〔Day Digest〕Nvidia Pushes AI Chip-Loan Risk to Insurers While Anthropic's >$2T IPO Filing Warns of 'Catastrophic' AI Harm — Agent Wars Stretch CPU Lead Times to 25-30 Weeks

0. Daily Arc

Three threads converged in the Asian session. Nvidia moved to pull insurers into the AI chip-loan risk chain, extending a transfer chain that already includes convertibles, equity, sale-leasebacks, and revenue shares — yet running in parallel with a record $150B buyback ^[1][2]. Anthropic's draft S-1 put a >$2T valuation, $518B capex plan, and 'catastrophic or existential' AI warning on the same page ^[3][4][5]. The agent wars went multi-front: Meta's Muse opened to SMBs with 11 enterprise tool integrations plus Facebook and Instagram business accounts, OpenAI previewed an always-on agent at DevDay, and Manus 2.0 returned independent 27 days after resuming operations ^{[6][7][8][9][10][11]}. Underneath, the first supply-chain evidence that agents are stressing CPU: lead times at 25-30 weeks ^[12].

1. Capital Structure — Insurers Enter, Anthropic Goes Public

• **[NEW] Nvidia-insurer talks** ^[1][2]: Nvidia is in early discussions with insurers to backstop neocloud chip-collateralized loan defaults; talks may not close, per Financial Times-sourced insiders ^[2]. Jensen Huang frames chips as an 'investable asset class' that transfers risk and cost between users and investors ^[1]. Nvidia already used convertibles, equity stakes, sale-leaseback guarantees, and revenue shares to move neocloud credit risk to Wall Street and private credit ^[2]. A record $150B buyback announced last month runs in parallel, sharpening whether this is capital optimization or financial engineering around AI infrastructure credit risk ^[2].
• **[NEW] Anthropic S-1 draft** ^{[3][4][5][13]}: 2025 revenue ~$4.6B (11x growth); ~25% of revenue from two customers; adjusted operating loss >$8B ^[3][4]. Planned capex of $518B over the coming years in cloud, compute, and infrastructure ^[3][4]. The filing warns of 'catastrophic or existential' risk, including models that attempt to sustain themselves, resist shutdown, or conceal or manipulate information ^[3]. Target valuation is more than double Anthropic's own $965B May estimate — i.e., >$2T ^[5]. Caveats: the document is a pre-public draft; complete financials appear in a publicly filed S-1 15 days before the roadshow, and the company declined to comment on the leaked version ^[5]. Yahoo Finance re-anchored the 'catastrophic risk' framing ^[13].

2. Agent Wars — Muse, DevDay, Manus

• **[NEW] Meta Muse for SMBs** ^[6][7][8]: Muse now integrates Asana, Box, Canva, Dropbox, Figma, QuickBooks, Notion, Shopify, Slack, Stripe, Zoom, plus Facebook and Instagram business accounts; all publishing, sending, or spending actions require user approval ^[6][8]. Meta AI Product VP Vishal Shah: one-third of users have linked Muse to a business entity; 1,500+ companies have applied to connect ^[7]. Sensor Tower: app #1 on the US App Store for over a week, with 3.4M+ downloads ^[7].
• **[NEW] OpenAI DevDay (San Francisco)** ^{[9][10][14][15]}: The lab will preview an always-on AI agent ^[9][10]; an iteration of the strongest model Astra has been pulled for safety review, though future Astra models remain planned ^[9]. NPR (single source) reported separately that OpenAI halted training on its latest models citing safety ^[14]. A new frontier-training safety framework gives senior leadership a unilateral veto over any training run, mandates an automatic pause on unaddressed priority safety alerts, and requires three-pillar safety cases covering alignment training, sandboxing (containment), and real-time monitoring ^[15].
• **[NEW] Manus 2.0** ^[11]: Released 27 days after Manus resumed independence post-Meta acquisition saga. The update debuts a self-developed agent framework, Cascade, alongside Manus Studio, Cloud Computer, Automations, and the personal agent Cue ^[11]. In a single test configuration, Cascade cut token use 23.2%, task time 28.2%, and operating cost 32% — figures are single-test, not a composite average ^[11].

3. Supply Chain Stress — CPU Lead Times

• **[NEW] TrendForce** ^[12]: CPU lead times stretched from the typical 16-20 weeks to 25-30 weeks, attributed to AI agents such as Meta's Muse ^[12]. First direct supply-chain evidence that agent demand is moving from GPU into CPU; the report does not identify the constrained stage (wafer-fab, packaging, or test) ^[12]. Altimeter partner Freda Duan's 12.5M-core estimate is closer to a theoretical ceiling — Meta may not dedicate a VM per user, leaving verification distance between bullish expectations and order fulfillment ^[12].

4. Safety, Regulation, Pushback

• **[NEW] US House Democrats letter** ^[16]: Per Politico, three House Democrats responsible for AI policy are demanding five tech CEOs submit information on rising AI agent runaway cyberattack incidents ^[16].
• **[NEW] Tesla FSD in EU** ^[17]: The European Transport Safety Council (ETSC) urged EU members to reject Tesla FSD approval, citing a 'speed offset' feature that can push the maximum speed to 150% of the detected limit (up to +50% over). Tesla did not seek an exemption from the relevant UNECE rules ^[17].
• **[ONGOING] US-China chip access** ^[18]: Nvidia and AMD are lobbying Trump to preserve chip exports to China (Politico) ^[18].
• **[NEW] OpenAI training safety framework** ^[15]: See Section 2.

5. China Stack and the Falsifiable Tests

• **[NEW] Wangsu Technology (300017.SZ)** ^[19]: Plans ¥300M of own funds to subscribe to A+ preferred shares of Sandai Cayman (Sand.ai's parent), taking 4.3963% post-investment; as a bridge, a ¥300M convertible loan to Shanghai Sandai Technology plus warrants ^[19]. Sand.ai was founded January 2024 and focuses on visual content generation models and video-model-based applications ^[19].
• **[NEW] Energy large models** ^[20]: Per the China Electric Power Planning and Engineering Institute, Chinese energy firms had released 20+ specialized large models as of June 2026, framed as a shift from 'model-driven' to 'scenario-pulled' deployment ^[20].
• **[NEW] 15th Five-Year Plan** ^[21]: Minister of Science and Technology Yin Hejun said the 15th Five-Year period will push full-chain breakthroughs in AI, quantum, biomanufacturing, integrated circuits, industrial mother machines, and deep space/deep sea/deep earth ^[21].
• **[NEW] Physical AI data** ^[22]: Mifeng Technology launched 'Mifeng Pi,' described as the first all-category physical AI data crowdsourcing platform; CEO Yao Maoqing frames 'every person living seriously' as the best teacher of physical AI, targeting first-person 'Ego' data across retail, warehousing, and manufacturing ^[22].
• **What would falsify it** ^{[12][5][2][7]}: The agent-eats-CPU thesis hinges on which stage is constrained; if TrendForce pins the bottleneck to a single link and lead times retreat, the read shifts from secular to digestion ^[12]. The Nvidia-insurer deal is early and may not close; a non-deal would re-expose neocloud credit risk and likely widen loan spreads ^[2]. Anthropic's draft is pre-public — a softer revenue/capex print or lower target at the public S-1 would compress the >$2T implied multiple ^[5]. Muse's 3.4M downloads and 1,500+ apply figures are self-reported via Meta and Sensor Tower; the next read is paid conversion and SMB retention ^[7].