〔Day Digest〕AI Capex Pivots From Chips to Rack-Scale Systems as SEMI Lifts Americas Wafer Spend Past $40.6B by 2029 — While JPMorgan Locks Out External AI Agents and Lao Gan Ma Trims Pickers 600→48

0. Daily Arc

The October 14 AI tape reorients around infrastructure, not silicon. SEMI projects Americas 300mm wafer-fab equipment spend more than doubling from $18.1B in 2026 to $40.6B in 2029, while global spend rises from $113B (2025) to $227B (2029) ^[1]. Bank of America's post-OCP note frames the pivot: rack-scale systems — power, cooling and network — have displaced single-chip throughput as the binding constraint, with Nvidia now optimizing for tokens-per-watt and AMD disclosing 95x monthly token-consumption growth ^[2]. On the same day, concrete adoption landed: Lao Gan Ma trimmed chili sorters from 600 to 48 on AI vision inspection and reported a record 5.4 billion yuan in 2025 revenue ^[3], while JPMorgan banned external AI agents from client accounts, with CEO Jamie Dimon calling the risk "too high" ^[4].

1. Capex Cycle and Equipment Pipeline

• **[NEW] SEMI forecast:** Americas 300mm fab equipment spend rises from $18.1B in 2026 to $40.6B in 2029, led by Intel, TSMC, Samsung and Micron ^[1]. Global 300mm construction spend goes from ~$48B (2026) to $63B (2027) to $68B (2028); 20 new 300mm lines start in 2026 and 18 more in 2027, with the U.S. taking 3 and 4 of those, respectively ^[1]. Figures were released at SEMICON West 2026 ^[1].
• **[NEW] ASML CEO Fouquet (Q3 video interview):** demand will stay supply-constrained through at least 2028; the 2027 NXE:3800F EUV tool lifts throughput from 230 to 260 wafers per hour, with material volume from 2028; immersion DUV tools shipping in 2027 are targeted at 330 wafers per hour ^[5].
• **[ESCALATED] System-level competition:** OCP summit discussion prepared technology roadmaps for 1MW racks, GW-scale data centers and multi-GW campuses, with Broadcom CEO Hock Tan flagging capital — not technology — as the binding constraint on AI expansion ^[2].

2. Industrial Adoption and Use Cases

• **[NEW] Lao Gan Ma:** AI visual inspection scans each chili every 0.3 seconds and air-ejects defective ones; key-process defect rate fell to 0.02%, chili-sorting headcount dropped from 600 to 48, and 2025 revenue hit 5.4 billion yuan — a record ^[3]. Deployment spans raw-material sorting, production QC and supply-chain scheduling at the Changming base in Guiyang ^[3].
• **[NEW] Pony.ai:** seventh-generation Robotaxi began road testing in Manchester with Uber and Verne, less than a week after the London announcement ^[6].
• **[NEW] Ant Lingbo:** signed strategic cooperation deals with Longsheng Weirui, Dema Technology, and De'er Xiangwang (robotics arm of De'er Co.) to push embodied robots into industrial logistics, anchored on a self-developed embodied-native model and cloud toolchain ^[7].
• **[NEW] Yuncong Technology:** launched an AI marketing team for cross-border e-commerce, combining AI infrastructure, agent operating system and growth operations ^[8].
• **[NEW] Doubao Work:** added a data import and sync function that pulls tasks, skills, connectors, memory and chat history from ChatGPT, Claude Code and Trae Work, preserving context ^[9].
• **[NEW] iFlytek:** the Global 1024 Developer Festival opens October 22 at the Beijing National Convention Center Phase II, with the launch of Starfire X2.6 and a new "Cloud+Edge" framework; parallel science expos run in Beijing and Hefei across education, healthcare, office and automotive verticals ^[10].

3. Policy, Regulation and Risk Control

• **[NEW] JPMorgan:** will block Meta's Muse, Instinct and other external AI agents from customer accounts; CEO Jamie Dimon said on the Q3 earnings call: "we will not let any external agent program into the banking system — the risk is too high" ^[4].
• **[NEW] Sunil Mittal (Chairman, Bharti Enterprises):** at the Bloomberg New Economy Forum in New Delhi, urged governments to set up AI regulatory frameworks immediately, drawing an aviation analogy: "it would be very foolish to think we can let AI develop to some stage and only consider how to respond after an accident" ^[11].
• **[NEW] Anthropic–Singapore:** signed an MOU to cooperate on AI safety, security, model evaluation and research ^[12].
• **[NEW] Shandong provincial government:** opinion to develop the data labeling industry, prioritizing smart labeling and 4D labeling for finance and healthcare verticals ^[13].
• **[NEW] Jiangsu:** information and communications "15th Five-Year" plan targets a fully built-out next-generation network and computing network by 2030 ^[14].
• **[NEW] Guangzhou Computing Industry Development Co.:** registered with 900 million yuan registered capital; legal representative Xiao Zhanhui ^[15].
• **[NEW] Qingdao:** released AI scenario innovation catalogue v2.0 ^[16].
• **[NEW] Xinhua:** coverage frames China's national integrated computing network evolving from compute clusters to "token factories" ^[17].

4. Macro Frame and Corporate Moves

• **[NEW] Li Xunbo (Yicai interview):** framed AI as a fourth technological revolution on a ~50-year Kondratiev cycle; cited 14–26% corporate efficiency gains, a 1,000x+ rise in China's daily token calls over two years, and double-digit export growth in January–August driven by AI-related equipment demand — though profits are concentrating in a handful of chip and equipment vendors ^[18].
• **[NEW] IMF Managing Director (Singapore event):** "Singapore is in a leading position in AI," and "in the current high-yield environment, financial conditions remain relatively loose" ^[19].
• **[NEW] Nvidia / Jensen Huang:** announced new investments in the French tech sector (France 24) ^[20]. Headline-only — treat as event marker, not a deal ledger.
• **[NEW] LSEG–Nvidia:** partnership to use Nvidia technology to process and analyze massive historical market datasets for financial institutions ^[21].

5. Contrarian / Tail Items and Source Quality

• **[NEW] Wuhan Tianyuan:** plans a 500 million yuan project in Chuxiong, Yunnan — 200 tons/year of 6N/7N high-purity indium and 100 tons/year of indium phosphide polycrystals; 100 million yuan equity, 400 million yuan bank loans, 12-month construction; the company itself flags "no technology or personnel reserve" and short-term immateriality to results ^[22]. Read as a forward supply bet on the AI/optical-compute chain, not a near-term earnings event ^[22].
• **[NEW] Thin-sourcing flags:** items ^[20], ^[23], ^[24], ^[25], ^[17], ^[16] are headline-only Google News or Xinhua relays with no body text — quote the headline, do not invent detail. The "OpenAI secret recording device" story appears twice (^[24] Economist and ^[25] Mashable) with different framings; treat as one event pending primary reporting. The EU "agentic AI" framing is an OpenAI-authored piece — single-source ^[23]. Quantify the capex story off ^[1] and ^[5], not off the secondary Google News headlines.