〔Day Digest〕Muse Tops U.S. iOS in Five Days, Lifting Meta ~12% and Zuckerberg +$25B; Jefferies Sizes PCB/ABF at $73B/$30B by 2030 - SoftBank's $50B Data Center IPO Delay Tests the Agent Capex Bet

0. Daily Arc

The day's tape rotated around Meta's Muse going viral—surpassing ChatGPT in U.S. iOS downloads within five days of its September 8 launch and topping the U.S. iOS App Store free chart by September 18 (10 days post-launch), lifting Meta nearly 12% and adding an estimated $25 billion to Mark Zuckerberg's net worth in a single session ^[1][2]. The agent-viral moment is repricing upstream infrastructure: Jefferies sizes AI PCB TAM at $7 billion (2025) growing to $73 billion by 2030 at ~60% CAGR, with ABF substrate rising from $8 billion to $30 billion at ~30% CAGR ^[1]. Yet capex conviction is cracking in parallel—SoftBank's U.S. data center arm, backed by OpenAI and Nvidia with no operational facilities yet, is slowing a ~$50 billion IPO and a $5 billion bond deal ^[3], while chip designer "Mr. Bubble" argues the real AI bottleneck has moved from chips to packaging, interconnect, and memory ^[4]. Net: an agent-era repricing of the upstream stack, with the durability of demand still the open question.

1. The Agent Pivot — Muse and the Consumer Test

• **[NEW] Meta Muse adoption:** topped the U.S. iOS App Store free chart on September 18, 10 days after its September 8 launch; cumulative downloads exceeded 2.5 million by Monday; surpassed ChatGPT in five-day U.S. iOS downloads per Sensor Tower ^[1][2].
• **[NEW] Stratechery's Ben Thompson:** personal agents' moat may be far stronger than chatbots'—once users feed personal data and integrate the agent into daily life, switching becomes much harder, posing strategic pressure on OpenAI and Anthropic ^[5].
• **[NEW] Kevin Kelly, speaking in China:** "in the next five years, 99% of AI may not directly face humans"—agents will call other agents, with humans interfacing with only 1–2 "manager" agents; engineer workflows already run 6–10 agents in parallel; within five years, capital flowing through the agent economy could exceed that through the human economy ^[6].
• **[ONGOING] SpaceXAI released Grok 4.7 on September 22 Beijing time:** scored 46 on the Artificial Analysis intelligence index (sixth place), lifting SpaceXAI into the global top four AI labs; strong gains on long-horizon agent and coding tasks, but web, 3D-scene, and visual-coding tests showed "benchmark-strong, deployment-flip" cases ^[7].

2. Infrastructure Repricing — PCB, ABF, Server CPU

• **[NEW] Jefferies:** AI PCB TAM to grow from $7 billion (2025) to $73 billion (2030), CAGR ~60%; ABF substrate TAM from $8 billion to $30 billion, CAGR ~30%; driver is the agent transition pulling inference, orchestration, and infrastructure-layer compute demand ^[1].
• **[NEW] Hygon Information released the CPU 1000 series, its first end-side compute chip:** built on the C86 architecture; President Sha Chaoqun said it extends the company's compute footprint from cloud to edge and end, with the series flexing CPU/GPU/NPU resources for embedded use ^[8].
• **[NEW] Deep Blue Auto + Intel AI BOX:** mounted on the Deep Blue L06 Special Edition, bringing PC-class end-side AI compute into the vehicle, with on-device LLM, voiceprint recognition, and AI panoramic sound; user behavior data stored locally only ^[9].
• **[NEW] Stifel maintained Buy on Oracle:** Project Jupiter data center construction rose to 26%, up 17 percentage points from July; New Mexico Supreme Court unanimously rejected two legal challenges to project permits; air-quality permit hearing still pending examiner and date; FY2027 Q1 revenue and EPS both beat ^[10].
• **[NEW] Firebird (San Francisco) brought a Hrazdan, Armenia data center online:** targets 300 MW capacity and 70,000+ Nvidia chips by end-2027, expected to rank among Europe's largest; 20% reserved for domestic use, balance allocated to foreign firms including PerplexityAI; site is ~45 km from Yerevan ^[11].

3. Capex Skepticism and the IPO Test

• **[NEW] SoftBank's U.S. data center arm (SB Energy) is slowing its IPO:** backed by SoftBank, OpenAI, and Nvidia, with no operational data center yet, it is still targeting a ~$50 billion valuation; a potential $5 billion bond deal also went cold amid AI-boom sustainability doubts ^[3].
• **[NEW] Chip designer "Mr. Bubble" interview:** AI compute infrastructure's real bottleneck has shifted from "making faster chips" to packaging, interconnect latency, and memory hierarchy—"Moore's Law economics have broken" ^[4]. *Thin sourcing: single podcast interview; no second-source corroboration of the bottleneck thesis.*
• **[ONGOING] Trillion-yuan flows into chips (First Financial):** this AI-driven cycle is restructuring "information production," not just distribution; the market debate is whether it is a long-cycle industrial dividend or another short thematic ^[12].
• **[NEW] Lei Jun (Xiaomi) clarification:** the "Lei Jun earned over 10 billion yuan on Unitree IPO" claim is false; Shunwei Capital (not Xiaomi) was the sole A-round investor in Unitree, and Shunwei shareholders benefited ^[13]. *Single source, interested party.*

4. China Going Global — Payment Rails, Policy, and the Pudong Stack

• **[NEW] Antom (Ant International) one-stop overseas solution for AI firms:** covers subscription, payment, tax, risk, and capital management; supports 300+ global payment methods across 200+ markets; signed partners include DeepSeek, Lovart, QwenCloud, and Kimi ^[14].
• **[NEW] Invesco's Christopher Hamilton (Asia-Pacific ex-Japan Head of Client Investment Solutions):** Chinese AI models need not penetrate the U.S. to succeed—China itself is a massive economy and industrial export powerhouse capable of deploying and scaling AI faster than any other nation; the market has historically underestimated Chinese policy impact, with current policy focus on AI tech and industrial-export sectors ^[15].
• **[NEW] Shanghai Pudong New Area released 11 "AI+Government" high-value scenarios:** four directions—serving enterprises, serving citizens, empowering the city, intelligent risk control—with a global open call for solutions launched in parallel ^[16].
• **[NEW] Goodwe Smart Electric Technology signed the "AI Full-Stack Cooperation Framework Agreement" with Alibaba Cloud:** collaboration on energy-industry large models, new-energy agents, and cloud services base ^[17].
• **[NEW] Alibaba's AutoNavi at the 2026 Yunqi Conference:** released the full-space unmanned system and Qianyu platform, organizing 80 million+ POIs, 5 million km of roads, and 7,500 sq km of 3D models into an AI-readable, computable spatial data system ^[18].

5. Policy, Content, and What Would Falsify the Agent Trade

• **[NEW] UK Parliament:** Business, Innovation, Science and Trade Committee invited Meta, Google, OpenAI, and Anthropic to an October 13 AI safety hearing on mandatory pre-release safety testing, mandatory serious-incident reporting, and standards-race risks; the UK AI Safety Institute was also invited ^[19].
• **[NEW] Anthropic and OpenAI jointly called on Australia to relax its ban on training AI models** ^[20].
• **[NEW] OpenAI and Microsoft executives' own quotes are being used by news outlets to threaten their AI copyright defense** ^[21].
• **[NEW] NVIDIA CEO Jensen Huang denied claims that AI would lead to human extinction** ^[22].
• **[NEW] Douyin's first AI Creative Competition:** 170,000 AI works over 72 days, 40+ billion views; Douyin announced a partnership with iQIYI with hundreds of millions in resources to extend top short content into drama and film; over 20 AI-native series have crossed 100 million views on Douyin in the past six months ^[23].
• **[NEW] Media sub-sector AI penetration exceeds 80%:** AI is now the core amplifier of hit content profitability ^[24]. *Single source, analyst note.*
• **[NEW] NPR:** "AI solved one of math's hardest problems; humanity learned nothing (so far)"—a sober counterweight to the agent-economy enthusiasm ^[25].
• **Falsifiable test:** whether Muse's viral ramp converts to durable infrastructure orders (server CPU, PCB, ABF) and whether SoftBank's IPO delay spreads to peers or stays idiosyncratic; binary catalysts next are the OpenAI copyright case and the UK safety hearing on October 13 ^[19][21]. The two views disagree openly: Jefferies sees agent-driven compute demand compounding at 60% CAGR ^[1], while "Mr. Bubble" argues the demand cannot be met by chips alone and that the binding constraint has already moved downstream ^[4]—print both, treat the second as thinner.