〔Day Digest〕AI Capex Spreads Beyond Hyperscalers: UBS Lifts S&P 500 to 8,100 by 2026, Gold to $5,000 — Even as CoreWeave Drops 1% Pre-Market on Slowdown Tag

0. The Daily Arc

The AI trade bifurcated: UBS strategist Bhanu Baweja's team broadened the capex thesis past hyperscalers and raised the S&P 500 year-end target to 8,100 with gold at $5,000 ^[1]. Yet CoreWeave slid 1.11% pre-market to $82.06 after Bernstein called it the single name most exposed to a training slowdown, following Monday's -6.75% close ^[2]. Bill Gates, separately, said the "do-good" window is 12-18 months ^[3]. Net: a spreading capex cycle running into a tightening slowdown narrative, with the China-US frontier-model gap described as a "super-competition" by Gates ^[4].

1. Market Read

• **[NEW] UBS AI breadth call** (Bhanu Baweja, UBS strategist): AI capex is converting into earnings across industrials, healthcare, utilities and financials, not just hyperscalers and chipmakers; S&P 500 target 8,100 by end-2026 (+~6% from Monday's close) and 8,900 by 2027; Stoxx 600 to 690 (+~8%) by year-end and 760 by end-2027 ^[1]. UBS calls the long-end rates view too hawkish, expects the first Fed cut in June 2027, and tags gold at $5,000 as the preferred fiscal/geopolitical/monetary hedge ^[1].
• **[ESCALATED] CoreWeave (CRWV.US)**: Pre-market -1.11% to $82.06 after Monday's -6.75% close ^[2]. Bernstein: ~25% of operational US power capacity and ~74% of contracted capacity sits in Tier 3/4 markets; take-or-pay backlog "won't be threatened" but unsold rural capacity demand could pull back if training slows ^[2].
• **[ONGOING] Anthropic $2T IPO**: Still on the table per Bloomberg headline flow ^[5]. Source quality: single headline relay, no body content provided ^[5].
• **[NEW] Risk-sentiment framing** (thin sourcing): Yahoo Finance flags a chart on how Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have "scared people about AI" ^[6]; USA Today op-ed says AI experts are sounding alarms and government won't listen ^[7]. Both are framing pieces, not data ^[6][7].

2. China AI Stack: From Silicon to Fund Vehicles

• **[NEW] MediaTek Dimensity 9600 Pro**: TSMC 2nm process, 2+3+3 all-big-core design, LPDDR6 and UFS 5.0; dual NPU, with the "super-performance NPU" supporting on-device 30B MoE and the "super-efficiency NPU" cutting always-on sensor hub power by 40%; first phones "soon" ^[8].
• **[NEW] Unisound U2-Flash**: ~266B total parameters with ~10B activated per inference (<4% activation ratio), positioned as a high-frequency, cost-efficient workhorse alongside Google's Gemini 3.8 Flash, Zhipu's GLM 5.3 Flash and DeepSeek's DS V4.1 Flash ^[9].
• **[NEW] Stepfun Step Audio 3 series**: Five models — Realtime, ASR, TTS, Gen and Music — live on the Stepfun open platform ^[10].
• **[NEW] iFlytek AStudio**: Desktop AI productivity platform built on the Spark X2.5 model ^[11].
• **[NEW] Runze Technology (300442.SZ)**: Plans 1 billion RMB (20.4081% of a 4.90001 billion RMB vehicle) as LP in the Xiamen Shidai Shenyuan venture fund targeting AI, embodied intelligence and frontier tech; cleared at GM office level, not board or shareholder ^[12].
• **[NEW] Inx (因克斯) Series B**: 300M+ RMB closed; investors include Citic Goldstone, Nice Group, Suzhou Venture, Huarui, Nanjing Jiaokong and Huarui; follow-on from Fosun, Shenzhen Capital, Huakong, Jinqiu and Puhua; Minglun Capital as long-term exclusive financial advisor ^[13].
• **[NEW] DeepKernel seed round**: Led by Shunwei Capital; sovereign AI infrastructure focus ^[14].
• **[NEW] Ant Group SingProbe**: LLM-native security guardrail adapted to 29 open-source models including Ling-3.0, GLM-5.2/5.3, Qwen and DeepSeekV4, integrated with SGLang and vLLM; code, models and benchmarks open-sourced ^[15].
• **[NEW] Tencent Cloud × U Mobile**: Strategic partnership to push 5G + AI integration in Malaysian enterprise ^[16].
• **[NEW] Alibaba Cloud**: General vouchers disabled for Qoder CN products from September 30, covering Qoder CN, QoderWork CN, Qoder CN CLI, QoderWake CN and Qoder Cloud Agents ^[17].

3. The Slowdown Trade — Winners and Losers

• **[NEW] Barron's winners-and-losers framework**: A training slowdown would first benefit the AI labs (Anthropic, OpenAI) on margin, with a price-war risk; Google, Meta, Microsoft and SpaceX gain relative catch-up; cloud providers (Amazon, Microsoft, Alphabet, Oracle, CoreWeave) are mixed as training rev decelerates but inference demand offsets ^[18]. Chinese labs are likely to stay outside any coordinated slowdown pact and may close the gap ^[18].
• **[NEW] Bill Gates' "do-good" window**: 12-18 months for AI to be steered toward health and other positive uses; "we shouldn't waste any time" ^[3].
• **[NEW] Bill Gates on China-US parity**: Both countries have 4+ institutions operating at state-of-the-art; the top-8 global models all outperform the best from four months ago; the dynamic is "super-competition" in talent and compute ^[4].
• **[NEW] Jensen Huang (NVIDIA CEO)**: In China, "no one talks about 'AI doomsday theory'" — the narrative is more pragmatic than in the US ^[19].

4. Governance, Safety and the Demand Backdrop

• **[NEW] Microsoft "Humanist AI Code of Conduct"** (MAI model): Sets the principle that "people are more important than AI"; MAI does not resist or evade human control, stays within authorized scope, does not originate independent goals, and surfaces limits with explanations ^[20].
• **[NEW] Shanghai "Zhi Ce Neng Shou" platform**: National-first generative-AI safety testing public service platform, launched September 14 at the Shanghai leg of the National Cybersecurity Week; claims ≥95% accuracy on automated judgement of illegal content, bias and factual errors ^[21].
• **[ONGOING] WTO forecast**: AI could lift global GDP 13.2% and global trade 40% by 2040, with digitally-deliverable services leading the trade gain ^[22].
• **[NEW] NVIDIA open-source deployment**: A major children's hospital uses open-source NVIDIA AI for cardiac care ^[23]. Source quality: corporate blog relay, no quantitative outcome disclosed ^[23].
• **[NEW] Editorial framing** (state media, treat as opinion): People's Daily says AI is moving from "used" to "used for value" in the real economy ^[24], and tracks two growth curves in China's compute build-out ^[25].

5. What Decides Next

• Falsifiable test for UBS breadth: the first industrials/healthcare AI-driven earnings prints versus the next hyperscaler capex guide ^[1]. If the S&P 500 fails to track 8,100 on a softer earnings mix, the breadth claim is wrong ^[1].
• Falsifiable test for CoreWeave: any disclosure that Tier 3/4 unsold capacity demand is rolling over; take-or-pay protects the near-term backlog but the rural optionality is the swing ^[2].
• Falsifiable test for China-US frontier parity: independent benchmark deltas between the top-8 models — Gates' view is that the spread is already small ^[4].
• Source quality control: the Barron's slowdown framework and the Gates "do-good" window are single-source relays ^[18][3]; the UBS targets are firm-published but are model-derived ^[1]; the Anthropic $2T IPO and Yahoo/USA Today framing items are headline-only ^[6][5][7].