〔Day Digest〕Astra Resets the Compute Capex Curve — B200 Lease at $6.42/hr (+11.65% m/m), Cerebras $25.4B OpenAI Backlog, AMD $5B into Anthropic, Yet MiniMax 1H26 Lifts the Model-to-App Revenue Gap

0. Day Arc

Astra, the Cerebras backlog, the AMD-Anthropic stake, and rising GPU lease rates form a single compute capex cluster ^[1][2][3][4]. Yet MiniMax's 1H26 — $117M revenue (+283.1% y/y, exceeding full-year 2025's $79M), loss $358M narrowed 11%, B-end open platform and AI enterprise services 63.4% of revenue — confirms the model layer still outpaces the application layer, and AI short-drama minute-prices collapsed from 5,000 yuan to a few hundred ^[5][6]. Net: a compute capex acceleration running ahead of monetization.

1. Model and Compute Layer

• **[NEW] OpenAI GPT-6 Astra:** launched Friday as the 'world's most intelligent and best-aligned model'; Goldman Sachs Delta-One head Rich Privorotsky called it the moment a leading large model 'truly stood out' — the bull case the AI bull market had been waiting for ^[2].
• **[NEW] Spatial reasoning:** OpenAI's applied research head said spatial reasoning was one of the last human edges over computers and Astra closes that gap ^[7].
• **[NEW] Cerebras:** $25.4B backlog, much of it from a single OpenAI agreement ^[3].
• **[NEW] AMD-Anthropic:** up to $5B commitment, with Anthropic's IPO prospectus reportedly days away ^[4].
• **[NEW] GPU lease rates (Sep 6):** B200 at $6.42/hr (+0.31% d/d, +5.94% w/w, +11.65% m/m); H200 at $4.55/hr (-1.73% d/d, +2.48% w/w, -7.52% m/m); H100 at $3.06/hr (+5.88% d/d, +6.62% w/w, +12.50% m/m); A100 at $1.09/hr (+2.83% d/d, +5.83% w/w, +4.81% m/m); RTX5090 at $0.63/hr (flat d/d, +18.87% w/w, +28.57% m/m) ^[1].
• **[NEW] Microsoft FY27 restructure (Morgan Stanley):** collapses historical segments into 'Agents and Infra' and 'Devices and Consumer'; described as accounting reclassification with no change in underlying economics; Overweight reiterated, $600 PT ^[8].
• **[NEW] Policy catalyst:** seven Chinese departments exploring 800V HVDC power architecture for compute facilities, deploying single-rack 100kW+ cabinets — AIDC power-architecture reform opportunity, per analysts ^[9].
• **[NEW] Financing pivot:** Calcalist frames Nvidia as no longer just an AI chip company but financing the revolution it built ^[10].
• **[ESCALATED] Inner Mongolia first 'compute Token loan':** issued — compute entering the collateralized-asset lexicon ^[9].
• **[NEW] Cloud Leopard Intelligence:** positioning Greater Bay Area-to-APAC as the backbone of a global compute network ^[11].

2. Application and Monetization Layer

• **[NEW] MiniMax 1H26:** revenue $117M, +283.1% y/y, H1 exceeding full-year 2025's $79M; loss $358M narrowed 11% y/y; B-end open platform and AI enterprise services at 63.4% of revenue — growth engine flipped from C to B ^[5]. A Tiger Sniff commentary framed it as the model-to-app layer gap widening, not closing ^[5].
• **[NEW] Office agent migration:** 17 Chinese desktop AI office agents hit combined monthly visits >60M in June vs 20M in March (3x in three months); Alibaba, Tencent, Baidu, ByteDance all restructured or launched products; Changan Auto deploying agents across R&D, manufacturing, and supply chain, with one packaging reviewer cutting workload ~90% via self-taught Qianwen Office scripts and a procurement manager building ~50 AI Skills ^[12].
• **[NEW] AI short-drama price collapse:** per-minute production quotes fell from 5,000 yuan to a few hundred, near cost; custom short-drama still 10,000-20,000 yuan/min; 2026 China AI drama/漫剧 market expected >40B yuan, +138% y/y; overseas AI micro-drama market >$4B, ~70% of the overseas micro-drama market ^[6].
• **[NEW] Virtual influencer monetization:** AI virtual character Fang Taozi's ad-quote range hit 168,000-258,000 yuan; AI short-drama male lead Duan Yan also opening e-commerce partnerships ^[13].
• **[NEW] Retail investors (WSJ):** Americans vibe-coding trading algorithms and handing portfolios to AI agents — the 'robot retail investor' ^[14].
• **[ONGOING] Organizational drag:** Tiger Sniff commentary argued AI may not be the cure for organizational problems, citing DingTalk — a Chinese AI-tool company — itself trapped in efficiency issues ^[15].

3. Physical AI and Drug Pipeline

• **[NEW] Physical AI convergence:** former Ministry of Science Vice Minister Li Meng said static-question understanding isn't true intelligence; only dynamic physical causality is; physical AI takes AI from the digital world to the physical world ^[16].
• **[NEW] First AI-assisted innovative drug approved:** Westlake University/Westlake Lab/Westlake Pharma's Istreve hydrochloride tablets (brand: Aipesiwei) received conditional NMPA approval for adult mild/moderate COVID-19; China's top-tier Class 1 innovative drug ^[17].
• **[NEW] Embodied-intelligence robot CE mark:** Changsha Changtai Robotics' autonomous winding robot became China's first industrial embodied-intelligent robot to receive EU CE certification ^[18].
• **[NEW] Cybercab training:** accumulated driving training equivalent to over 16,000 human life cycles; omnidirectional sensing, never tired or distracted ^[19].
• **[NEW] Shanghai SASAC framework:** directing supervised enterprises to shift AI applications across four dimensions — coverage from pilot to full, focus from compute supply to networking + R&D + scenario outcomes, form from Q&A assist to agent execution (digital labor in investment advisory, design, supply chain scheduling), implementation from fragmented to ecosystem coordination ^[20].
• **[NEW] Guangdong AI core industry:** scale exceeds 300B yuan ^[21].

4. Politics, Governance, and Risk

• **[NEW] Trump data-center bet vs midterms:** reporting that President Trump's strong support for AI data centers could deepen GOP mid-term risk; ~70% of Americans don't want such facilities near home; Democrats tying infrastructure expansion to cost-of-living concerns; advisors urging Trump to emphasize AI job benefits ^[22].
• **[NEW] Asia-Pacific Media Forum (Shenzhen, Sep 5):** former PBoC deputy governor and former IMF deputy MD Zhu Min vs UTS adjunct professor Mark Beeson on AI diffusion and global governance — optimist vs skeptic framing ^[23].
• **[NEW] Dean Ball (OpenAI):** 'We should not think about AI as a race' ^[24].
• **[NEW] National anti-fraud AI app:** launched in Shanghai under Ministry of Public Security CID guidance, combining LLM, multimodal, and agent technology for risk analysis and scam-pattern recognition ^[25].
• **[ONGOING] Source-quality flag:** the Cerebras $25.4B figure, AMD $5B figure, and ~70% Trump-midterms poll number are single-source social-relay items; quote the bands, not points ^[22][3][4]. The MiniMax 1H26 model-to-app gap framing is editorial commentary rather than a stated numerical ratio ^[5].