Premarket Sep 15: oil, 5% 10y, Fed

The lead

S&P futures ~flat — ES ~7,623.50 (−0.02%) vs Monday close 7,619.98 (−0.48%); NQ ~29,131.75 (−0.07%) (Markets Insider ~9:25am ET).

• 10y tagged ~5.02% Tuesday — highest since 2007 on the Bloomberg print — after Monday’s close near 4.96% / touch ~5.01–5.014%.
• Oil held three digits on Saudi East-West pipeline / Houthi risk: Reuters near the open had Brent ~$105.74 / WTI ~$101.66 after session highs near ~$108.43 / ~$104.21; Markets Insider ~9:25am showed Brent ~$106.41 / WTI ~$103.72.
• FedWatch ~92–93% for a Wednesday 25bp hike (CNBC / Reuters); Chair Warsh presser is the near event.
• Empire State printed 7.6 in September vs 20.6 in August; prices paid rose further (WSJ / Dow Jones Newswires).
• Operators pushed back on an AI cliff: Tan reaffirmed Broadcom’s $115bn FY27 / $230bn FY28 AI semi path on CNBC; Samsung Taylor / HBM wires and Wells DC moratoriums stayed live overnight.

Open question: Does the cash open finish Monday’s AI de-rating, or do oil + a ~5% 10y into Wednesday’s Fed dominate while semis digest operator guidance vs essay language?

Positioning

AI capex durability — operators still defending the multi-year path. Tan’s “not in the least” on FY27/FY28 AI semi forecasts (CNBC; discussion) treats pacing as governance talk, not a sequential guide-down. Shotwell sees no drop in compute-rental demand; Musk frames Terafab as build-or-fail-to-scale (All-In). Huang rejects slowdown (All-In); Baker keeps undersupply through 2028 as base case (a16z). Wells moratoriums +175% are delay risk on megawatts — not a hyperscaler guide-down. Oil + ~5% 10y are competing discount-rate pressure into Wednesday, not a theme falsifier by themselves.

HBM supply binds — contested on new axes, not cleared. Samsung dual-sourcing custom HBM base dies via TSMC (discussion) is customer-access / NVHBM plumbing — not a third stack supplier ending the constraint; replies flag TSMC advanced-logic wafer allocation as scarcity transfer. Astra loop transformers (discussion) open whether next-frontier depth-via-loops is HBM-sparing or sequential-FLOP-heavy — unsettled. Taylor early ramp on Tesla AI5 (discussion) is consistent with advanced-node scarcity, with Aju Press pilot caution as the check.

Inference margin inversion — still open via spend ≠ tokens. OpenRouter’s ~70% spend / ~27% tokens split (discussion) shows high ASP at frontier labs can coexist with either widening or compressing margins. Baker’s shortage frame allowing token prices to rise sits awkwardly against repeated-price-cut GM expansion. No falsifying lab margin print overnight.

Enterprise agent stall — lightly contested. March Jensen All-In (archive) describes production-flavored agent anecdotes without a Fortune 500 $ or headcount figure. Recent Jensen doomer episode stays at lab/tooling altitude.

Top stories

10y ~5.02% into Fed — oil still three-digit. Bloomberg print highest since 2007; Reuters oil on East-West/Houthi risk; Goldman scenarios still cite Brent toward $120+ if Gulf outages persist; FedWatch ~92–93% for a 25bp hike Wednesday (CNBC / Reuters); CNBC notes oil–10y correlation at strongest since 2019.

Empire State cools — prices paid still hot. NY Fed Empire State business conditions index 7.6 in September vs 20.6 in August (WSJ expected ~15); prices paid and prices received both rose five points — energy-inflation vs demand split for the session (WSJ / Dow Jones Newswires).

Broadcom’s Tan — AI semi targets unchanged amid pacing debate. On CNBC Mad Money, Hock Tan said Anthropic-led “pace the frontier” chatter has not changed Broadcom’s $115bn FY27 / $230bn FY28 AI semiconductor forecasts; Anthropic on track as largest custom customer in 2027–28 (CNBC; news).

Anthropic Broadcom TPU path — 1/5/10 GW, pay-for-it unanswered. Tanay Jaipuria: Ironwood 1 GW in 2026, TPU v8i 5 GW in 2027, line of sight to 10 more GW in 2028; Anthropic to top Google as Broadcom XPU customer next year; visible reply: how do they pay if not generating cash (discussion).

Samsung Taylor ~30% util — Tesla AI5 cited as pull-forward. Korean media via jukan05: utilization ~20% → ~30%, earlier than November plan; Tesla AI5 across auto/robotaxi/robot/DC; Aju Press same week: pilot lines / test wafers, mass production early next year — treat 30% as directional (news).

Samsung dual-sources custom HBM base dies with TSMC. ZDNet Korea / jukan: Samsung Foundry + TSMC “two-track” for cHBM base dies by customer request; Memory Division designs when TSMC fabs (news).

Wells — active DC moratoriums +175% over three months. Carl Quintanilla quotes Wells: national active DC moratoriums up 175% over three months, momentum expected to continue amid frontier-lab security concerns; primary Wells note not opened (news).

The conversation

Anthropic 1/5/10 GW path — financeability still open. Tanay’s schedule matches Broadcom call language; thin replies challenge cash. Carried read: roadmap load-bearing if supply holds; pay-for-it unresolved (discussion).

Fed hike odds — ~90% priced vs 50/50 room. Balchunas 90% vs 38% juxtaposition; replies include 50/50 personal odds. Carried read: priced hike has the better of the pre-meeting tape; Wednesday settles the binary (discussion).

Broadcom FY27/FY28 AI targets — rebuttal to “cycle end” tape. jpinsights weights Hock Tan’s $115B/$230B path and Anthropic as largest custom silicon customer in 2027 over weekly end-of-cycle takes. Carried read: operator numbers win the mechanism fight overnight; delivery and CSP concentration remain open (discussion).

Samsung Taylor early ramp — real foundry recovery or pilot optics? ~30% utilization / Tesla AI5 pull-forward vs Aju Press pilot framing. Carried read: direction early is credible; do not equate utilization prints with proven 2nm mass economics (discussion).

Samsung–TSMC HBM base-die dual-source — bind eased? ZDNet Korea two-track; replies flag TSMC wafer allocation. Carried read: customer-chain access, not HBM unconstrained (discussion).

GPT-6 Astra looped depth — HBM vs FLOP intensity? SemiAnalysis “basically confirmed” loop transformers; replies split HBM-sparing vs sequential-compute-heavy. Carried read: named research claim is load-bearing; semiconductor implication unsettled until write-up or model card (discussion).

OpenRouter spend ≠ tokens — three labs ~70% spend / ~27% tokens. Monday evening chart; thin replies. Carried read: spend ≠ tokens — frontier APIs as high-ASP slice (discussion).

In their words

Hock Tan just reaffirmed Broadcom’s ~$115B AI semiconductor revenue target for FY27 and ~$230B for FY28.
— JP Insights (@jpinsights), X, 2026-09-15 03:54 ET — Operator multi-year path vs Monday’s cycle-end tape.

According to Korean media reports, Samsung Foundry's Taylor fab… has recently reached a utilization rate of 30%.
— Jukan (@jukan05), X, 2026-09-15 05:00 ET — Falsifiable utilization claim on the US foundry ramp.

Samsung is collaborating not only with Samsung Foundry but also with TSMC on custom HBM base dies, depending on customer requirements.
— Jukan (@jukan05), X, 2026-09-15 04:02 ET — Base-die dual-source; ZDNet Korea is the wire behind it.

The 3 labs still attract ~70% of all spend while only seeing 27% of tokens used
— Peter Walker (@PeterJ_Walker), X, 2026-09-14 17:12 ET — Spend ≠ tokens on OpenRouter.

It's as much an AI business as it is a space business.
— Gwynne Shotwell, All-In, 2026-09-14 — Mix claim behind the compute-rental durability read.

Slowing down is definitely the wrong strategy.
— Jensen Huang, All-In, 2026-09-14 — Explicit anti-cliff posture after Monday’s pacing tape.

Podcasts & interviews

All-In — Jensen (Doomer Hoax) — BULLISH / asr. Rejects extinction-probability framing; cites ~$400bn AI-native VC with ~80% on open models; NVIDIA posture “go up as far as needed, as low as possible.” Recap · Watch

All-In — Elon & Shotwell (Terafab / Starship) — BULLISH / asr. Shotwell: SpaceX as much AI as space by revenue, compute rental with no demand drop; Musk: Terafab as build-or-fail-to-scale, Starship catch ~50–60%. Recap · Watch

All-In — Jensen (Physical AI / Inference, Mar 2026) — BULLISH / asr. Archive in-window: agentic ~10,000x compute step, Vera Rubin factory $/token argument, Physical AI near-$10bn line. Recap · Watch

a16z Show — Gavin Baker (AI demand outrunning compute) — BULLISH / asr. Undersupply through 2028, sub-one-year compute paybacks, hybrid open-weight routers; token prices allowed to rise under shortage. Recap · Watch

TBPN — The AI Slowdown Debate — NEUTRAL / metadata-only. Live + Diet cut; guests include Acquired’s Rosenthal/Gilbert. No transcript ingested. Recap

What to watch

Fed decision Wednesday 2:00pm ET — path language over the print. Statement, SEP/dots, Chair Warsh presser at 2:30; markets ~92–93% priced for 25bp (CNBC / Reuters); path language matters more than the print against ~5% 10y and three-digit crude (discussion).

Cash open breadth and SOXX path — concentration vs mean-revert. Whether Monday’s concentration (chips/power) continues or mean-reverts while Mag7 dispersion holds; Empire State prices-paid heat already in.

Broadcom / Anthropic confirmation path — targets vs pay-for-it. ~$115B FY27 AI semi revenue, Anthropic as largest XPU customer, and any funding/revenue print that answers the pay-for-it objection (Tan news, financeability).

Samsung Taylor follow-ups — utilization denominator and AI5 quals. Samsung IR or Korean follow-ups on Taylor utilization denominator and Tesla AI5 qualification (discussion).

Samsung/TSMC cHBM confirmation — base-die dual-source in volume. Any Samsung/TSMC confirmation of cHBM / NVHBM base-die dual sourcing in volume (discussion).

Astra / SemiAnalysis write-up — looped vs dense HBM/GB. Measured tokens/$ or HBM/GB under looped vs dense baselines (discussion).

DC permitting / moratorium follow-through — delay vs cancelled spend. Whether Wells +175% prints into hyperscaler site delays vs cancelled spend (news).