〔Day Digest〕Samsung Taylor Books Out Pre-Trial on 2nm as Citi Lifts 2030 Server CPU TAM to $237B From $137B; Goldman Says Two AI Catalysts, Not CPI, Decide the Week

0. Weekly Arc

The center of gravity rotated from inflation to compute: Citi lifted 2030 server CPU TAM to $237B from $137B ^[1], Samsung's Taylor fab booked out pre-trial on 2nm orders from Tesla and Broadcom with yields near 80% ^[2], and Arm's Rene Haas called chip scarcity the industry's #1 bottleneck inside an "absolutely supply-constrained" cycle ^[3]. Goldman strategist Lee Coppersmith argued the real weekly driver isn't Tuesday's CPI but two AI catalysts — OpenAI's Astra release and the Communacopia+ conference — against a near-1% S&P straddle and net leverage at the 4th percentile ^[4]. China is layering its own stack in parallel: Arm's first Chinese CPU customers named, a Supreme Court AI IP framework, and a Token-compute loan in Sichuan ^[5][6][7]. Net: a supply-constrained compute cycle colliding with a tripling of the demand forecast and a positioning tape that has paid for almost no volatility.

1. Compute Supply: Fabs, Packaging, and the Bottleneck

• **[NEW] Samsung Taylor sold out pre-trial:** 2nm bookings from Tesla, Broadcom and others fill the Texas plant before late-September/early-October trial production; yields rose from ~60% early in 2026 to ~80% recently, with mass production targeted for next year ^[2].
• **[NEW] Arm CEO Rene Haas (paraphrased):** AI demand for chips and memory is "unprecedented" for the industry; a fab costs "tens of billions of dollars" and takes 2-3 years; the industry is "absolutely supply-constrained" and "this will continue for some time" ^[3].
• **[NEW] Samsung Yokohama packaging R&D:** Samsung opened an advanced packaging research center in Yokohama under a roughly ¥40 billion five-year plan first announced in 2024 ^[8].
• **[NEW] Arm integrates a neural accelerator into Mali GPU for the first time**, debuting Mali G2-Ultra NX inside CSS for Mobile 2; Arm is following Qualcomm's Sept. 2 Adreno Neural Fusion reveal, signaling mobile GPUs are absorbing on-die AI compute ^[9].

2. Demand Repricing: TAM, Data Centers, and the China Stack

• **[ESCALATED] Citi server CPU TAM:** 2030 forecast raised to $237B from $137B — a 73% upgrade within the same model ^[1].
• **[NEW] China data center capacity:** Operational IT capacity is set to more than double from 2025 levels by 2030, with new builds tilting to the north and northwest where land and power are more available; latency-sensitive mature markets grow more slowly ^[10].
• **[NEW] Arm's first Chinese CPU customers:** Lenovo, ByteDance's Volcano Engine, and New Tsinghua Unigroup; Volcano Engine plans to ship the first Arm AGI CPU-based agent sandbox solutions, and a joint AGI innovation center will cover agent orchestration, CXL memory pooling, rack-scale architecture, and edge AI ^[5].
• **[NEW] DeepSeek V4.1 Flash enters mid-version internal testing:** new model architecture, native multimodal support, faster speed, lower cost ^[11].
• **[NEW] Tencent Docs launches AI Workbench:** WorkBuddy Agent core combined with Tencent's editor engine and Skills, enabling AI to modify live files rather than only generate content ^[12].
• **[NEW] Baidu Xiaodu:** Super Xiaodu covers 55M families, with daily deep voice interactions up nearly 300% YoY, per CEO Li Ying ^[13].
• **[NEW] Insilico Medicine's Rentosertib:** first AI-designed anti-aging candidate to complete clinical evaluation; published in Nature Biotechnology with Harvard Medical School, Stanford, and the Broad Institute; claims up to 6-year biological-age reversal, with six independent aging clocks pointing the same direction ^[14].
• **[NEW] Songyan Dynamics launches embodied-AI sub-brand Scalabot,** focused on robot world-understanding and interaction ^[15].
• **[NEW] Winning Health joins Nvidia's Software Co-Pioneer China program:** collaboration on compute support, model training, and engineering deployment, though specifics are NDA-bound ^[16].
• **[NEW] Taicheng Light:** FAU (Fanout Array Unit) products developed for high-core-count, high-spec customer needs are now in batch shipments, per the company's interactive-platform response ^[17].
• **[NEW] Sichuan's first "Token compute scenario" loan** disbursed by ICBC's Chengdu branch to a Chengdu High-Tech Zone AI firm, collateralized on compute workloads rather than real estate ^[7].
• **[NEW] Supreme People's Court AI IP framework:** the Opinion and accompanying interpretive guidance were published on Sept. 7; they clarify AI-generated-content copyright status, free open-source AI provider liability safe harbor, and AI-invention patentability, while leaving contested questions open for further case-law development ^[6][18].
• **[NEW] MOFA briefing (spokesperson Mao Ning, Sept. 8):** China is hosting a new AI capacity-building workshop at Peking University, Sept. 7-12, advancing the AI Capacity-Building Inclusive Plan and the proposed World AI Cooperation Organization ^[19].

3. The Catalyst Stack: Goldman Names the Real Driver

• **[NEW] Goldman Lee Coppersmith:** two AI catalysts — OpenAI's Astra launch and the Communacopia+ tech conference (OpenAI on stage Tuesday 9:30 a.m. PT) — are the week's decisive variables, not Tuesday's CPI ^[4].
• **[NEW] Positioning data (Goldman prime book):** U.S. fundamental long/short net leverage at the 4th percentile, gross at the 27th percentile; VXSMH (semiconductor implied vol) collapsed from ~65 in July to 36; Goldman's tech long/short momentum basket still sits ~50% below its late-June high ^[4].
• **[NEW] S&P 500 weekly straddle priced at ~1%** — the market is paying for almost no volatility into a catalyst-dense week ^[4]. The contradiction: positioning is unusually quiet, but the catalyst stack is unusually dense.

4. Tail Risks and Source Quality

• The "coiled spring" framing is fragile: a soft Astra reveal or a hot CPI print could flip the setup into a gamma squeeze the wrong way ^[4]. The $18.6B headline tied to Nvidia, Palantir, and Meta (single source, Yahoo Finance) is unverified beyond the headline ^[20].
• China-specific catalysts carry execution risk: DeepSeek V4.1 Flash is mid-version internal testing only ^[11]; the Supreme Court AI IP framework leaves contested questions explicitly open ^[6]; and Sichuan's "Token compute scenario" loan is the first of its kind, so precedent is thin ^[7].
• Math-AI (Axios) and AI-Pentagon contracts (The Intercept) are headlines only, with no underlying figures cited ^[21][22]. The UK investment-trust exposure to Anthropic and OpenAI is a Morningstar piece without a quantified basket ^[23].
• The institutional flow read on high-end PCB, optical modules, and storage — with HDI in mass production at 9- to 11-stackup levels, claimed industry-leading — is a single-source social relay ^[24]; treat as a positioning tell, not a forecast.
• Cognizant's U.S. AI hiring plan (1,500 new graduates in 2026, scaling Frontier-certified engineers and ops specialists to 15,000, with the Synapse retraining target raised to 2M by 2030) is a single corporate announcement, with no third-party validation ^[25].