AI Industry 2026-09-30 中文 PDF

NIGHTLY INTELLIGENCE BRIEF

〔Day Digest〕OpenAI Eyes $30B Pre-IPO at $1.4T as Kimi K3 Cracks Codex Billing, Samsung Targets 66% HPC by 2029 - Capital Surge vs. One-Page Safety Pact

OpenAI opened early talks for a $30B pre-IPO round at a ~$1.4T target valuation, with a 2027 listing expected, while Kimi K3 became the first Chinese open-source model to enter OpenAI's Codex enterprise billing channel. The capital push is broad: ACM Shanghai's order book hit 17.073B yuan (+88.2% YoY), ST Doushen signed a ~$119M GPU compute lease, Aoni Electronic a 1.821B yuan GPU card purchase, and Meta is using R&D tax credits to claim tens of billions in breaks on AI data centers. Yet the policy floor is hardening — US tech CEOs signed a one-page 'Superintelligence Agreement' at the White House, and Bank of England Governor Andrew Bailey called for more pre- and post-deployment model testing. The Samsung Foundry roadmap to a 66% HPC mix by 2029 and Hubei Starchain's first 3D advanced packaging line at 7B+ yuan decide whether supply keeps pace with the order book.

0. Daily Arc

Capital is racing ahead of the policy perimeter. OpenAI opened early talks for a $30B pre-IPO round at a ~$1.4T target valuation, with a 2027 listing expected [1]. On the same day, ACM Shanghai's order book hit 17.073B yuan (+88.2% YoY) [2], ST Doushen signed a ~$119M GPU compute lease [3], and Aoni Electronic signed a 1.821B yuan GPU card purchase [4]. Yet the safety frame caught up: US tech CEOs signed a one-page 'Superintelligence Agreement' at the White House [5], and Bank of England Governor Andrew Bailey called for more pre- and post-deployment testing [6]. Net: a compute-buying binge that the regulatory floor is only beginning to acknowledge.

1. Capital and Compute Pipeline

  • **[NEW] OpenAI pre-IPO**: early-stage talks for a $30B round at a ~$1.4T target valuation, targeting a 2027 IPO; no term sheet signed [1]. (Single source via Sina Finance relay — treat the valuation as indicative.)
  • **[NEW] ACM Shanghai**: order book 17.073B yuan as of Sept 29, +88.2% YoY versus its voluntary prior-year disclosure; cites 'high-speed growth' in semi equipment [2].
  • **[NEW] ST Doushen (300010.SZ)**: wholly-owned sub Doushen Hong Kong signed a 60-month ~$119M (~801M yuan) cloud service agreement — GPU compute leasing, deployment, and O&M — with 'VB client'; frames the deal as the move from 'AI applications' to 'compute infrastructure' as a second growth curve [3].
  • **[NEW] Aoni Electronic**: wholly-owned sub Aoni Intelligent signed a 1.821B yuan GPU compute card procurement contract with 'C company' [4].
  • **[NEW] Meta R&D credit push**: per NYT, Meta is classifying tens of billions in AI data centers as 'experimental R&D' to claim the federal credit; documents show the policy cut its 2024 tax bill by nearly $4B [7]. BPM partner Andre Shevchuck called the AI-data-center-as-R&D classification 'outside the norm' [7]. Internal Meta accountants warned of legal risk; the IRS has previously challenged similar reclassifications [7].
  • **[NEW] Baotou storage-compute integration**: Inner Mongolia Baotou project breaks ground, converting 'storage' into 'tokens' via co-located compute [8].

2. Foundry, Packaging, and Hardware Roadmap

  • **[NEW] Samsung Foundry**: Senior VP No Mi-jung told IR that customer count (already ~3x 2017) is targeted to ~4x by 2029, with HPC revenue share rising from 28% today to 66% by 2029 (vs 5% in 2017) [9].
  • **[NEW] TSMC Texas**: evaluating additional Texas investment beyond its $265B Arizona commitment, per two sources [10]. (Thin sourcing — two informants relayed; no confirmation.)
  • **[NEW] Hubei Starchain 3D advanced packaging**: Phase 1+2 combined investment >7B yuan; billed as China's first 3D advanced packaging mass production line; once fully built, it can carry 40-50 chip SKUs in pilot/risk production and verify 30-40 domestic semi equipment and material sets [11].
  • **[NEW] Lishan Tech 7G100 GPU**: completed tape-out, manufacturing, and packaging; pro and consumer SKUs launched; sales ramp and next-gen architecture are now the milestones [12].

3. Models, Distribution, and Applications

  • **[ESCALATED] Kimi K3 in OpenAI Codex**: Baseten announced enterprise users can run Kimi K3 inside Codex with charges folded into existing OpenAI spend commitments — first Chinese open-source model into OpenAI's enterprise procurement channel [13][14]. (Earlier, Amazon Bedrock onboarded Kimi K3 under a revenue-share model — first revenue-share deal between a Chinese model firm and a global top-tier cloud [13].)
  • **[NEW] Doubao**: ahead of Golden Week, added one-shot natural-language booking for flights, trains, ride-hailing, and navigation [15].
  • **[NEW] WuXi AppTec**: said on the interactive platform that AI in innovative drug R&D is in 'exploration and trial' stage, with future efficiency gains and lower barriers expected [16].

4. Embodied AI

  • **[NEW] Humanoid shipments, H1 2026**: IDC counts ~25,000 units globally (+432.1% YoY), with China at >19,000 (~77.9% of global) [17]. SAG separately counts ~19,100 globally (+272% YoY), with Chinese makers contributing 97%+ [17]. Both rank Zhiyuan above Unitree for H1 2026 shipments and revenue; Zhiyuan confirms >8,600 units shipped, ~35% global / 45%+ China share [17]. Last year Unitree led both volume and share [17].
  • **[NEW] Power grid embodied AI**: China Southern Power Grid signed a strategic framework with Unitree for grid O&M deployment; North China Electric Power University Energy Internet Research Center Deputy Director Wang Yongli pegs 2026 power-sector embodied AI investment to exceed 10B yuan [18]. Existing pilots cover switchgear inspection and infrared thermography [18].
  • **[NEW] Linjiedian OmniHand 3 Ultra**: completed foundational Su embroidery tasks — twisting thread, splitting silk, threading needle — under professional guidance [19].
  • **[NEW] OPC track**: 2026 AI Future Design Competition OPC (super-individual) track launched in Shanghai by MIIT's Industrial Culture Development Center, with hardware prototype and software tracks [20].

5. Policy Floor and Source Control

  • **[NEW] White House Superintelligence Agreement**: signed Sept 29 by US tech CEOs; one-page voluntary pact for stronger internal controls, independent external audit, and board oversight of frontier models; Trump called it 'morally binding' [5][21]. Same day, Trump directed federal agencies to prefer 'Super Intelligence' over 'Artificial Intelligence' in official communications [21]. Elon Musk, Jensen Huang, Tom Brown (Anthropic co-founder), and Gavin Baker (Atreides Management CIO) appeared together on stage [21]. Musk cited the US average ~500 GW power load and his bet that each 1% rise in power use maps to ~1% GDP [21].
  • **[NEW] Bank of England Governor Andrew Bailey**: called for more pre- and post-deployment AI model testing before regulation [6].
  • **[NEW] OpenAI 'Dots'**: industry safety focus shifting to 'What did my AI assistant do now?' per Axios [22]; Fortune frames the move as 'OpenAI fights Muse with Dots' [23].
  • **Source quality control**: the $1.4T OpenAI valuation [1], TSMC Texas expansion [10], and the 'Dots/Muse' narrative [22][23] are single-source or thinly sourced — quote as indicative, not confirmed. H1 humanoid shipment figures diverge materially between IDC (~25,000) and SAG (~19,100) — print both and flag the spread [17]. The Meta R&D credit push [7] carries a named tax adviser's dissent and an internal warning of IRS challenge; the upside is not risk-free.

SOURCE TRAIL

Citations

23 citation records

  1. [1]
  2. [2]
  3. [3]
  4. [4]
  5. [5]
  6. [6]
  7. [7]
  8. [8]
  9. [9]
  10. [10]
  11. [11]
  12. [12]
  13. [13]
  14. [14]
  15. [15]
  16. [16]
  17. [17]
  18. [18]
  19. [19]
  20. [20]
  21. [21]
  22. [22]
  23. [23]

    Google News — AI 产业OpenAI fights Muse with Dots - Fortune ↗