〔Day Digest〕AI Capex Pivots from Silicon to Wattage: ~75GW Behind-the-Meter Orders and a ~$284M LG CNS GPU Buy Stack Up, Microsoft Targets ~38GW by 2032 - Power Now Sets the Pace

0. Day Arc

AI capital expenditure is rotating from silicon to wattage. SemiAnalysis counts ~75GW of behind-the-meter power orders globally, ~20GW of which were signed in Q2 alone, with LG CNS committing 381.4 billion KRW (~$284M) to NVIDIA gear and Microsoft targeting ~38GW of data center capacity by 2032 ^[1][2]. Yet the consumer side is cracking: 90%+ of AI short-drama producers are loss-making ^[3][4], and OpenAI's CFO argues that, as intelligence becomes a commodity, "judgment" is the new scarce asset ^[5]. Per the SemiAnalysis read, compute is no longer constrained by GPU supply but by electrons ^[1].

1. Power Replaces Silicon as the Bottleneck

• **[NEW] SemiAnalysis via Yicai:** behind-the-meter power for AI compute has stacked up to ~75GW of confirmed orders globally, with ~20GW signed in Q2 2026 alone; gas generation and storage are moving into hyperscale site planning ^[1].
• **[NEW] LG CNS:** will spend 381.4 billion KRW (~$284M) on NVIDIA GPUs and infrastructure, the parent LG Group's IT arm doing the deal ^[2].
• **[NEW] Microsoft (single source / unverified):** reportedly targeting ~38GW of global data center capacity by 2032, more than triple current levels ^[1].
• **[NEW] Google:** announced a €13 billion AI infrastructure investment in Finland over two years, alongside a 22-year nuclear power agreement ^[1].
• **[NEW] NVIDIA plus Australian data center partners:** plan to drive up to 2GW of AI infrastructure capacity in Australia by 2027, against a current national data center base of ~1.6GW ^[1].
• **[NEW] (outlier, single source) Rapidus via Nikkei:** the Japanese advanced-chips company is weighing a chip fab on the Moon around 2040 ^[6]. Treat as signal-of-ambition, not roadmap.

2. Commercialization Reality Check

• **[NEW] DataEye via Yicai:** of 221,900+ AI short dramas launched in H1 2026, only 1,055 (0.47%) crossed 100 million views; using 50 million views as the break-even line, fewer than 1.3% recoup costs; about 90% of AI short-drama companies are loss-making ^[3][4].
• **[NEW] Industry consolidation:** active producers fell from 1,216 in Q1 2026 to 698, a 42% drop ^[3]; the Shanghai exchange issued a regulatory warning to Longban Media (605577.SH) for under-disclosing AIGC revenue and impact on results ^[4].
• **[NEW] Taobao Flash Sale (淘宝闪购):** launched the "Wending" (问鼎) chain-F&B AI management system, positioning it as a 7x24 "AI operating partner"; says it has helped 1,000+ brands accumulate 100 million incremental members with a 20% lift in repurchase frequency ^[7].
• **[NEW] Zibo municipal "AI + Manufacturing" 2026-2028 plan:** targets 10+ industrial LLMs, 3+ industry corpora, 10+ industrial agents, 50+ application scenarios, 200+ smart factories, six AI capability centers, and one embodied-robot training ground by 2028 ^[8].
• **[NEW] Databricks:** crossed $7 billion in annual recurring revenue, up from under $1 million when CRO Ron Gabrisko joined; the company attributes the run to data context, governance, usage pricing, and a rebuilt sales motion ^[9].
• **[NEW] Zhipu (智谱):** plans to place ~21.97 million H shares at HK$714 each ^[10].
• **[NEW] Moonshot AI (月之暗面, single-source wire):** reportedly targets $2 billion in annualized sales by end-2026 ^[11]. Treat as guidance, not confirmed revenue.

3. Security and Safety

• **[NEW] Ant Group AI Security Lab:** its Tiangong Cyber model and Tiangong Harness topped the CyberGym benchmark with a 98.5% vulnerability-verification success rate, fusing DeepSeek V4 Pro and Qwen 3.8 Max into the harness ^[12].
• **[NEW] Anthropic:** says Russian hackers used Claude AI to automate malware evasion ^[13].
• **[NEW] Anthropic (same campaign):** separately warned of attempts to use AI to build biological weapons ^[14]. Sourcing is third-party reporting; quote both threats, not the magnitude.

4. Investment Logic Shifts

• **[NEW] a16z podcast (Sep 10, single venue):** partners Jen Kha, David George, and Accolade's Aram Verdiya argue AI is pushing the venture "power law" from an industry feature to a systemic market rule; only 20 of ~3,000 US VCs have delivered sustained 3x net returns over 20 years, and AI crossed the $100 billion revenue scale unusually fast ^[15].
• **[NEW] OpenAI CFO Sarah Friar on Consello's Grey Matter:** as "intelligence" commoditizes, "judgment" becomes the scarce asset; many firms are halving back-office headcount and redeploying to customer-facing roles ^[5].
• **[NEW] Zheshang Securities:** maintains Buy on Sanrenxing (三人行), citing its JV with Zhonglian Data to build out an AI compute-center platform; FY26-28 net profit estimates 0.27/0.34/0.39 billion yuan, on 36x/29x/25x ^[16].
• **[NEW] (single source, headline only):** The American Prospect reports OpenAI hired Senator Chuck Schumer's daughter away from Amazon ^[17]; Fortune Tech covers Anthropic's bioweapon block, Altera's IPO preparation, and OpenAI courting Wall Street ^[18]. Treat as signal, not substance.

5. Frontier Notes

• **[NEW] Huawei director Yang Chaobin at the 2026 China Unicom partner conference:** "computing power substrate" is the new infrastructure of the intelligent era; China Unicom has deployed Ascend compute and 384 super-nodes, and is adapting Ascend 950 to Unicom Cloud ^[19].
• **[NEW] ZTE president Xu Ziyang, same venue:** network connection targets are shifting from "Internet of People" to "Internet of Intelligence," with the value ruler moving from bits to tokens, and demand tilting toward large-uplink and deterministic guarantees ^[20].
• **[NEW] Qiyuan Robot (启元机器人) at the Bund Conference:** showed a health-check app on the Q1 humanoid that captures heart-rate and SpO2 from a one-minute handshake, co-built with Hangzhou Feikuo Technology ^[21].
• **[NEW] Xiaomi ecosystem:** established Beijing Patika Intelligent Technology, covering AI software and intelligent robotics R&D and sales, via Xiaomi subsidiary Hanxing Venture Investment ^[22].
• **[NEW] Ant Lingbo CEO Zhu Xing:** pushed back on "needing cash or compute" as the financing motive, citing talent incentives and multi-cycle resilience as the real reasons ^[23].