〔Day Digest〕AI Buildout Widens: 3.2 Pb/s Fiber, 3,000+ Satellites and 108 Game-Company AI Investments - RMB 6.63bn Component Outflows Meet RTX 5090 Above RMB 50,000

0. Daily Arc

• **[ONGOING] Capacity versus near-term pricing:** The packet's center of gravity was a widening gap between announced AI capacity and the immediate price and capital-flow backdrop for AI hardware: Shanghai set a 3,000-plus satellite and 150 Tbps network target, a Chinese laboratory set a 3.2 Pb/s fiber record, and A-share hardware drew heavy outflows ^[1][2][3][4].
• **[NEW] Enterprise demand provided the counterweight:** Bank of America expects Azure demand to stay above supply during FY2027, while Schneider Electric, Nvidia and HPE introduced edge-agent products ^[5][6]. The falsifiable test is whether those launches and the Shanghai network milestones produce sustained use rather than only announcements ^{[5][6][1][2][7]}.
• **[NEW] Source-quality flag:** The edge-agent and Nvidia architecture items are headline-level, the Anthropic lawsuit item is also headline-only, the Vietnam chip item cites unnamed sources, and the Anthropic/OpenAI IPO item is an unnamed market report ^{[5][8][9][10][11]}. These leads are thinner than the more detailed Bank of America forecast, Shanghai plan excerpts, Huaqiangbei merchant interviews and laboratory transmission release ^{[6][12][1][2][7][3]}.

1. Infrastructure Buildout

• **[NEW] Shanghai satellite roadmap:** Shanghai's Municipal Commission of Economy and Information Technology and five other departments issued a 2027-2030 satellite-internet action plan targeting more than 3,000 satellites in orbit, a core industry above RMB 80 billion, more than 100 converted research results, more than 100 companies to be cultivated and more than 100 flagship applications by 2030 ^[1][2].
• **[NEW] Network specifications:** The plan calls for VDES deployment, more than 150 Tbps of network capacity, inter-satellite laser-link networking, less than 5 ms rerouting latency, greater than 99.9% switching success, more than 30 POPS of AI computing power per satellite, direct-to-phone connectivity, space-based computing and integration of space-air-ground networking with communications, sensing, computing and intelligence ^[2][7].
• **[NEW] Fiber record:** China Information and Communication Technology Group's Optical Communication Technology and Network National Key Laboratory, with Pengcheng Laboratory and FiberHome Fujikura Optical Fiber Technology, said it completed a real-time bidirectional 3.2 Pb/s transmission over multi-core fiber, described as a world record for single-fiber real-time bidirectional capacity ^[3]. The test equates to more than 20,000 4K movies downloaded per second, 32 times the roughly 100 Tb/s ceiling of conventional C+L-band single-core G.652 fiber; it also validated 400G S-band and 1.2 Tb/s C+L-band optical modules, and the result was submitted to the Journal of Lightwave Technology ^[3].
• **[NEW] Cambricon product signal and rumor denial:** Cambricon Chairman and General Manager Chen Tianshi said the company's next-generation smart-processor microarchitecture and instruction set will target training and inference for natural-language, video-and-image-generation and vertical-domain models, with product competitiveness sought in programming flexibility, ease of use, performance, power consumption and die area ^[13]. Separately, Chen said the company had never discussed a forecast of more than 1 million chip shipments next year and directed investors to its Shanghai Stock Exchange and company disclosures ^[14].

2. Enterprise AI and Capital Allocation

• **[NEW] Microsoft outlook:** Bank of America expects Azure demand to remain above supply during FY2027, forecasts FY2027 capital expenditure including finance leases at $220.7 billion, up 51.9%, revenue near $391.9 billion and EPS of $19.70, and retains its Buy rating and $600 target ^[6].
• **[NEW] Edge-agent launches:** Schneider Electric, Nvidia and HPE introduced edge AI agents for industrial monitoring, cybersecurity and energy optimization, while Nvidia separately detailed an architecture for secure enterprise AI agents ^[5][9]. The packet contains only headline-level descriptions of these launches, with no customer, deployment or financial detail ^[5][9].
• **[NEW] CPU evaluation catches up with agents:** Arm introduced a CPU evaluation methodology for agent workloads, AMD has proposed a framework covering tool execution, data retrieval and task orchestration, and Intel has tested Xeon 6 for agent throughput and concurrent execution ^[15]. The mechanism is that agentic work adds database access, code execution and tool calls to GPU-heavy inference, making CPU behavior part of the end-to-end workload ^[15].
• **[NEW] SK Hynix return package:** Industry sources say SK Hynix is preparing an additional KRW 50-60 trillion shareholder-return package, equivalent to $35-42 billion, which would combine with a completed KRW 40 trillion buyback cancellation for record 2026 shareholder returns of KRW 100 trillion; the board is expected to finalize the package on October 28 alongside Q3 results ^[16].
• **[NEW] Corporate signals:** Zhiyuan Innovation (Shanghai) Technology's registered capital rose from about RMB 91.959 million to RMB 96.825 million ^[17]. Separately, a wholly owned 360 subsidiary joined Yuanhuan Intelligent as a shareholder; Yuanhuan's registered capital increased to RMB 1.2753 million ^[18]. A person close to ByteDance denied that Zhou Chang, one of Seed AI's core R&D leads, had resigned ^[19].

3. Compute Markets and Constraints

• **[ESCALATED] China hardware risk-off:** A-share performance was structurally split: the Shanghai Composite gained 0.19% to 3697.52, the Shenzhen Component fell 1.19%, ChiNext fell 1.71% and the STAR 50 fell 2.23%, with 3033 decliners and RMB 1.61 trillion of turnover; component and semiconductor net outflows were RMB 6.63 billion and RMB 4.87 billion ^[4]. The AI hardware chain was the principal area of selling, while the cited overseas reference points were the Nasdaq down 0.21% at 25955.34, Nvidia at $210.96 and the Hang Seng Tech Index down 0.23% ^[4].
• **[NEW] Consumer GPU inflation:** At Shenzhen's Huaqiangbei market, merchants said high-end consumer GPU quotes changed daily: the RTX 5060 traded around RMB 4,000 versus about RMB 3,000 two weeks earlier; the RTX 5070 traded around RMB 8,000 and some above RMB 9,000, with one seller reporting three increases and a nearly RMB 1,000 rise over a week; the RTX 5080 rose RMB 6,000-7,000; the RTX 5090D traded around RMB 30,000; and the RTX 5090 exceeded RMB 50,000 ^[12]. The report linked the broader squeeze to continued AI data-center demand, but the price evidence comes from merchant interviews at one market ^[12].
• **[ESCALATED] Bottleneck migration:** SemiAnalysis founder Dylan Patel said at the October 23 CASPA annual conference that the hard bottleneck in AI infrastructure construction had shifted from power to chip packaging substrates, citing long-standing underinvestment across substrate, PCB-material and upstream-material chains and cautious expansion by core Japanese suppliers ^[4]. The supplied item ends before its two-year outlook is complete, so no numerical forecast from that passage is usable ^[4].
• **[ONGOING] Thin market headline:** Barron's headlined that Nvidia shares were hurting while Musk's large AI plans could provide support, but the packet supplies no figures or supporting mechanism ^[20].
• **[NEW] Thin sovereign report:** A media report citing unnamed sources says Vietnam's goal of starting chip pilot production by end-2027 is likely to be missed ^[10]. The item provides no revised timetable or supporting detail ^[10].

4. Adoption, Robotics and Open Models

• **[ESCALATED] Gaming becomes an AI investment pocket:** Gamma Data's 2026 China Game Industry AI Development Report says the 50 game companies ranked by revenue initiated 108 AI investment cases in the first half, already above the 89 recorded for full-year 2025; 80% disclosed AI industry plans and 56% had created AI laboratories or specialist teams ^[21]. The report was cited as the direct catalyst for a game-stock rally as funds rotated out of computing hardware and precious metals ^[21].
• **[NEW] Chinese foundation models dominate Korean releases:** A media report citing a Rockbridge Network Korea study says Korean developers published 1,845 models on Hugging Face from January through August 2026, with 49.3% based on Chinese models and Alibaba's Qwen alone at 46.8%, versus 35.9% for US models and 14.9% for Korean models ^[22]. The cited item gives no methodology for overlap among the reported categories, so the percentages should be quoted individually rather than added ^[22].
• **[NEW] SuperMimic compresses training cost:** A University of California, Berkeley SuperMimic paper compressed full-body humanoid-controller training compute by more than 1,000 times, from 128 GPUs for one week to one H100 for three hours; the prior best public system, SONIC, required about 21,000 GPU hours, and the paper was included in CoRL 2026 and posted on arXiv on October 20 ^[23]. Nomura International (Hong Kong) analysts Frank Fan and Donnie Teng said falling motion-software barriers shift competition toward hardware consistency, warning that companies able to preserve joint accuracy in mass production may hold the more durable advantage ^[23].
• **[NEW] Heavy-load embodied robotics:** Zhiyuan formally released the safety-grade heavy-load embodied robot Spirit G2 Max, which it called the industry's first robot for long-range, multi-task heavy-load work ^[24]. The unit has a 20 kg single-arm load, 50 kg dual-arm peak load and repeat-positioning accuracy within +/-0.5 mm ^[24]. It is in validation and deployment on consumer-goods, e-commerce and food-production lines for depalletization, unpacking and loading, empty-carton recovery and warehouse inbound work; on Huguang's robot-manufacturing factory's Qiyuan T1 line, it matched a cycle of one robot every 2.5 minutes ^[24].

5. Legal and Political Risk

• **[ESCALATED] AI enters the US midterm campaign:** As the November 3 US congressional midterms approached, researchers identified 220 political ads nationwide that used AI tools at least in part, with more than $137 million in total spending, versus less than $50,000 spent on identified AI-generated ads in 2024; 81% were sponsored by Republican candidates or groups, 12% by Democrats and the remainder by independents, and many attack ads were not labeled as AI-generated ^[25]. The Wesley Media Project said comprehensive counting is difficult and that the overall number is likely understated ^[25].
• **[NEW] Thin litigation lead:** A Euronews headline asks why Elvis Presley rightsholders are suing Anthropic, but the packet gives no allegations, court venue or claimed damages, leaving the matter's financial scale unquantified ^[8].
• **[ESCALATED] IPO positioning:** Market reports say political risk is pushing Wall Street to prepare quickly, with some institutions urging Anthropic and OpenAI to bring possible IPOs forward rather than delay over AI-bubble concerns ^[11]. Large investors are reportedly concerned that a Democratic-controlled Congress could restrict major AI transactions and question executives about unemployment and deepfakes, while Meta could conduct an unusually large secondary share sale in coming months to fund AI buildout ^[11]. Because the report relies on unnamed market sources and conditional political scenarios, it should be treated as a positioning signal rather than a confirmed corporate transaction ^[11].