〔Day Digest〕Shanghai Lays AI-Compute Industrial Blueprint as Xiaomi Drops 3nm D100, Samsung Maps 3D zHBM, and Goldman Lifts 2027 WFE 45% — the Memory Wall Becomes the Binding Constraint

0. Daily Arc

A coordinated Chinese compute-supply push ran into a US-side memory squeeze in Monday's session. Shanghai's 15th Five-Year industrial blueprint put AI compute at the center of industrial policy, targeting world-class chip clusters and a chip-model-cloud autonomous ecosystem ^[1][2]; Xiaomi unveiled the D100, China's first 3nm smart-driving high-compute AI chip ^[3]; and Samsung laid out a three-stage HBM roadmap culminating in 3D-stacked zHBM ^[4]. Goldman Sachs raised global WFE 2026-2028 to $150B / $218B / $281B, with 2027 y/y lifted to 45% from 32% ^[5]. Yet Micron HBM design architect Sreeramaneni warned the "memory wall" is widening — AI compute grows 3x per 2 years versus HBM under 2x ^[6]. Capital is following the supply thesis: Alibaba Chairman Joseph Tsai and CEO Wu Yongming bought HK$120M of stock a day after an HK$80B AI-funded placement ^[7], XPeng Robotics closed $900M+ at a $6.3B post-money ^[8], and "China's Palantir" Zhongshu Ruizhi closed a strategic round with national-fund backing ^[9]. Pivot: which cracks first — the memory bottleneck or embodied-AI demand.

1. Policy Stack: Shanghai's 15th Five-Year Blueprint

• **[NEW]** Shanghai municipal government's 15th Five-Year Plan on accelerated new industrialization and a modern industrial system names intelligent computing chips, next-generation large models (architecture, training optimization), embodied AI, and a chip-model-cloud autonomous stack as priorities, with a "Moli Jushen" action for smart robotics ^[1].
• **[NEW]** Same plan targets international-advanced levels for high-performance computing chips, general-purpose processors, memory, interconnect, and smart sensors, organized as a "one-body-two-wings" cluster in Zhangjiang, Lingang, and Jiading ^[2].
• **[NEW]** Hong Kong Financial Secretary Paul Chan, in his blog, said frontier tech will pressure the labor market and Hong Kong must accelerate reskilling, especially for youth, with full government support to SMEs ^[10].

2. Compute Supply: Chips and the Memory Wall

• **[NEW]** Xiaomi's Xuanjie D100: 3nm process, 20-core high-performance CPU + 16-core high-compute NPU, up to 160GB unified memory, can locally deploy 200B-parameter models; R&D complete, commercial launch next year ^[3].
• **[NEW]** Samsung, at Hot Chips, unveiled a three-stage HBM roadmap presented by DRAM design team member Sangwook Han. The terminal stage — zHBM — stacks DRAM directly on GPU/TPU compute dies, eliminating the 2.5D interposer. Claimed gains vs HBM4E: 70% power reduction, 230% DRAM bandwidth increase, 100W per-module savings, 8.3% additional GPU power headroom ^[4].
• **[NEW] (single-source warning):** Micron HBM design architect Sreeramaneni warned the "memory wall" is worsening — AI accelerator compute grows 3x per 2 years while HBM grows under 2x per 2 years, risking deepening memory shortages; new compute-storage boundary-blurring designs are needed ^[6].
• **[ESCALATED]** Goldman Sachs, in an Aug 23 report, lifted 2026-2028 WFE to $150B / $218B / $281B (y/y 36% / 45% / 29%, from 32% / 32% / 12%). Foundry WFE raised to $58B / $84B / $109B (y/y 45% / 45% / 30%, from 30% / 30% / 16%), driven by an 8-percentage-point uplift to each year of TSMC's capex assumption ^[5].

3. Capital and Vertical Adoption

• **[ESCALATED]** Alibaba Chairman Joseph Tsai bought 720K shares at ~HK$112 (~HK$80M) and CEO Wu Yongming bought 350K shares at ~HK$111.6 (~HK$40M), totalling ~HK$120M / 1.07M shares a day after Alibaba's HK$80B new-share placement fully earmarked for full-stack AI and AI infrastructure ^[7].
• **[NEW]** XPeng Robotics' first external round raised $900M+ at a $6.3B post-money (~$43B yuan), a record single-round PE figure for Chinese embodied intelligence; led by IDG Capital with Gaorong Ventures, plus strategic capital from Tencent and Alibaba; XPeng Group retains control. Use of funds: mass production and physical-AI model R&D ^[8].
• **[NEW]** Beijing Zhongshu Ruizhi ("China's Palantir") closed a "hundreds of millions of yuan" strategic round — its third such round in roughly a year — with China Internet Investment Fund, Suchuangtou, the National Social Security Fund, Financial Street Capital, ICBC Capital, and Kunlun Capital joining existing shareholders ^[9].
• **[NEW]** US legal-AI unicorn Harvey launched Tenet, built on Moonshot AI's open-weight Kimi K3 via post-training — the first Chinese open-weight backbone to enter a top-tier US vertical-AI training stack. Harvey: $11B valuation, ~$3.5B annualized revenue, raising at ~$15.5B; previously reliant on closed-source OpenAI/Anthropic, serving 2,400+ institutions and 200,000+ lawyers ^[11].
• **[NEW]** Alibaba DAMO Academy, with Shengjing Hospital (China Medical University) and others, published liver-cancer AI model DAMO LiON in Nature Medicine; in a 2-month prospective trial it caught 15 previously missed malignant tumors, mostly ~1cm ^[12]. At the second World Humanoid Robot Conference, the Tiangong Omni humanoid (Tiangong Team) won the 400m obstacle race in 4:11.44, with Zhiyuan and Tiangong teams taking 2nd and 3rd ^[13].

4. Enterprise AI: FDE Wars on the "Last Mile"

• **[NEW]** OpenAI in May set up a Deployment Company dispatching forward deployment engineers (FDEs) into enterprises; in June, Anthropic partnered with IT services firm DXC to train "tens of thousands" of Claude-certified FDEs for banking, aviation, insurance, and manufacturing; on Aug 18, Tencent Cloud renamed its agent-platform engineer certification to ADP FDE ^[14].
• **[NEW]** Minglamp Technology CEO Wu Minghui, at the 2026 World Robot Conference, said the new FDE role "goes deeper" than traditional deployment — agents must be wired into real business workflows while field learnings are fed back to the platform ^[14].

5. Falsifiers and Tail Risks

• Pairing ^[6] and ^[4]: if HBM4E production scales faster than the under-2x/2yr trajectory Sreeramaneni flagged, the memory-side bottleneck eases; if zHBM slips, the 2027-2028 WFE beat-and-raise in ^[5] loses its memory justification.
• Anthropic is reportedly probing abnormally elevated error rates in Mythos 5, Fable 5, and Opus models ^[15] — single-source from 财联社, treat as unverified.
• Harvey at $11B valuation against ~$3.5B ARR ^[11] prices in roughly a 3x revenue multiple, an order of magnitude above legacy SaaS; the falsifier is net-dollar-retention prints as the open-weight Chinese backbone scales.
• Source-quality note: ^[16] (NYT, Nvidia minicomputers in Moscow drones), ^[17] (Axios, Anthropic mission-vs-money interviews), ^[18] (The Neuron, Anthropic IPO), and ^[19] (Motley Fool chip pick) appear in the packet as headlines only; cite as directional, not data.