2026 estimate (CNY)
publicly released
share (2025)
investment (after US & UK)
1. The DeepSeek Shock — What Actually Happened
On January 20, 2025, DeepSeek released R1, an open-source reasoning model that matched OpenAI's o1 on multiple benchmarks while claiming a training cost approximately 1/30th of GPT-4. Within 24 hours, Nvidia shed ~$600 billion in market cap — the largest single-day loss in US stock market history.
The real disruption wasn't the benchmark scores. It was the combination of three factors: (1) a highly efficient Mixture-of-Experts (MoE) architecture, (2) reinforcement learning for chain-of-thought reasoning instead of vast RLHF compute, and (3) full open-source release under a permissive license. These reset global assumptions about the cost of frontier AI.
2. The Model Wars — China's 8 Major AI Players
Over 100 Chinese LLMs have been publicly released. Here are the eight you need to know:
Largest installed user base via Baidu Search. Deeply integrated with maps, Apollo autonomous driving. Enterprise API widely adopted by foreign companies in China.
Strong multilingual performance, including Japanese. Qwen2.5 (2025) and Qwen3 (2026) released. Backed by Alibaba Cloud's infrastructure for enterprise deployments.
Successor to R1. World-class reasoning at industry-low inference cost. API adoption surging globally among developers and enterprises seeking GPT-4 alternatives.
Powers TikTok/Douyin recommendation AI. Specializes in long-context, video understanding, and live commerce integration. Largest Chinese consumer AI user base.
Leader in ultra-long context (up to 2M tokens). Preferred for legal document analysis and research papers. Accelerating global expansion in 2026.
Integrated into the WeChat ecosystem. Reaches 500M+ users via mini-programs and WeCom (企业微信). Strong for internal enterprise workflows in China.
Domain-specific models for healthcare, finance, and manufacturing running on Ascend chips. Fully GPU-independent stack. Dominant in government and SOE contracts.
Government-backed academic model. Top scores on STEM benchmarks. Serves as the reference open-source implementation for China's national AI strategy.
3. The Chip Race — Innovating Under Sanctions
With US export controls effectively banning Nvidia H100, A100, and H20 for China, Beijing has elevated domestic chip development to a national strategic imperative. The landscape is evolving rapidly:
| Chip | Maker | Performance | Key Points |
|---|---|---|---|
| Ascend 910C/D | Huawei | ~60–80% of H100 | Mass production began 2025. Manufactured by SMIC (7nm). Adopted by Baidu, Alibaba, Tencent cloud services. Government and SOE preferred choice. |
| Ascend 910B | Huawei | A100-class | Previous generation with established deployment track record in Chinese cloud AI services. |
| Suiyuan T20 | Enflame (燧原科技) | — | Tencent-backed. Optimized for AI inference at scale in cloud environments with focus on cost efficiency. |
| MLU580 | Cambricon (寒武纪) | — | Only listed AI chip pure-play on A-shares. Full product line from edge to cloud. Widely deployed in China's smart city infrastructure. |
| MTT S4000 | Moore Threads (摩尔线程) | — | Proprietary MUSA architecture. Targeting both gaming GPU and AI training market. Growing developer ecosystem. |
Huawei's Ascend 910C entered mass production in 2025 and is now being deployed in major Chinese cloud platforms. The performance gap with Nvidia is narrowing, and some analysts project domestic chips will supply over 50% of China's AI training demand by 2027.
4. The Regulatory Framework — 3 Rules Global Companies Must Know
China has led the world in AI-specific regulation since 2021. Companies using AI in Chinese operations need to understand these three frameworks:
5. AI Application Hotspots — 5 Industries to Watch
Manufacturing & Robotics
The fusion of AI and humanoid robotics is accelerating dramatically. Unitree Robotics, Fourier Intelligence, and UBTECH are moving into factory-floor production phases. Foxconn, BYD, and BOE have already begun pilot deployments. Full-scale factory integration is expected to become mainstream in 2026.
Healthcare & Drug Discovery
AI applications are advancing across tumor diagnostics, drug discovery, and telemedicine. Baidu has expanded AI-assisted CT interpretation partnerships with top-tier hospitals. Chinese biotechs like Insilico Medicine and XtalPi are applying AlphaFold successors to structure-based drug design — creating partnership opportunities for Japanese pharma companies.
Financial Services
Nearly every major Chinese bank, insurer, and securities firm has deployed "financial LLMs." Ping An's OneConnect and CMB's "Capricorn AI" are leading examples. Risk assessment, fraud detection, and customer service automation are the primary use cases.
Education
After the 2021 crackdown on private tutoring, AI-powered learning is returning as a B2B product for schools. TAL Education (好未来) and Yuanfudao (猿辅导) are pivoting to institution-facing personalized learning platforms.
Autonomous Vehicles
Baidu Apollo has accumulated 30M+ km of autonomous driving data in Wuhan and Beijing. Robotaxi commercial deployment is expanding to new cities in 2026, creating competitive pressure — and potential partnership opportunities — for Japanese automakers.
6. Six Business Opportunities for Global Companies
7. Five Actions to Take Now
- 1 Run an internal AI literacy sprint: Have your China-based teams trial major Chinese AI tools (Kimi, Doubao, ERNIE Bot, Qwen). Set a target of practical daily use within 90 days.
- 2 Conduct a regulatory mapping exercise: Audit AI systems currently in use or under evaluation against the Generative AI Measures and Algorithm Regulation. Involve your legal team.
- 3 Screen Chinese AI partner candidates: Identify Chinese AI companies active in your industry vertical and explore POC (proof of concept) opportunities.
- 4 Benchmark DeepSeek and Qwen for internal tasks: API costs are 1/5–1/10th of GPT-4. Test on meeting transcription, translation, code review, and document summarization.
- 5 Build an AI risk policy for China operations: Define what internal data can/cannot be fed into AI systems. Create a China-specific version aligned with PIPL and local AI regulations.