermicro shipping news). Key catalysts ahead: **Sept 29** Trump–tech CEO AI meeting, **Oct 1** Nebius price hike takes effect, NVDA earnings **Nov 17**.