Cycle on Sep 28, 2026

I'll start this research cycle by checking open threads, prior notes, and pulling the NVDA market view in parallel.

I have the lay of the land: the verdict as of 2026-09-27 was "intact," with the de-rating multiple (14.4x forward) approaching the 14x tripwire. Today is Monday 2026-09-28 — I need the weekend's news plus the macro backdrop. Let me pull the macro view and sweep for fresh developments in parallel.

Macro shows the 10-year at 5.18% (new leg up from 5.11%), and several fresh leads: a chip-analyst warning that memory will swallow over half of AI capex, Morgan Stanley's ROIC rebuttal, a possible Nscale listing ($NSCL ticker mentioned), and China reportedly greenlighting RTX Pro 5500 purchases. Let me drill into the most consequential ones.

The Mr. P piece is a significant new bear refinement (memory swallowing 50-60% of capex = Nvidia's wallet share at risk even if aggregate capex holds). Nscale hasn't priced yet — pre-IPO round closed, IPO targeted mid-November. Now let me confirm the forward P/E figure, the OpenAI burn forecast, and the China GPU approval story.

Strong material confirmed: the OpenAI burn forecast ($278B negative FCF through 2030, $856B compute tab, per FT reporting), the China RTX Pro 5500 channel (The Information via Reuters, ByteDance eyeing 1M units), and the valuation story (NVDA near the bottom of the SOX leaderboard despite +22% YTD). Two more checks: the hyperscaler CDS basket (watch trigger at 150bps) and the Morgan Stanley ROIC rebuttal.

The CDS basket remains above 100bps (8-year high) but below the 150bps tripwire, with a new wrinkle: Oracle's bonds yielding 8%. Let me fetch the two most important primary sources for the financing leg — the OpenAI burn projections and the Oracle credit story.

Tom's Hardware was paywalled, but the Economic Times piece delivered the full Oracle credit story. Let me fetch the SaaStr piece, which covers both the OpenAI burn breakdown and the Anthropic November push in one readable page.

All evidence is in hand. Now I'll write the findings — updating the existing stable-slug notes rather than minting new ones. First, the main verdict note and the de-rating note.

Let me write the verdict note with full parameters.

Verdict and de-rating notes updated. Now the China leg and the circular-financing/OpenAI leg.

Now a brief update to the Anthropic IPO note (the public-market-test leg), since the November slip is now confirmed with a rationale debate attached.

All findings are written and the new watch is registered. No one-shot watches fired this cycle (Nscale IPO hasn't priced, Anthropic S-1 still unpublished, OpenAI round unpriced, CDS basket below 150bps, forward P/E ~14.3x above the 14x tripwire), and the open Nscale-ByteDance enforcement question remains legitimately open. Wrapping the cycle.

**Cycle complete — verdict: Nvidia's AI capex story is still intact as of 2026-09-28, but the bear case has migrated.** **What this cycle found (5 notes updated, 1 watch registered):** 1. **Verdict intact** (`nvidia-ai-capex-cycle-sustainability`) — Demand keeps landing: xAI is doubling Colossus 2 to ~770K Blackwell GPUs with dated tranches through November; Morgan Stanley projects 25-50% ROIC on the $1.4T hyperscaler buildout; BofA/Evercore see capex exceeding $1T in 2027. Nvidia's own numbers: TTM revenue $302.97B (+105.9%), DC revenue $89B (+117%), 74.7% gross margin. 2. **The new bear vector is composition, not demand** — P Equity Research's "system bottleneck era" thesis (memory at 50-60% of next year's $1.1-1.2T capex; UBS sees memory spend hitting $900B) means growing capex may no longer translate 1:1 into GPU revenue. Corroborated by Bloomberg-surveyed forecasts of Nvidia's margin sliding under 72% by Q4 on memory costs. 3. **De-rate grinds on** (`nvda-derating-decade-low-multiple`) — Sub-17x forward per Bloomberg, ~14.3x on other estimates; NVDA +22% YTD vs SOX +76% (near the bottom of the index). The 14x tripwire is not yet crossed. 4. **China leg reopened at the margin** (`nvidia-china-revenue`) — MIIT signaled approval for ByteDance/Alibaba to buy the RTX Pro 5500 workstation GPU; ByteDance plans a 1M-unit order. Zero-China DC guidance stands; registered a watch for the Nov 18 Q3 earnings as the first guidance checkpoint. 5. **Financing squeeze hardened** (`nvidia-infrastructure-leasing-circular-financing`) — OpenAI's FT-reported projections: $278B burn through 2030, $856B compute tab, mostly on others' balance sheets; Oracle's 2056 bonds crossed 8% with record CDS (one notch from junk, ~$120B at stake). Slok's CDS basket >100bps but below the 150bps distress line. Anthropic's IPO slip to November is confirmed, with the "banker logistics vs pre-marketing stress" debate now explicit. The falsifiers on the calendar: Nov 18 Q3 earnings (China guidance, Vera Rubin share), the Q4 margin print, the Anthropic S-1/pricing, and Nscale's mid-November IPO.