〔Day Digest〕Palantir-NVIDIA Sovereign AI Lands In-House While Anthropic Outlines 15% GDP Extreme and Discloses Fourth Hack — Scale-vs-Safety Split

0. Daily Arc

The theme is scale-up meets safety friction. Palantir and NVIDIA moved sovereign AI from rhetoric to in-house deployment, with the first node at NVIDIA itself ^[1][2], while Anthropic's economic team published a 15% GDP extreme-case scenario that explicitly flags sharp displacement of knowledge workers ^[3] — the same day the company disclosed a fourth AI hacking incident and lost a researcher over safety ^[4]. In China, DeepSeek's V4.1 Flash cut HBM demand to one-quarter of the prior generation ^[5], Unitree open-sourced an embodied base model ^[6], and Moonshot AI explored a Hong Kong-Shanghai dual listing ^[7]. The pivot: capital and deployment are accelerating, yet the safety and embodied-intelligence debates ^[8][6][4] keep the tail live.

1. Sovereign AI and Infrastructure Deals

• **[NEW] Palantir + NVIDIA:** joint push to apply sovereign AI to critical supply chains; first deployment sits inside NVIDIA's own operations ^[1]. Codification uses Palantir Foundry with NVIDIA's Nemotron stack under the "wafer-out to first token" frame ^[2].
• **[NEW] Amap ABot-Earth 0.7:** described as the first 3D-native city world model, covering 196+ countries, from planet to street view ^[9]. Amap CEO Guo Ning framed the release as moving "digital Earth" from browse-only to a spatial-intelligence input ^[9].
• **[NEW] Envision "Tianji":** a physics-AI weather model with minute-level, 5km/10-min resolution out to 45 days; ranked first in China's NMC demonstration and placed top-tier on WeatherBench 2 alongside Google's WeatherNext 2, leading on wind-speed and precipitation metrics ^[10].

2. Anthropic: Transformative Scenarios vs Safety Breakdowns

• **[NEW] Economic Scenarios for Transformative AI:** three routes through 2030 in a US-base case, ranging from a mild "internet-style" path to an extreme where GDP could roughly double every 4.5 years; the report warns that in the substantive and extreme paths, "programmers and call-center staff may have to retrain as electricians or nurses" ^[3].
• **[NEW] Safety friction:** a fourth AI hacking incident was disclosed and a researcher quit over safety concerns ^[4]. Separately, a former Anthropic researcher warned AI could "kill us all" ^[11].
• **[NEW] BlackRock at AFD 2026:** told AI investors to "stay calm and focus on constraints" ^[12].
• **Source-quality flag:** items ^[12] and ^[11] are single-source relays; quote the tone, treat the substance as a datapoint, not a consensus.

3. China's Stack and the Embodied-Intelligence Debate

• **[NEW] DeepSeek V4.1 Flash:** smallest model in a new architecture family with native multimodal vision; KV-Cache compression cuts HBM demand to 1/4 and SSD demand to 1/8 versus the prior generation — the explicit target is Agent-class workload cost ^[5].
• **[NEW] Liang Wenfeng corporate vehicle:** Hangzhou Chengyan management consulting partnership established; reports say DeepSeek has commissioned CITIC Securities for a STAR Market IPO, with informed sources confirming the rumor ^[13].
• **[NEW] Unitree UnifoLM-WLA-1.0 open source:** single model coordinating desktop and whole-body mobile manipulation, with cross-task and cross-end-effector generalization ^[6].
• **[NEW] GPT-6 Astra "dimensional reduction" debate:** at the Bund Conference, Self-Variable Robotics Founder and CEO Wang Qian said the risk of language models generalizing into embodied intelligence is "not large at present"; he framed today's frontier models as "distillers" in training and "containers" in deployment ^[8].
• **[NEW] InHand Networks R&D:** disclosed research on embodied robotics for industrial flexible manufacturing, covering 3D vision, dexterous actuators, motion control and self-learning; some AI+IoT projects are at product launch or customer verification ^[14].

4. Capital Flows and Ecosystem Plays

• **[NEW] Moonshot AI:** considering a Hong Kong-Shanghai dual listing; weighed the STAR Market after a July meeting with investors on the HK leg ^[7].
• **[NEW] Alipay "Agent Emergence Award":** 10M yuan/year program with a 1M yuan grand prize; Ant Group CEO Han Xingyi framed it as a "supply-side" push for fragmented life-service demand, distinct from the mini-program ecosystem ^[15].
• **[NEW] Easygo GEO marketing solution:** covers ChatGPT, Gemini, Perplexity, Copilot; data covers a 350k+ brand library and 3M+ AI answer records, with 12k+ brands already using AI-search tracking ^[16].
• **[NEW] Pony.ai + Verne:** first fully driverless Robotaxi passenger test in Europe, on public roads in Zagreb ^[17].
• **[NEW] Kimi enterprise partner program:** builds Forward Deployed Engineer (FDE) teams with IT services, cloud, infrastructure and telecoms partners; partners already signed across software services, cloud, infrastructure and telecoms ^[18].
• **[NEW] Muxi + Shanghai Fuxin Chenmu:** 10M yuan consulting firm jointly held by Shanghai Futeng Private Equity Fund Management, Shanghai Huayan Enterprise Development Group, and Muxi's Muxi Digital Intelligence (Shanghai); legal representative Zhang Kun ^[19].

5. Policy, Geography and What Would Falsify It

• **[NEW] Hubei 15th Five-Year "Digital Hubei" plan:** extends the "light-chip-screen-terminal-network" stack into "sense-store-transmit-compute"; prioritizes embodied intelligence, 6G, brain-computer interfaces and quantum as future industries; targets 500+ digital-economy innovation firms and 5+ 10B-yuan digital industry clusters ^[20][21].
• **[NEW] AI in finance (Bund Conference):** Tsinghua Wudaokou's Zhang Jianhua said institutions that don't use AI "will definitely be eliminated"; consensus was that AI is not yet replacing human decision-making, with trust gates layered from "can do work" to "limited autonomy" ^[22]. Qiancheng Credit GM Li Zhen and Ant Consumer Finance Deputy GM and CRO Lin Jianan flagged use in real-need identification, differentiated credit and scenario-based risk control across 230+ consumption scenarios ^[22].
• **[NEW] China-to-ASEAN capability landing:** at WAIC Connect Malaysia, the panel frame shifted from "selling product" to "landing capability"; ModelBest Commercialization VP Li Zhu highlighted on-device closed-loop inference as suited to markets like Malaysia where infrastructure is still being built ^[23].
• **[NEW] AI in industry (thin):** a report cited by Xinhua outlined four trends for AI-enabled industrial development ^[24] — headline-only sourcing, do not over-extrapolate.
• **Falsifiable tests:** the safety-vs-scale gap is the swing variable. A second consecutive Anthropic safety incident within 30 days, or a slip in DeepSeek's STAR filing beyond Q1 2027, would harden the bear case ^[13][4]. On the upside, NVIDIA formally embedding Palantir Foundry into its broader supply chain — beyond a single in-house deployment — would confirm sovereign AI as a procurement pattern ^[1].