〔Day Digest〕AI’s Buildout Spans Record DRAM, a 3 Million-Robot Plan and a 7% Model-Stock Rebound — Hardware Cost Pressure Makes Governance the Test

0. Weekly Arc ^[1][2][3][4]

The World Internet Report 2026 frames the long run: AI is forecast to contribute $22.3 trillion to the global economy by 2030, equal to 3.7% of global GDP, while the Industry 4.0 market is projected to expand from $239.47 billion in 2026 to $801.49 billion in 2034 ^[2]. The immediate arc is more constrained: October global DRAM contract prices set a record, NAND held at the prior month’s high, and institutions still expect fourth-quarter DRAM gains, with MLC and SLC paths expected to diverge ^[1]. Against that, the session added three product-and-deployment signals: Apple opened mainland beta access, Ant expanded to 40 connected services, and JD announced a five-year procurement program of 3 million robots, 1 million driverless vehicles and 100,000 drones ^[5][6][7][3]. The cited equity reaction was narrow: Zhipu and MiniMax each gained 7% on company-specific commercialization and usage news ^[4]. Net arc: AI buildout is visible in infrastructure plans and services, but hardware costs and agent accountability increasingly determine whether that buildout creates durable economics ^{[1][8][9][10]}. Source note: the Taiwan semiconductor and Nvidia allocation datapoints are headline-only in this packet, while the 7% stock move rests on one market brief ^[11][4][12].

1. Memory and Compute Economics ^{[1][13][14][9]}

• **[NEW] Memory pricing:** October’s global DRAM contract average set a new record, while the NAND Flash average stayed at the previous month’s highest level; institutions expect DRAM prices to rise quarter on quarter in the fourth quarter, and MLC and SLC trends are expected to diverge ^[1].
• **[NEW] Hardware repricing:** Alibaba Cloud said global supply-chain changes and rising raw-material costs had significantly increased hardware production costs, prompting price increases on the Wuying Pro-NS01, NE01, 23.8-US01, 21.5-US05, Ultra-AX01 and Pro-AS06 models; it said other Wuying models would be repriced in stages ^[9].
• **[NEW] Utilization argument:** OpenAI CEO Sam Altman said only about 5% of the time available in DRAM installed in personal computers was being used, argued that concentrating memory in cloud data centers would improve utilization, and then clarified that he wanted global memory output increased enough for people to install more memory in their computers ^[14].
• **[ONGOING] Compute supply:** Sugon Chairman Li Jun said China’s computing industry should specialize and collaborate openly, using chips, base software, models and industry applications to convert industrial advantages into actual computing-power supply while reducing repeated investment ^[13].

2. Product Rollout and Industrial Deployment ^{[5][6][15][16][3][7][17][18]}

• **[NEW] Apple in mainland China:** Apple Intelligence’s mainland-China beta opened with visual intelligence, and the Dewu app was among the first Chinese applications to connect to that capability ^[5]. Apple also opened the iOS 27.3 Developer Beta to mainland users, planned a December iOS 27.3 release, and said its Alibaba partnership would connect Qwen to Siri, Writing Tools and Visual Intelligence ^[6].
• **[NEW] JD’s physical-AI plan:** JD Technology Committee Chairman Cao Peng said JD would procure 3 million robots, 1 million driverless vehicles and 100,000 drones over five years under the “JD Physical AI Acceleration Plan,” creating an intelligent supply-chain network ^[3].
• **[NEW] Enterprise surfaces:** Ant’s Lingguang added connectors for podcast subscriptions, Quark Cloud Drive, Hello, McDonald’s and WeChat Read, taking the total number of connected services to 40 ^[7]. Baidu Aicaigou’s forthcoming “AI Push” upgrade covers six AI-search channels—Baidu AI, Google, Bing, Tencent Yuanbao, Kimi and Nano AI—alongside an authoritative-source matrix, intelligent analysis and intelligent publishing ^[15].
• **[NEW] Adoption bottleneck:** The Wuhan Industrial Innovation Development Research Institute and Alibaba Cloud launched the Token Wuhan and Wuchuang Qwen dual platforms to address model selection, interface debugging, environment deployment and joint-testing friction faced by local small and midsize businesses and research institutions that often lack dedicated AI developers ^[16].
• **[NEW] Education model:** Hailiang Education Service’s eCareer released OpenSpan, described as the first national K-12 career large model, at the Wuzhen Open Source Pavilion ^[17].
• **[NEW] Regional cooperation:** A China-Korea embodied-intelligence forum in Seoul brought government, research-institute, university and company representatives together to discuss core technology, industrial applications, commercialization challenges and technical standards ^[18].

3. Enterprise Agents Move From Copilot to Execution ^{[19][8][20][21][10]}

• **[NEW] Training-data boundary:** Tencent said the limited cybersecurity data used during HY4 training was restricted to defensive material related to proof-of-concept generation and vulnerability patching ^[19].
• **[NEW] Governance mandate:** Hong Kong Monetary Authority President Yu Weiwen said AI had advanced from a question-answering tool to agents capable of independently executing tasks, and announced the SAGE project to research agent governance and responsibility arrangements ^[8].
• **[NEW] Financial workflow deployment:** Sunrate AI Chief Technology Officer Li Yilong said the company had deployed agents in cross-border payment workflows for compliance review, account opening, payments and foreign exchange; agents had moved from primarily providing chat-based customer service in the prior year to authorized task decomposition, orchestration and execution in the current year ^[8].
• **[NEW] Architecture framework:** Tianrongxin released a “three-domain joint defense, unified control” agent-security framework built around assets and supply chain, operations and behavior, and data and content, with governance mapped to objects, identities, behavior and data ^[20].
• **[ESCALATED] Security timeline:** Gartner warned that one quarter of security vulnerabilities could stem from agent abuse by 2028, while Gartner Research Vice President Gao Ting said 2027’s focus would shift from building and applying AI to managing autonomous systems and redesigning enterprise technology architecture ^[10].
• **[NEW, brief release]** The Wuzhen summit forum also published an Agent Development Self-Regulatory Convention, but the packet provides only the release title and no operative provisions ^[21].

4. Market Signal and Capital-Market Caution ^{[22][23][11][4][12]}

• **[NEW] Hong Kong model shares:** Zhipu (02513.HK) and MiniMax-W (00100.HK) each rose 7% in afternoon trading; the cited catalysts were better-than-expected commercialization data for Zhipu’s platform and API business and a recovery in call volume for MiniMax’s open models ^[4].
• **[ONGOING] Funding signal:** Demand for an ETF holding Anthropic private shares cooled after SpaceX’s record listing; earlier in 2026, the Procure Space ETF drew large inflows, the Tema Space Innovation ETF (ticker NASA) passed $3.2 billion in assets, and both products’ assets and trading volumes fell after SpaceX began trading ^[22]. The item described the ETF-demand slowdown as a warning signal for the AI-sector trade ^[22].
• **[NEW] Executive hire:** Uber President Jill Hazelbaker was set to join OpenAI as chief corporate affairs officer, taking charge of global communications and brand strategy amid frequent senior-team turnover and anxiety about a potential IPO; the role had been vacant since the beginning of the year ^[23].
• **[NEW] Thin supply-chain headlines:** A Caixin headline says Taiwan plans further restrictions on semiconductor exports to the mainland and heavier penalties for illegal transshipment, but the packet contains no article body to verify the scope or timing ^[11]. A Bloomberg headline relayed by Yahoo Finance says Nvidia’s China clients will receive 1% of a $70 billion RTX-chip order, but the packet provides no supporting article text ^[12].

5. What Decides Next — and What Would Falsify It ^{[1][22][8][3][9][20][4][10]}

• **Memory test:** The next checkpoint is whether DRAM prices rise quarter on quarter in the fourth quarter and MLC/SLC pricing diverges as forecast; failure of either condition would weaken the memory-pricing thesis ^[1].
• **Hardware test:** Further Alibaba Cloud price adjustments across additional Wuying models would confirm that supply-chain costs are reaching a wider hardware footprint ^[9].
• **Deployment test:** JD’s figures are procurement targets, so the operational checkpoint is delivery against 3 million robots, 1 million driverless vehicles and 100,000 drones over five years rather than the announcement alone ^[3].
• **Agent test:** The durability test for autonomous agents is whether SAGE yields concrete responsibility rules and whether Tianrongxin’s four governance dimensions become operating controls rather than framework language ^[8][20]. Gartner’s 2028 one-in-four vulnerability warning is the packet’s dated security marker ^[10].
• **Market test:** The equity test is whether Zhipu’s platform/API commercialization and MiniMax’s open-model call recovery persist beyond the 7% move; the Anthropic-ETF slowdown is the counter-signal ^[22][4].
• **Source control:** Treat items 12 and 25 as headline-only until the underlying reports are available; items 17 and 18 contain only titles and are not used for substantive inference; treat the 7% move as a single-brief datapoint rather than a broad market trend ^{[11][24][25][4][12]}.