Weekend Sep 26: S&P +1.2%; GS ROIC; China tokens
S&P +1.2% week / NQ +2.1%; 10y ~5.18%; Brent ~$98. Overnight: GS $1.42T ROIC math, China 57–67% OpenRouter tokens, ORCL CDS 237bps.
The lead
S&P +1.2% on the week — closed Friday 7,743.41 (+0.51% session) vs prior Friday 7,650.50; Nasdaq +2.1% week / +0.48% Fri to 27,068.72; Dow +0.3% week / +0.93% Fri to 51,828.62; Russell 2000 −0.8% week (AP via WTOP; Yahoo). First winning week in three; tech carried, small caps did not.
- Rates still the long-end problem. 10-year finished ~5.18% (Yahoo settle 5.184) after poking ~5.23 — a ~19-year closing-high zone into SoftBank settle Sep 29 (Yahoo TNX; WSJ).
- Oil — Brent held off three-digit on Hormuz-reopen hopes. WTI ~$92.8 / Brent ~$97.7 Friday; AP: Brent dropped below $98 and helped yields ease off the highs (WTOP; Yahoo CL; Yahoo BZ).
- Breadth — NQ led; R2K red; Friday Mag7 split. Nasdaq’s +2.1% week vs Dow +0.3% / Russell −0.8% kept concentration live (AP); Friday cash was MSFT/AAPL-led with SOX +1.4% and META/ORCL soft (Yahoo SOX; Friday settle).
- Quiet overnight archive (post-close → Saturday AM): Goldman via ZH — ~$1.42T / $11.6B per GW clearable for 15% ROIC (59% in cloud backlogs) vs $920B/yr opex floor at 0 ROIC; CNBC — Chinese models 57–67% of OpenRouter tokens week of Sep 14; Oracle Jupiter CHANGE — Bernstein reads FM as not structural, while Sep 25 ORCL CDS hit a record 237bps; Latent Space OpenRouter / Stripe longform shipped.
Open question: Into SoftBank settle Sep 29 / OpenAI ~Oct 1, does a sticky ~5.18% 10y plus record ORCL CDS still force Stargate/DC credit pressure — or does the oil fade + GS clearability math + SOX reopen stick once aftermarket HY spreads print?
Positioning
AI capex durability — clearability math printed; CDS and rates keep the break path live. Overnight Goldman-via-ZH puts ~$1.42T cumulative rev / ~$11.6B per GW as clearable for 15% ROIC with 59% already in cloud backlogs — the durable-spend case — while the same note’s $920B/yr floor at 0 ROIC and jukan’s rates-as-obstacle camp keep the falsifier (token deflation into open models, or credit rationing on off-balance-sheet shells) visible. Oracle Jupiter update: Bernstein frames the FM as optionality on a Q4’28 OpenAI-primary campus (pipeline slip Aug’26→Feb’27), not demand destruction — but Sep 25 ORCL CDS +10 to a record 237bps undercuts Barclays’ prior “neutral credit” read. Break remains a top-four sequential guide-down or RPO cancel, not the notice alone.
Inference margin inversion — token mix migrating to cheaper Chinese open models on the router. CNBC citing OpenRouter/Vercel: Chinese models 57–67% of OpenRouter tokens week of Sep 14 (from 6–13% in February) and 55% of Vercel usage in August (from 11% in January) — while U.S. frontier models still attract more overall spend. Same distribution layer in Latent Space’s OpenRouter / Stripe treatment (10T+ tokens/day claim; marketplace price competition after invoice shocks). Consistent with cost-per-token pressure; a break is still disclosed lab GM compression across consecutive price cuts, not router mix alone.
Top stories
S&P +1.2% week / Nasdaq +2.1% / Dow +0.3% / R2K −0.8% — first winning week in three; rates and oil set the shape. Tech carried; small caps lagged. Friday oil fade and softer yields helped cash green into the close (AP via WTOP; Yahoo history).
10y ~5.18% / Brent ~$98 — long end sticky; Hormuz hopes held the crude unwind. Yahoo TNX settle 5.184 after a ~5.23 poke; Brent held under three-digit on reopen-hope headlines after Thursday’s ~$107 print (Yahoo TNX; Yahoo BZ; WSJ).
Goldman — $1.42T / $11.6B per GW clearable for 15% ROIC; $920B/yr even at 0 ROIC. Overnight ZH TL/DR of a GS sensitivity table; Klaudnin3 maps rent-the-GW (AAPL) vs shovel owners; jukan flags elevated rates as a major AI-investment obstacle (discussion).
Chinese models — 57–67% of OpenRouter tokens (week of Sep 14); 55% of Vercel (August). Up from high-single / low-teens earlier in 2026; spend still U.S.-frontier-led; House committees investigating (CNBC; news).
Oracle Jupiter CHANGE — Bernstein: FM not structural; CDS printed a record 237bps. Sell-side via X: Q4’28 online, Bloom up to 2.8GW / 1.2GW contracted, minimal FY27 revenue impact already guided; credit market: Sep 25 ORCL CDS +10 to 237bps vs Barclays’ Sep 24 “neutral” (news).
The conversation
Goldman hyperscaler ROIC — clearability has the better of pure “capex doomed,” but sensitivity + rates keep breaks live. ZH overnight math ($1.42T / $11.6B/GW / 59% backlog) vs $920B/yr at 0 ROIC; Klaudnin3 rents vs owns; Stalwart asks why META pays more interest to stay asset-light; jukan rates bind. Carried read: clearability-at-15% is the durable-spend case; token deflation or off-BS credit rationing is the break (discussion).
Podcasts & interviews
Latent Space — OpenRouter / Stripe (Atallah + Midha) — BULLISH / transcript. Neutral multi-model router as critical infra; claimed 10T+ tokens/day, 10M+ devs, ~9% WoW; Stripe deal sold as token-fraud / agentic-fraud security (blocked dollar-volume ~10× MoM); Mistral price war as first marketplace proof; focus over fine-tune/memory adjacencies. Recap · Watch
What to watch
SoftBank settle Sep 29 / OpenAI tranche ~Oct 1. Whether aftermarket spreads and CDS hold with 10y still ~5.18% and ORCL CDS at a record print.
Oracle–Blue Owl FM negotiation / Jupiter 2028 path. Whether FM is agreed (possible rent delay) or walked back; NM air-quality decision Nov 23; whether CDS cools after the 237bps print (news).
Goldman clearability vs next Mag7 / Oracle guide. Whether FY27–28 AI revenue or ROIC language matches the $1.42T / $11.6B-per-GW hurdle — or cuts the path (discussion).
OpenRouter / Vercel token-mix durability. Whether Chinese open-model share keeps climbing after U.S. lab cheaper-model responses, and whether spend share follows tokens (news).
Hormuz / Iran follow-through. Whether reopen-hope headlines stick or reverse — Brent already off Thursday’s three-digit spike (WSJ).
October FOMC vs sticky ~5.18% 10y. Whether FedWatch hangs elevated through SoftBank aftermarket with the long end still near multiyear highs.
The Open/Close · Research commentary, not investment advice. Positions may be held in securities mentioned.