S&P +0.5% / SOX +1.4%. 10y ~5.18%; Brent ~$98. AKAM held green, ORCL soft. DoD Anthropic bar upheld; China DC model.

2-1 D.C. Circuit keeps one DoD supply-chain risk designation on Anthropic; conflicts with N.D. Cal. vacatur of a parallel label.

New China Datacenter Model maps 1,000+ sites / 60+ operators; China >24GW vs US 56GW; ByteDance ~1/5 capacity, mostly leased.

ES +~0.2%; 10y ~5.18%; Brent ~$98 on Hormuz hopes. Akamai–Anthropic $11.6B; WSJ Jupiter “cracks.” Trump–Xi optics, no chip deal.

Bloomberg via CNBC/TechCrunch: Oracle sent Blue Owl force majeure on Stargate Project Jupiter; seeks payment delay if 2028 slip; ORCL −4%.

Akamai 8-K/PR: Anthropic commits ~$11.6B over seven years for dedicated cloud/CPU capacity; optional +$9B; ~$5.5B related capex.

S&P ~flat / SOX −0.3%. 10y ~5.16%; Brent ~$107. Oracle Jupiter FM (−3.5%); Texas AG water probe. SoftBank print digested.

SoftBank PR: ~$11.1B notes priced — USD 8.625%/9.25%/9.75%; EUR 7.125%/8%; issue Sep 29; funds $10B OpenAI tranche (Oct 1).

Paxton OAG investigating data-center developments that did not answer TWDB water-use surveys; initial counties include Harris, Travis, Dallas.

SoftBank issued ~$11.1B OpenAI HY. S&P futures ~−0.5%; 10y ~5.13%; Brent still >$100 into Trump–Xi dinner.

Bloomberg via wires: BlackRock-backed AIP and IFM in exclusive talks for Stack Infrastructure Asia-Pacific DCs, up to ~$25B; Blue Owl seller.

The Information via Reuters: DeepSeek ARR ~$1B (more than double), price hikes 2.3–4.5×; Liang says >70% compute to training.

GS Chief Credit Strategist: hyperscaler IG $250bn/'26→$400bn/'27; $6tn capex '26–'30; hyperscalers 40% of AI issuance; little crowding-out; IG absorb.

Ep.033: 77 providers ranked; Nebius joins CoreWeave platinum; Google gold; Azure/AWS down; GPU-hour backwardation; NVDA backstop ~$588bn→$2tn; SLAs.

S&P −0.75% / NQ −1.13% / SOX −1.2%. 10y ~5.11% (2007 high). Oil rebound; Barr more hikes. Ramp 5.73%; Claude ART enzyme.

Ara Kharazian adds AI share of non-payroll business spend: 5.73% in August, almost entirely tech + technical services on Ramp’s panel.

Anthropic primary: Claude agents spotted array-associated reverse transcriptases in jumbo phage DNA; lab verified; function still open.

S&P soft (−0.1%) after flat Tuesday. SoftBank HY >$20B interest. JPM 16Hi→2029 relay; Samsung D0a hybrid bonding; Trump–Xi today.

jukan05: JPM HBM model now sees 16Hi adoption 2029 earliest. Replies split on earlier slip vs prolonged 8/12Hi scarcity; no JPM primary in-thread.

ZDNet Korea: D0a label shifts from VCT to W2W hybrid bonding (B1b); VCT becomes D0a-V; B1b under review for HBM5 (~2029).

InstructGPT coauthor: frontier chat/RLHF APIs wrong for software; Jev as code-consumed System One; >1T tokens/day machine traffic; dark data + agents.

S&P flat / NQ +0.45% / SOX +2.1%. Brent $99.25; 10y ~4.97%. Barkin keeps hike door open. Opus ~40% cheaper; Sol/Luna −50%.

Same-day primaries: Anthropic Opus 5.5 ~40% cheaper vs Opus 5; OpenAI Sol/Luna API list −50% vs GPT-5.6 promo.

ES ~+0.1% after Nasdaq record; Brent soft near $100, 10y ~4.9%. SoftBank ~$11B OpenAI HY, Abbott TCEQ halt, EMIB/Penang, Alibaba Zhenwu.

SemiAnalysis ChipBook: Korea→Malaysia HBM +467% YoY, ~$3B/2mo via Penang. Replies split: customers fleeing CoWoS scarcity vs queue transfer / qualify risk.

Apsara: T-Head Zhenwu V900 (3× M890, commercial Q1 2027); Cloud >20GW by 2032; Qwen path to 5–10T params. CEO Wu says demand outpaces supply.

Sep 21 letter directs TCEQ to pause all data-center permits until ERCOT audit finishes; compliance update due Oct 19. Extends August grid freeze.

Rogo CEO: o1 Pro→Opus 4.5 unlocked junior-analyst work; next 2–5y is firm reinvention; harness/compliance/last-mile beat raw models for buy-side.

Unconventional AI CEO: Google-scale token energy already ~12GW; ~50% of token cost is power; aims 1000x efficiency in ~3.5y via dynamical chips.

Rayliant CIO: China models on-par/open-source; energy grid edge; hardware rents until overcapacity; Mag7 CapEx arms race; S&P is one-tree, not diversifier.

Ex-IMC semis options MM on flow/V, NVDA–AMD relative vol, DeepSeek corr blowups; 50-name AI stack book at ~2×; models-beat-S&P claim needs harness.

Light Street CIO on AI5 semis concentration, NVDA ~85% share, demand ahead of supply, 10–20y stack cycle, agents→~5× tokens; DC politics as education risk.

S&P ~flat Fed-hike week; 10y ~5%, WTI $100. Overnight: OpenAI −$278B FCF path, Anthropic >$100B ARR / Nov IPO talk, Nscale S-1, Accenture eval.

NYT via Bloomberg: Anthropic annualized revenue expected above $100B this year; separate WSJ/Reuters lanes point to a possible November IPO window.

Anthropic primary: Faculty/Accenture as first embedded evaluator under Amodei’s pacing plan; each side expects ≥$1B over 5y; Anthropic funds Accenture.

FT (via Reuters): OpenAI July presentation projects −$278B FCF 2026–2030, ~$856B compute/infra spend vs ~$840B cumulative revenue; cash out ~2028.

CNBC on Nscale prospectus: H1’26 $140.6M revenue / $1.02B net loss; $56.4B RPO; 25k active GPUs; NYSE ticker NSCL; Nvidia among investors.

Databricks CEO on pacing PR vs cyber risk, four-test RSI bar, ontology/Genie as the adoption bind, Uni Gateway cost control + GLM shift.

Stefano Ermon on Mercury ≈ Haiku/Flash/mini speed tier, ~10× decode vs AR at GPT-2 scale, OpenCall leaving Cerebras for NVDA GPUs, 20–30% latency wedge.

Witching Friday mixed-narrow: S&P +0.17% / NQ +0.39% / Dow −0.18%. 10y ~5.00%; WTI $100.30. CleanSpark Meta HY ~$2.28B 4× bid.

Bloomberg: CleanSpark ~$2.28B HY deal drew ~$10B orders; Sandersville GA site leased to Meta sub Anviran on $6.6B / 20y; ops aimed Q4 2027.

Futures muted into witching; 10y ~4.98%. Crusoe $3.9B, SoftBank Arm loan, labs hunt 20–30 MW DCs; Dwarkesh–Brown swarms.

@ParadisLabs cites MS: AMD MI450/MI455 stay 100% TSMC CoWoS in 2027; mid-point ~600k units; Venice may take ~80k CoWoS-L if OSAT ramps lag.

CNBC relays Reuters: exploratory talks for SK Hynix to make memory in the U.S. via Intel Ohio lease and/or a JV that could include major cloud firms.

Reuters: SK Hynix's Solidigm considering a U.S. NAND flash fab with upstate New York leading; separate from Hynix–Intel Ohio memory talks; no plan set.

Bloomberg: SoftBank raised its Arm share-backed margin loan by $5B to $25B; separately expanded a credit line to $6.5B as AI funding needs mount.

Reuters/company: Crusoe raised $3.9B Series F at $30.9B post-money; cites >$140B contracted value, >6 GW contracted / 1 GW live; Spark modular sites.

CNBC sources: both labs seeking smaller ~20–30 MW capacity deals (UK/Nordics/US) for speed-to-usable capacity as inference can run on split clusters.

OpenAI's Noam Brown on 10k-agent Navier-Stokes solve (130B tokens/88h), Ultra Mode multi-agent, Codex $7–8k/day internal, RSI ≠ 100x overnight.

Bounce stuck — S&P +1.13% / Nasdaq +1.62%. Huang ~2x chips; Nvidia–Brookfield $2B. Oil settles ~$102; 10y ~4.95%.

Nick Dorsey from AI Infra Summit: vendor roadmaps overlap, Astera/CXL example — today's technical leads feel wobbly under industry-wide acceleration.

SemiAnalysis says MI355X is quickly closing perf/TCO vs GB300 in agentic inference via SGLang/MoRI/UMBP; replies split on InferenceX $/M-token snapshots and software vs silicon.

First Squawk/Bloomberg relays: Nvidia commits $2 billion to a Brookfield AI fund, extending the prior BAIIF / Radiant infra partnership.

CNBC: Jensen Huang said Nvidia expects to sell twice as many chips next year as this year; comes after ~70% FY28 growth / ~$673B outlook.

Futures rebound after Fed selloff; WTI under $100; 10y eases under 5%. Generac–Amazon power; Gerstner offtake; OpenRouter share flip.

SemiAnalysis maps 300+ U.S. datacenter moratoriums: 20GW inside restricted local boundaries, 1,525MW actually slips, 2.3GW nationwide including New York.

Market rumor: CPO cuts ultralow-loss CCL demand. Jeff Pu says upgrade stays — CPO not on compute board soon; NVSwitch still needs high-spec PTFE CCL.

Baker cites OpenRouter: OpenAI 20%→50% vs Anthropic since June. Talia notes token vs revenue share diverging; Baker agrees OpenAI is discounting.

Rhodium via CNBC: all Chinese AI models combined ~10% of OpenAI+Anthropic ARR; cites DeepSeek $0.5B to ByteDance $4B vs OpenAI $40B / Anthropic $65B.

CNBC: McKinsey/SEMI Foundation see up to 157,000 U.S. semiconductor worker shortfall by 2030; Samsung Austin EVP flags pipeline; Asia hiring still leads.

Generac filing: long-term Amazon data-center generator supply with ~$2.4B initial deliveries in 2027–28; Amazon warrants vest toward $8B payments.

Rune Kvist (ex-Anthropic) on $40M Series A: AIUC-1 quarterly agent standard, Lloyd's-backed policies, Waymo/Air Canada liability — eval+insurance stack.

Altimeter's Brad Gerstner on All-In: earnings-driven tape, offtake must fund Mag5 capex, Dylan 43GW too hot (~25GW), lab RR as takeoff switch. [asr]

Unanimous 25bp hike; Warsh hawkish. S&P −0.45% / Dow −1.21% / SOX +0.63%. Oil off highs; Amazon–Generac DC power; Apollo CDS caution.

Futures modestly higher into Fed; 10y eased under 5%; oil still three-digit. Hynix–Intel memory talks; pacing reframed as more compute.

jukan05 says the load-bearing angle is a possible Hynix–Intel JV with hyperscalers hungry for memory; replies stress prior denials, cycle cost, and that talks ≠ wafers this quarter.

Gavin Baker argues Anthropic/OpenAI "pacing" is more compute on alignment, monitoring, and evals at the cost of slightly lower margins — not a capex cut. Roon frames pacing as asymmetric margin compression; xEBITDA argues safety spend can raise total semiconductor demand.

Emergency ep on Amodei pacing: OpenAI CoT monitoring ~20% of rollup compute; safety spend likely raises, not cuts, infra demand; HF as shot across bow. [asr]

Nadella on All-In: broad diffusion over mystical slowdown; ~30m enterprise Copilot users of ~250–300m TAM; kit ~60% of cost; Quincy DC ~400–500 MW. [asr]

S&P −0.45% as 10y tagged ~5.04%; SOX bounced while indexes fell. Ramp Astra 13% vs Fable 8%; AWS Bahrain unrestorable; Kyber slip to 2028.

CNBC: six months after Iran-war drone damage, AWS says it cannot restore Bahrain-hosted resources/data; one UAE availability zone also unrecoverable, with two other UAE zones still being worked.

Paradis relays a Goldman Sachs Nvidia 800VDC expert call: Kyber NVL144 systems for Rubin Ultra slipped to 2028 (>12 months past a 2027 target); 2030 US 800VDC share cut to 21% from 25%.

Ramp's Ara Kharazian says Astra takes 13% of enterprise AI spend vs Fable at 8%; he frames Anthropic's pace-the-frontier call as already losing adoption. Replies split on whether Ramp is the right sample.

Futures ~flat into Fed; 10y tagged ~5.02% and oil stayed three-digit. Tan reaffirms Broadcom AI path; Anthropic GW financeability still open.

SpaceX President Gwynne Shotwell says SpaceX is as much an AI business as a space business by revenue, with compute rental 'a heck of a business' and Starlink at ~1.5–2% penetration; Elon Musk joins from Memphis to push cross-lab model peer review, handicap Starship ship-catch at ~50–60%, frame Terafab as build-or-fail-to-scale, and non-deny a Tesla–SpaceX combination.

Huang on All-In: extinction %s unscientific; ~$400bn AI VC ~80% open-model; NVIDIA goes "as deep as needed." Trump brands DC opposition a hoax. [asr]

SemiAnalysis posts that GPT-6 Astra is "basically confirmed" to use loop transformers — deeper passes over layers without growing parameter count. Visible replies split on whether that is HBM-sparing or sequential-compute-heavy. Video body unread; root text truncates mid-quote.

Carl Quintanilla quotes Wells saying active data-center moratoriums rose 175% nationally over the last three months, with momentum expected to continue amid frontier-lab security concerns. Primary Wells note not opened.

jukan05 relays Korean media that Samsung is collaborating not only with Samsung Foundry but also with TSMC on custom HBM base dies depending on customer requirements. Primary article not opened; root post truncated.

jukan05 relays Korean media that Samsung Foundry's Texas Taylor fab started earlier than a planned November start, with utilization crossing ~20% last month and approaching ~30% this month. Primary outlet not opened this pass.

On CNBC Mad Money, Hock Tan said Anthropic pacing chatter has not changed Broadcom's $115bn FY27 / $230bn FY28 AI semi forecasts; demand "very durable."

jukan05 relays Korean media that Taylor utilization hit ~30% (from ~20% last month), pulled forward vs a November plan on Tesla AI5 demand, with some expecting full utilization by year-end. Aju Press same week frames pilot lines and mass production early next year — a utilization-vs-pilot split.

jukan05 and ZDNet Korea report Samsung will use both Samsung Foundry and TSMC for custom HBM base dies by customer request, with Memory Division owning design when TSMC fabs the die. Replies frame it as avoiding turnkey isolation; wafer-capex allocation to TSMC remains the open bind.

After Monday's chip de-rating, JP Insights weighs Hock Tan's reaffirmed ~$115B FY27 / ~$230B FY28 AI semiconductor outlook — and Anthropic as largest custom silicon customer in 2027 — against weekly end-of-cycle takes. Visible replies favor operator numbers; demand vs delivery remains open.

NVIDIA CEO Jensen Huang tells the All-In hosts that agentic workloads drove a ~10,000x compute step in two years, that a higher-capex Vera Rubin factory can still deliver the lowest token cost via ~10x throughput, and that Physical AI is already a near-$10bn NVIDIA line while open-weight agents redefine the desktop OS — with China licenses restarting and consensus growth paths rejected as undersized.

Atreides CIO Gavin Baker tells David George that AI fundamentals accelerated through July–August while related equities drew down; argues sub-one-year compute paybacks and thin heavy-user penetration make undersupply through 2028 the base case, with NVIDIA’s financeable stack and hybrid open-source routers as the durable structure.

Tanay Jaipuria relays Anthropic's Broadcom TPU roadmap — 1 GW Ironwood in 2026, 5 GW TPU v8i in 2027, line of sight to 10 more GW in 2028 — with Anthropic as Broadcom's largest XPU customer next year. Thin replies challenge cash generation and call the Google overtake "wild."

Eric Balchunas flags a 90% vs 38% juxtaposition on rate-hike odds into the Fed meeting. Replies split between trusting economist/market hike pricing after a hot inflation print and a personal 50/50 that Wednesday can still go either way.

Metadata-only: TBPN's Sep 14 episode (full + Diet cut) titled The AI Slowdown Debate, with guests including Nico Wittenborn, Scott Keogh, Mitchell Green, David Rosenthal, Ben Gilbert, and Faraj Aalaei.

OpenRouter's Peter Walker shows three major American labs attracting about 70% of spend while accounting for only 27% of tokens. Thin reply chain; the chart is the claim. Google's weight depends on whether the metric is spend or tokens.

Session closed SPX −0.5% / Nasdaq −0.6% with positive breadth; damage in SOXX, chips, and AI-power. Baker reframes pacing as more alignment compute and lower lab margins; HBM 4-Hi stays contested sell-side vs SemiAnalysis.

Metadata-only capture of SemiAnalysis Weekly Ep. 030: Myron Xie and Jordan Nanos on Rubin Ultra shipping 192GB HBM versus a 1TB preview, supply-driven decontenting, and why less memory per chip can still be the right call.

Metadata-only: Steve Eisman with Vincent Daniel and Porter Collins on bonds, Treasury buybacks, OpenAI risk, gold, shorting mechanics, and two live short ideas.

Metadata-only: Latent Space interviews Richard Socher (Recursive / You.com) on recursive self-improvement as the next major AI step; show notes flag AI×finance conference promo.

jukan05 cites TrendForce that 4-Hi HBM is not enough and doubts suppliers would make it anyway; Jeff Pu concurs that datacenter de-spec to 4-Hi is unlikely. SemiAnalysis the same day argues 4-Hi maximizes tokens per HBM wafer for inference — a live supply-vs-architecture split.