〔Overnight Brief〕AI Safety Chorus Swells as TSMC Brands Models 'Three-Year-Old Superman,' Anthropic's Slowdown Call Meets Trump-Huang 'Completely Aligned' Line - No-Pause Coalition Intact

0. Weekly Arc

The overnight tape carried an open split: US AI giants publicly asking for a slower iteration on safety grounds ^[1], with TSMC co-COO Y.J. Mii saying AI in semiconductor R&D is a "three-year-old superman" - powerful but unsafe near sensitive data ^[2]. The acceleration camp fired back in unison: OpenAI's Sam Altman "very confident" on self-regulation ^[3], Nvidia's Jensen Huang "we don't need any new laws" ^[4], House Speaker Mike Johnson "we cannot pause AI development" ^[5], and Treasury Secretary Scott Bessent confirming Trump and Huang are "completely aligned" ^[6]. The disagreement is over tempo, not direction.

1. Policy Narrative

• **[ESCALATED] Safety camp:** US AI giants publicly call for slower AI iteration on safety grounds; the "computing-power metals" complex gave up its valuation premium as US Treasury yields rose ^[1]. Anthropic CEO Dario Amodei is publicly pushing for a slowdown until safety mechanisms are in place ^[7].
• **[NEW] White House contact:** Trump's aides met an Anthropic executive to discuss AI safety ^[8]. Treasury Secretary Scott Bessent said he has seen no accountability findings yet from the Hugging Face incident ^[9], and that Trump and Huang are "completely aligned" ^[6].
• **[ONGOING] Acceleration line:** OpenAI CEO Sam Altman, at Salesforce's Dreamforce in San Francisco with Marc Benioff, said he is "very confident" the industry can self-police ^[3]. Nvidia CEO Jensen Huang: "we don't need any new laws" ^[4]. US House Speaker Mike Johnson: "we cannot pause AI development" ^[5].
• **[NEW] (single-source) Senate pressure:** Senator Bernie Sanders is urging AI company CEOs to stop donating to anti-regulation political candidates ^[10].

2. Key Data and Market Read

• **[NEW] Cross-sector product drops:** Huawei launched what it calls the world's first 3D data center; Apple rolled out an AI version of Siri; Anthropic released a Claude variant for financial advisors; iFlytek launched AStudio desktop AI workstation and invested in humanoid-robot firm Mengwu Intelligent; Ant Group took a stake in embodied-AI firm Daimeng Robotics; Samsung invested $231M in Euclyd positioning for Nvidia GPU alternatives; Zidong Taichu open-sourced ZDTaichu5.0-9B ^[11][12].
• **[NEW] ByteDance H1:** Revenue rose 30% y/y to $120B; net profit fell to $20B as AI investment accelerated ^[13].
• **[NEW] OpenAI Canada:** First Canadian news deal signed ^[14]; OpenAI is exploring Canadian data-center partnerships citing energy and land advantages, per George Osborne, OpenAI's head of country affairs and former UK Chancellor of the Exchequer (2010-2016), who called the Carney government "one of the more explicit AI application plans among Western governments" ^[15].
• **[NEW] KPMG UK cuts:** ~200 additional roles to be cut in the UK consulting business, touching AI and adjacent teams ^[16].
• **[NEW] Frontier research:** Princeton Plasma Physics Laboratory and Princeton University published in Nuclear Fusion an AI framework that can predict and control fusion plasma in ~20 ms under hardware safety constraints ^[17]. US scientists using an AI model reported in Nature the discovery of a previously "invisible" human protein, TM184C ^[18].
• **[NEW] Mobile OS race:** Honor, vivo, OPPO are racing to ship system-level agents that can invoke multiple apps from a single sentence; ByteDance and ZTE Nubia's "Doubao Phone 2" opens sales September 16 with pre-orders over 370,000 units ^[11][19].
• **[NEW] Vertical AI agents:** Meituan launched a "craftsman Agent" for hairstylists that compresses 11 manual posting steps into 2; platform daily active stylists rose 16% y/y; the share of hairstylists posting monthly rose from 21.9% to 39.8% ^[20].
• **[NEW] China policy and deals:** The "15th Five-Year Plan" for intelligent-connected new-energy vehicles and the mandatory national standard "Safety Requirements for Autonomous Driving Systems of Intelligent Connected Vehicles" were recently issued; LLM and agent penetration spans cabin, vehicle control, and supply chain ^[21]. The Sichuan provincial government signed a strategic cooperation agreement with China Mobile focused on AI ^[22]; Shanxi, UBTech, and Hehao Ruisheng signed a cooperation agreement to build an embodied-intelligence innovation center and an industrial/mining special-robot project ^[23].
• **[NEW] Industry forum:** The 2026 Lujiazui Finance Salon fintech session opens September 18 in Shanghai Pudong, focused on domestic compute infrastructure, financial LLM applications, and risk/compliance ^[24].

3. Contrarian and Tail Risks

• The "slowdown" call is CEO-level signaling, not legislation: Altman says self-regulation works ^[3], Huang says no new laws needed ^[4], Johnson says no pause ^[5]. The falsifiable test is whether the Trump-Anthropic meeting ^[8] produces binding rules, or whether another Hugging Face-class agent incident ^[25][9] forces Washington's hand.
• Source quality control: the "AI giants call for slowdown" headline and the "computing-power metals" de-rating are single-source via Securities Times/Yicai ^[1]; the Anthropic CEO slowdown quote is via The Guardian without direct attribution in the feed ^[7]. Treat the safety chorus as a sentiment flag, not a confirmed market move.
• The KPMG UK cuts ^[16] and TSMC's R&D caution ^[2] are early corporate signals that some buyers are pulling back from frontier-AI integration into sensitive workflows - watch for replication in financial services ^[24] and chip design.