Cycle on Oct 9, 2026

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

I have the full picture of open threads (many dated watches), the note set, and NVDA's current market view. Key observations to pursue: 1. **NVDA fell 2.9% on Oct 8** — need to find why (first notable down day in the recent run). 2. **Firmus Grid IPO postponed** — extends the equity-exit-channel story. 3. **Anthropic S-1** was expected early-to-mid October — check if it dropped. 4. **Macro leg**: 10-yr Treasury watch at 5.5% — need current level. Let me search for the last 24 hours of news and pull the macro view.

Major development found: **the Oct 8 selloff was triggered by a report that OpenAI's annualized revenue (~$50B) is well below the ~$68B figure previously circulated** — a direct hit to the demand leg. Anthropic's public S-1 still hasn't hit EDGAR. The 10-yr sits at 5.28%, below the 5.5% tripwire. Let me read the primary articles.

The Oct 8 selloff is now clear: FT reported OpenAI's annualized revenue at ~$50B vs the ~$68B/$70B previously circulated — a reporting-methodology dispute that hit the whole AI complex (NVDA -3%, ORCL -6%, CRWV -8%, XLK worst day since Sept 14). Now let me check the financing-leg watches: the SpaceX $40B debt raise, the Volta $5B loan pricing (~Oct 14), and the Firmus Grid IPO postponement.

The financing-leg watches have updates: SpaceX's $40B raise is confirmed in talks (not yet priced), and 9fin reports the Volta $5B loan is **oversubscribed** with bids due Oct 14. Firmus Grid's IPO collapse adds to the equity-exit-channel story. Let me read the 9fin and Firmus articles to confirm details before writing.

The 9fin piece is rich: Volta oversubscribed confirmed, but also sharp secondary markdowns on AI data-center bonds (Tract -11 points, Coatue/Fluidstack 8.875%→11%+ yields), Texas permit freeze fallout, and a Brookfield/Naver ~$9B GPU loan. Let me get the Firmus details and check the Google Finland data-center halt.

I have strong evidence across all legs now. One more verification: TSMC's quarterly revenue (+50% YoY claim) matters for the demand leg and the TSMC watch. Let me confirm.

I have all the evidence needed. Writing the five note updates now — each a revision of an existing slug, grounded in this cycle's sources.

Verdict and concentration notes updated. Now the IPO-window, financing-web, and permitting notes.

Now the permitting-constraint note update (Google Finland halt), plus a new watch for the Nebius junk-bond pricing — the next clean test of junk-market appetite for AI paper.

All five notes are written and the Nebius watch is registered. Checking the watch ledger one more time: no one_shot watch fired this cycle (SpaceX $40B still in talks — close expected 2027; Volta bids due Oct 14, not yet priced; Anthropic S-1 still not public; 10-yr at 5.28% below the 5.5% tripwire), and no ongoing threshold was crossed (TSMC +50% vs the <20% trigger; SpaceX CDS last print 194bps vs the 250bps trigger). So no `record_event` calls are warranted. Ending the cycle.

Cycle complete. Here's what this round found on **"Is Nvidia's AI capex story still intact?"**: **Verdict: intact for a tenth straight read — but the character of the risk changed.** Five notes updated: 1. **The first demand-leg scare landed and was absorbed** (`nvidia-ai-capex-cycle-sustainability`). On Oct 8, the FT reported OpenAI's annualized revenue was ~$50B vs the ~$68B previously circulated — knocking NVDA -2.9%, Oracle -5.9%, CoreWeave -7.8%, and XLK to its worst day since Sept 14. But it was a reporting-methodology dispute (the $68B included gross partner revenue), with OpenAI still growing 77% run-rate in Q3 — and TSMC printed record Q3 revenue of NT$1.49T, **+50% YoY, the same day**, beating its own guidance. 2. **The concentration thesis got its live test** (`ai-customer-concentration-openai-anthropic`). One revenue clarification at one private customer repriced the entire AI complex in a session — Eisman's "single point of failure" mechanism demonstrated, then resolved benignly. Meanwhile both top customers face public skepticism at once (New Constructs values Anthropic at $150B vs its $2T IPO target). 3. **The equity exit channel failed its biggest test** (`ai-ipo-window-equity-exit-channel`). Nvidia-backed Firmus Grid pulled its US$5.5B Australian IPO after investors refused the A$11 price — "a compelling story... just not a compelling valuation" (UniSuper). Public-equity checks are now failing *before* private-credit checks. 4. **The financing web keeps building — but the secondary market is marking down** (`nvidia-infrastructure-leasing-circular-financing`). SpaceX's $40B Apollo-led raise confirmed in talks (close 2027); Volta's $5B GPU loan **oversubscribed** at ~11% (bids Oct 14); Brookfield/Naver ~$9B Korean loan. But Tract's Nvidia-leased bonds are down 7-11 points and Coatue/Fluidstack paper went 8.875%→11%+ — all on **permitting risk**, the new marginal driver. 5. **The permitting constraint went global** (`hyperscaler-capex-surge-2026`): Google ordered to halt two Finnish data centers, joining NM, TX, NV, and OK in the constraint stack. **Watch registered:** Nebius's $3-4B junk bond (Goldman bookrunning) — the next clean test of junk-market appetite. Key dates ahead: Volta pricing Oct 14, Anthropic S-1/pricing late Oct–Oct 31, Microsoft Oct 27, Nvidia earnings Nov 18.