Chinese AI Labs Collective (2023)

Class card for the Chinese frontier AI lab ecosystem that crystallised 2023-2025 around a shared operational grammar: domestic AI sovereignty + open-weight strategy + cost-efficiency push that reframed global frontier AI competition. Key instances: DeepSeek (founded July 2023, Hangzhou, by Liang Wenfeng / High-Flyer Quant hedge fund); Alibaba Qwen series (Sept 2023 onwards, Alibaba Cloud R&D); Moonshot Kimi (founded Oct 2023, ~$3.3B valuation 2024, long-context specialisation); Baidu ERNIE Bot (Mar 2023, first government-approved Chinese chatbot); Zhipu GLM (Tsinghua University spin-out, est. 2019, GLM-4 Jan 2024); 01.AI Yi (Kai-Fu Lee, founded 2023); MiniMax abab. Atlas seed DM-38: fragmentation MEDIUM (multiple competing labs), ontological_doubt HIGH, regime dawn→day. Motor: push (DM default — independent cost-efficiency and capability pushes; no unified Progress telos). Wallerstein split: geographic semi_periphery (China, restricted from Nvidia H100/H200/B100 by US chip export controls Oct 2022 + Oct 2023 + Dec 2024) / functional core (Chinese labs define open-weight frontier benchmarks as of Jan 2025). Operational inflection: DeepSeek-V2 May 2024 ($5.8M training cost claim); DeepSeek-V3 Dec 26 2024 (671B MoE params, 14.8T tokens, $5.6M training cost); DeepSeek-R1 Jan 20 2025 (reasoning model rivalling OpenAI o1) → surpasses ChatGPT on iOS App Store rank Jan 27 2025 → NVDA stock -17% ($600B single-day mkt-cap drop). Alibaba Qwen-72B Sept 2023; Qwen-2.5-72B Sept 2024; Qwen3 Apr 2025 SOTA open-weight. Substrate incorporeal + cognitive + semiotic: operating grammar is model weights, training pipelines, open-weight publication, and the cost-efficiency narrative that reframes AI affordability globally. Heavy compute substrate (Alibaba Cloud, domestic H800/H20 GPU clusters, Huawei Ascend) is partially internalised — unlike OpenAI/DeepMind which fully externalise to Azure/GCP. dm_current = late_modernity: massive capital burn + ontological_doubt HIGH (multiple labs; fragmentation MEDIUM; governance tension between open-weight strategy and Chinese AI policy). tep_paradigm_primary = tep5_ict (existing ICT stack) + tep_paradigm_secondary = tep6_ai_biotech (AI irruption); contested = true (tep6 onset ambiguous). SCHEMA STRESS: class-vs-instances — treat DeepSeek, Qwen, Kimi, ERNIE as InstanceRef (no instance qualifies for promotion per §B.1: they don't yet each have ≥2 of distinct motor/wallerstein/scope, own snapshot ≥2 phases, AND ≥2 cross-era target references). V0.2 GAP: wallerstein_position_trajectory field may need additional entries as Chinese labs approach full core-functional status post-2025.