〔Day Digest〕GPT-6Luna's $0.10/M Print Undercuts DeepSeek as 1.6T DSP Gap Widens to 20M+ Units in 2027 and Burry Lifts AI-Infra Shorts — Price War Meets Supply Crunch

0. Daily Arc

Two contrary signals hit the AI trade in one session. Michael Burry disclosed a 'sizable' increase in short positions in Micron, Nebius, the iShares semiconductor ETF and Palantir ^[1], while the physical supply chain points the other way: Nomura's Sep 22 expert call maps a 300-400K unit 1.6T DSP shortfall in 2026 expanding past 20M units in 2027, with no meaningful capacity release before H2 2027 ^[2]. Layered on top, OpenAI and Anthropic both reset token prices in the same trading session, with GPT-6Luna undercutting DeepSeek V4.1 Flash ^[3]. Net: a contrarian short, a structural shortage, and a price war, all live at once.

1. The Short Signal

• **[NEW] Michael Burry (Substack post):** disclosed a 'sizable' increase in short positions in Micron, Nebius, the iShares semiconductor ETF and Palantir ^[1]. The bet is the AI-infrastructure tape, not AI demand itself ^[1].

2. The 1.6T Bottleneck

• **[NEW] Nomura AI infrastructure expert call (Sep 22):** 2026 1.6T DSP demand 20M+ units vs 16-17M shipped, gap 300-400K; 2027 demand 60-70M vs 40-50M supply, gap 20M+; meaningful capacity release only H2 2027 ^[2]. Driver is 3nm wafer capacity and yield, not demand softness ^[2]. NPO/CPO evolution is shifting value from DSP toward TIA and driver ICs ^[2].
• **[ONGOING] Eoptolink (Xin Yisheng):** 1.6T in 'continuous ramp'; Thailand plant in continued expansion for H2 delivery ^[4]. Q3-on-Q2 shipment direction and overseas guidance are qualitative ('continuous growth'), not quantified ^[4].
• **[ONGOING] Hengmingda:** AI compute and communications segment gross margin 24.99%, up y/y; capacity insufficient, new capacity expected H1 2027 ^[5].
• **[NEW] Samsung Electric (subsidiary Aux Intelligent Technology):** signed a US data center energy storage oil-immersed transformer contract worth $132M (~8.85B RMB), 6.16% of 2025 audited revenue, labeled an 'important breakthrough' in US data center distribution ^[6].

3. The Token-Price Reset

• **[NEW] OpenAI:** launched GPT-6Sol at $2 input / $10 output per million tokens and GPT-6Luna at $0.10 / $0.50, both ~50% below prior-generation promo ^[3]. Three price tiers (Astra, Sol, Luna) now span enterprise reasoning to high-frequency tasks ^[3].
• **[NEW] Anthropic:** released Claude Opus 5.5 in the same session, ~90 minutes after the OpenAI event ^[3].
• **[NEW] Cross-vendor comparison:** GPT-6Luna input at $0.10/M sits below DeepSeek V4.1 Flash's off-peak price and Xiaomi's MiMo-V2.6-Flash ^[3]. Prior-generation Luna output was $1.20/M, now $0.50/M ^[3].

4. Materials, Macro and Policy Tailwinds

• **[NEW] Morgan Stanley:** initiates on CCL and copper foil, lifting global CCL TAM from $19B in 2025 to $47B by 2030 at a 20% CAGR vs consensus $35-40B; AI/data center accounts for 90% of incremental ^[7]. Overweight on Elite Material, Taiyo, Gold Circuit and Kingboard Laminates with targets NT$8,400 / NT$2,700 / NT$730 / HK$75 ^[7].
• **[NEW] JPMorgan:** previously initiated Kingboard Laminates at Overweight with a HK$65 target, modeling 8x EPS growth over 2025-2028 on CCL pricing, capacity and a 75% expansion in weaving-equipment scale ^[7]. Two-anchor buy-side view on the same name.
• **[NEW] CAICT:** 2025 global AI-related product trade reached $4.18T, +21.9% y/y, contributing 42% of total global trade growth ^[8]. Flags data-security and tech-sovereignty risks from model training and API calls ^[8].
• **[ONGOING] MIIT (vice minister Yu Xiaohui):** 15th Five-Year plan to 'appropriately advance' next-gen comms network, dual-10G broadband, low-orbit satellite internet, international network layout, and compute-network synergy ^[9].

5. The Agent Layer and Real-World Data

• **[ESCALATED] Alibaba at Apsara Conference:** CEO Wu Yongming said 'the representative product of the machine-intelligence era has not yet appeared'; debuted QwenBook hardware, QwenNote, and the Enterprise Context infrastructure ^[10]. Qwen Office VP Shu Junliang framed the office-agent thesis as 'Context is All You Need' ^[10].
• **[NEW] ByteDance Doubao Work:** added Goal mode and Plan mode for task differentiation ^[11].
• **[ESCALATED] Meta Muse:** launched Sep 8, hit #1 on both US iOS and Android free charts within 13 days with 2.5M+ downloads; Wallstreetcn analysis argues the 'new entry point' thesis does not hold, drawing the parallel to the 2016 Messenger bot arc ^[12].
• **[NEW] Kaiwang Data:** launched the Lighthouse Plan, recruiting 100 real-scene partners over 100 days to build a 1M-hour real-scene dataset across retail, office, home, logistics, industrial and F&B; collection, labeling and QA by end-2026 ^[13]. Headline scenarios already mapped to identified customer demand ^[13].
• **[NEW] iFlytek:** released Spark-ASR-2.0 with a non-autoregressive + LLM-enhanced autoregressive hybrid, claiming gains in Chinese-English mixed, dialect, terminology, high-noise, low-volume, fast-speech, child and contextual recognition ^[14].

6. Labor, Governance and Falsifiable Test

• **[NEW] Barclays 3A Framework (Sep 22):** only ~20% of core workplace skills are highly replicable by AI; cognitive exposure concentrated in computer/engineering/science, physical in construction/production/maintenance ^[15]. US hiring share in most-exposed occupations has fallen from ~30% in 2022 to ~25% ^[15]. AI-skill share in postings rose from ~2% in 2019 to 6-10% by Aug 2026 across US/UK/France/Germany ^[15].
• **[ONGOING] Regulatory/legal:** residents filed suit over OpenAI data center approval ^[16]; NPR frames an 'AI freeze' as potentially consolidating incumbents ^[17]; Business Insider profiles an Anthropic scientist on the AI-doom debate ^[18].
• **[ONGOING] CAICT digital inclusion report:** AI model performance on math reasoning, multimodal understanding and PhD-level science QA 'approaching or surpassing' human baseline, with the disruption now targeting 'cognition' itself ^[19][20].
• **Falsifiable test and source quality control:** the next 1.6T shipment print (Eoptolink Q3 ^[4], Hengmingda H1 2027 capacity release ^[5]) against Nomura's 60-70M 2027 demand forecast ^[2] is what confirms or breaks the supply gap; the Burry disclosure is single-source ^[1], the 1.6T figures come from a single expert-call note ^[2], and the Morgan Stanley vs consensus $35-40B CCL TAM gap is a model bet, not yet realized ^[7]. Quote each as a band, not a point.