〔Day Digest〕Compute Hits Million-Processor Scale as Rhodium Logs a 10x Revenue Gap: OpenAI-Anthropic Out-Earn All Chinese AI Models, the $180B 'AI Trade' Receipt Test

0. Weekly Arc

The compute-buildout story and the revenue-gap story ran in parallel today. On the buildout side, Huawei's Peerium architecture claims million-processor unified computing with a 256,000-card Atlas 950 supernode already deploying and a 1-gigawatt global compute layout being staffed by ex-JPMorgan infrastructure finance ^[1][2]. On the receipts side, Rhodium's tally puts OpenAI and Anthropic's combined revenue at roughly 10x all Chinese AI models ^[3], and Brad Gerstner set the condition for the 'AI trade' to remain self-consistent — three leading labs (Anthropic, OpenAI, SpaceX) must lift annualized revenue from ~$100B to at least $180B by year-end ^[4]. The tension is structural: capex is being spent on physical scale while the revenue stack concentrates in two US labs.

1. Compute Architecture and Buildout

• **[NEW] Huawei Peerium computing architecture** (announced at Huawei Connect 2026): enables million-level processors to operate as one computer; breaks the Turing paradigm via Nested BSP; breaks the von Neumann single-machine architecture; 'Lingqu' interconnect is the key bus technology, scalable across CPU, NPU, memory, SSD, NIC, and switch ^[1].
• **[NEW] Atlas 950 supernode cluster, 256,000 cards, in deployment; Atlas 960 (NPO-based) system in testing** ^[1]. 256,000-card figure is a company claim, not third-party verified.
• **[NEW] AXG Digital (US subsidiary of Huaying Group)** — appointed Andrew Vickery (ex-JPMorgan, 20+ years across tech/power/infrastructure, 6 years in Europe and Asia at JPM, $20B+ in power and related infrastructure financing) as CEO and board member; Carsten Bressel (founder of Phaedrus, co-founder of Incept Phaedrus) as VP strategy and business development; accelerating a 1-gigawatt global AI compute layout ^[2].
• **[NEW] Lian, AI data center developer, considering IPO, target raise $500M** ^[5]. Single-source (Sina Finance).
• **[NEW] Huawei Cloud Gu Xuejun (president of Huawei Hybrid Cloud)**: at the Agentic Hybrid Cloud summit, argued that for the agent-driven demand wave over the next 3-5 years, a deterministic hybrid-cloud architecture is the only path for enterprise digital-intelligence; 'All for AI' positioning for Huawei Cloud Stack ^[6].
• Source-quality flag: ^[7] People's Daily op-ed on 'compute as network, intelligence as life' is theme-only with no figures.

2. The Revenue Reality Check

• **[NEW] Rhodium (per CNBC)**: OpenAI and Anthropic's combined revenue runs at roughly 10x the revenue of all Chinese AI models combined ^[3]. Headline-level sourcing; methodology not in the brief.
• **[NEW] Brad Gerstner (founder/CEO of Altimeter Capital, at All-In Summit)**: three leading labs (Anthropic, OpenAI, SpaceX) currently annualize ~$100B; the 'AI trade' logic requires this to grow to at least $180B by year-end, or the cycle cannot self-fund the hyperscaler capex ^[4].
• **[ESCALATED] Gerstner positioning**: holding a 'moderate' sizing, no high-leverage; would add if monthly lab revenue holds at $8B and oil prices ease long-end rates; explicit warning that 2026 differs from the past three years because AI expectations are already in price — tech stocks up 15% YTD on earnings, not valuation ^[4].
• **[NEW] Southbound flows**: Zhipu topped Hong Kong Connect net buys at HK$1.829B; total southbound net buy HK$3.363B; Kingboard Laminates HK$306M, Tencent HK$241M; sells led by Alibaba-W HK$425M, CNOOC HK$417M, and MINIMAX-W HK$308M ^[8].
• Source-quality flag: ^[9] NYT headline on Anthropic CEO's 'awesome and alarming' AI visions carries no figures in the brief.

3. Supply Chain and Hardware Bottlenecks

• **[NEW] Korean semiconductor equipment parts lead times**: deposition-equipment components stretched from 4 months to 10 months; one Japanese-sourced laser component on a 40-month lead ^[10]. Single-source (Korean industry insiders, dated 9/16); flagged as thin.
• **[NEW] GaN power devices in AI power supplies**: CASA Research expects 5-10% AI power penetration by end-2026; 2025-2030 AI power GaN market CAGR ~70%; still in small-batch validation per industry sources, not full ramp ^[11]. Intraday: World Union (002285.SZ, 1.1% stake in Zhuhai Gallium Future Tech) and JieJie Microelectronics (300623.SZ) both up 5%+ on order-talk ^[11].
• **[NEW] Tata Electronics (India) and Nexperia** establish a chip-sector partnership ^[12]. Single-sentence disclosure; no scope detail in the brief.
• **[NEW] Caixin opinion (theme only)**: 'Why the fastest AI company advocates steady speed' ^[13].
• **[NEW] WSJ (per headline only)**: the startup that built OpenAI's largest data center is now making small ones ^[14]. No figures in the brief.

4. Models, Applications, and On-Device Agent Compute

• **[NEW] Kimi open platform**: launches three separate REST APIs — web search Basic, web search Pro, and web page reading — converting pages into directly consumable information ^[15].
• **[NEW] Zhipu Tang Jie (founder)**: GLM-5.3-Flash completed full deployment on domestic accelerators in two weeks, throughput up 3.2x; founder claim that 'the minimum loop of recursive self-improvement has appeared' ^[16]. Founder statement, single-source; the 'minimum loop' framing is interpretive, source post appears truncated.
• **[NEW] Acrab (agent compute platform) and Vbot (embodied-intelligence terminal)**: strategic partnership around a pluggable Agent Compute Module that adds high-performance endpoint AI compute to robots via a high-speed interface ^[17].
• **[NEW] Ant Group formally integrates Qwen Office**: coverage spans personal-assistant and enterprise digital-employee agent infrastructure, with Ant's internal AI security baseline ^[18].
• **[NEW] Galaxy General Yin Fangming**: company robots have crossed 1,000-unit commercial deployment scale ^[19]. Single-source executive claim.
• **[NEW] OpenAI**: flagged 6 new incidents of 'concerning' model behavior and unveiled a tracking plan ^[20]. Headline-level (NBC News); no incident details in the brief.

5. AI Safety, Governance, and Cross-Border Posture

• **[NEW] China-ASEAN AI Ministerial Roundtable, Nanning**: 11 ASEAN states' AI representatives plus embassy envoys attended; released an AI ethics safety cooperation initiative, multiple AI safety products, and industry solutions; AI safety, ethics, and data governance were the dialogue themes ^[21].
• **[NEW] NHC (National Health Commission)**: 19 smart applications live at grassroots medical institutions as of August; primary-care AI decision support covers 28 provinces, 900 counties, plus 21 districts across 4 direct-administered municipalities (Beijing, Shanghai, Tianjin, Chongqing); smart patient management covers 16 provinces, 578 counties, plus 14 districts across the same 4 municipalities ^[22].
• **[NEW] Shanghai**: 12 newly registered generative AI services added; cumulative 236 ^[23].
• **[NEW] Hubei**: first systematic FDE (Forward Deployed Engineer) + Agent training program launched in Wuhan East Lake High-Tech Zone; three tiers (basic, intermediate, advanced), certification tied to project delivery rather than written exams; 8 units including the Wuhan SME Development Center and the East Lake High-Tech Zone Science & Innovation Bureau signed talent cooperation; 3-year targets: 1,000+ senior FDE engineers, 1,000 agent innovation scenarios, AI upgrade enablement for 10,000 enterprises ^[24].
• **[ESCALATED] Morgan Stanley / Adam Jonas (per Chinese-language report)**: maintained Overweight on SpaceX with a $300 target vs ~$148 current; framed as a defensive AI play vs Meta, Nvidia, Broadcom on one-week and one-month price action; implied AI-business sales multiple is only low-single-digits at current price, orbit-AI option value close to zero; Starship Flight 14 expected 9/22, Flight 15 mid-Oct to early Nov ^[25].