〔Day Digest〕ChiNext AI ETF +5% on CPO Shortage and Goldman's 224.9% Target Upside as Astra's 99.9% ARC-AGI-3 Reignites Capex — But AMD/Intel-Nvidia Split Flags Trade Rotation

0. Weekly Arc

The AI capex thesis re-lit on a three-way catalyst: CPO equipment-shortage reports, Goldman's doubled target on a leading Chinese optical-module name, and OpenAI's Astra benchmark refresh — pulling the ChiNext AI ETF target index (159382) up 5.05% intraday ^[1]. Yet the chip tape is splitting — AMD and Intel soared as Nvidia lagged ^[2] — and China is parallel-tracking: MIIT unveiled a RMB 3.8 trillion information-infrastructure plan through 2030 ^[3] while iFlytek released Spark X2.5 on a fully domestic compute stack ^[4] and Sichuan's first Token factory lit up in Ya'an ^[5].

1. Hardware and Capex

• **[NEW] CPO shortage + Goldman doubles target:** international customers are placing follow-up orders for silicon-photonics packaging and inspection equipment, with Hiwin confirming pull-in ^[1]. Goldman initiated H-share coverage, reiterated Buy on the A-share, and raised the optical-module target to RMB 2,645 — 224.9% upside from the RMB 814 reference ^[1]. Of 82 A-share CPO companies that have filed H1 results, 52 posted YoY net-profit growth and 800G optical modules are ramping fast ^[1]. A Cailianshe telegraph frames CPO equipment as the 0-to-1 phase, with 2026-2030 a "second-derivative positive" up-cycle as customer orders stack ^[6].
• **[NEW] 800G/1.6T PCB supply-chain validation:** a Nvidia + CPO + PCB supplier is shipping 800G and 1.6T optical-module PCBs in volume; Q2 800G-and-above optical-module revenue up 15x QoQ; M9-grade and PTFE material mass-production capability now in place ^[7].
• **[NEW] Beijing Creative (002350.SZ) subsidiary Cyber Engine:** won the RMB 285 million Thailand data-center diesel-generator module framework bid, ~13.25% of 2025 audited revenue; positive impact on 2026 and beyond, formal contract pending ^[8].
• **[ONGOING] Xihe Technology silicon-photonics IPO:** replied to first round of inquiry; pre-IPO valuation RMB 3.701 billion (Dec 2025); controlling shareholder Wu Aimin holds 45.3944% voting rights; backers include Yuanhe Yuandian, Shenzhen Dacheng, Langmafeng, Yaheng (Sequoia); high customer concentration flagged ^[9].

2. Models, Agents, and Regulation

• **[NEW] OpenAI GPT-6 Astra page update:** ARC-AGI-3 99.9%, FrontierMath Tier 4 ~97.6%, Terminal-Bench 4.0 57.9% (vs GPT-5.6 Sol 37.3%, Claude Fable 5.1 55.8%), Agents' Last Exam 59.3% (vs Claude Opus 5 55.5%); rolling out to Plus/Pro/Business/Enterprise and entering API, Azure, and AWS Bedrock "in coming days" ^[10].
• **[NEW] iFlytek Spark X2.5:** MoE architecture, fully domestic compute stack, focus on code and agent capabilities; live on the iFlytek Open Platform ^[4].
• **[NEW] Tencent Hunyuan Hy4 preview:** targeted fix for "long thinking and excessive self-verification under complex tasks" ^[11].
• **[NEW] Baidu Wenxin Kuaima (Baidu Comate) merged into Baidu Damate (DuMate):** Comate has served 10 million+ developers; by end-2025 AI-generated code at Baidu exceeded human-written, supporting 99%+ of engineers' daily R&D ^[12].
• **[NEW] WeChat internal test "Xiaowei AI Social":** a user's "Xiaowei" agent can contact a friend's "Xiaowei" agent to negotiate plans; the two AIs converse, and the user intervenes only at authorization and decision points; messages do not surface in normal chat windows ^[13].
• **[NEW] AI-agent dual-authorization dispute:** King & Wood Mallesons Shanghai partner Chi Cheng and Fadada chief legal officer Mei Zhen argue GUI-path agents execute user-authorized cross-app actions and do not require platform-level dual authorization; the friction should be resolved by compliance design and interop APIs, not by pre-authorization mandates ^[14][15]. Shanghai Academy of Social Sciences assistant researcher Qu Ran argues consumer choice rights extend to choosing decision-making tools, and that platform "one-size-fits-all" blocking undermines the comparison-shopping model ^[16].

3. Compute Build-Out and Consumer AI

• **[NEW] MIIT 15th Five-Year Plan (Information and Communications Industry Development Plan):** 13 main indicators including RMB 4.1 trillion industry revenue by 2030, RMB 3.8 trillion cumulative information-infrastructure investment, 7% average annual telecom-business-volume growth, 50 5G/5G-A base stations per 10,000 people, 9,800 EFLOPS of intelligent computing power, and 95% 5G/5G-A user penetration ^[3].
• **[NEW] Sichuan "Tianjiang·Baize Token Factory":** built by Tianfu Jiangxi Laboratory and Shutian Data Group in the Ya'an Economic Development Zone; ~3 trillion tokens this month, targeting 10 trillion tokens/month and top-tier domestic status by Q1 2027 ^[5].
• **[NEW] IFA 2026 Chinese-brand push:** TCL, Haier, Midea, Xiaomi, Roborock, and Dreame dominated the Messe Berlin entry-plaza ads; Samsung was notably absent from the main hall for the first time in years; NIQ GfK global strategic-client lead Cai Ling observed AI moving from product feature to home-intelligence entry point, with cleaning robots expanding into multi-scenario home executors ^[17].
• **[ONGOING] OpenAI data-center noise complaints (Savannah):** residents questioning noise impacts ^[18] — single-source local NIMBY, no operator response in the feed.
• **[NEW] Anthropic US$50bn compute build-out:** Fluidstack is leading the rollout per AI Magazine ^[19] — single source, link only, no financial detail.

4. Market Structure and Falsification

• **[NEW] Chip trade split:** AMD and Intel soared as Nvidia lagged — a Yahoo Finance headline raises whether the AI trade has structurally changed ^[2] (single source, no underlying factor attribution).
• **[NEW] Oriental Harbor (Aug 2026 view):** the past-two-month semiconductor-hardware correction is characterized as "rebalancing" rather than systemic risk; argues mainstream LLMs have completed initial agent-ification, unlocking new demand layers in education, healthcare, legal, software tools, video/games, and 3D printing ^[20].
• **[NEW] CFA roundtable on Agentic AI:** CFA Institute's Rhodri Preece, Brian Pisaneschi, and James Tate described a workflow shift from "code-driven" to "skills-driven" agents using Markdown-defined Skills and MCP servers; flagged loss-aversion bias in models as a governance blind spot ^[21].
• **Falsification test:** whether Astra's broader rollout ^[10] translates into sustained inference demand and re-accelerated capex; whether the CPO equipment-shortage reports ^[1][6] and the 800G+ revenue ramp ^[7] hold into Q3 prints; whether the AMD/Intel-Nvidia split ^[2] is rotation or a structural re-rating of non-Nvidia AI silicon.