Inside China’s AI Registry
China’s Cyberspace Administration (CAC) maintains the world’s only comprehensive, publicly accessible registry of generative AI tools (GAT). Every public-facing generative AI service—whether text, image, audio, video, or multimodal—must register before deployment. In April, Trivium posted an Excel file that lists all the GATs in this registry. The Excel file captures essential metadata including registration number, approval date, tool name, company name, province, functional description, and business model (B2B or B2C). As of April 2025, it contains 3,739 generative algorithmic tools (GATs) from approximately 2,353 unique companies, growing by 250-300 entries monthly. This provides unprecedented visibility into China’s AI ecosystem that no other country offers.
Support the Gradient Flow Newsletter 
The CAC registration requirement creates both constraints and opportunities. While it adds compliance overhead, it enforces traceability, security audits, and content safety alignment. The system evolved from 2021’s recommendation algorithm provisions, expanded to deep synthesis in 2023, and now covers all generative AI. For teams in China, this means that—because registration is obligatory and the registry doubles as a live census—they must also satisfy strict content-filtering rules (blocking politically sensitive outputs) and undergo tougher, sector-specific audits in areas such as healthcare and finance. However, it also provides competitive intelligence through public metadata and clear compliance frameworks for market entry.
A ‘Massive Scrum’ of Foundational Models
Over 50% of registered tools are foundation models! Unlike Western markets consolidating around OpenAI, Anthropic, or Google’s models, China has hundreds of companies—from tech giants like Alibaba and Baidu to startups and state-owned enterprises (SOEs)—all building proprietary LLMs, image generators, and multimodal systems. Key drivers include: no clear market winner has emerged yet, companies refuse to build on competitors’ technology stacks, government emphasis on technological self-reliance, and the historical pattern of hyper-competition preceding market consolidation in Chinese tech sectors.
This fragmentation creates both challenges and opportunities. On the positive side: diverse foundational models offer multiple integration options, intense competition drives rapid innovation, and niche opportunities exist in underserved verticals. Challenges include: interoperability issues between incompatible systems, risk of partnering with models that may not survive consolidation, and inefficiency from duplicated development efforts. Practitioners should design modular architectures that avoid vendor lock-in and focus on vertical applications where differentiation matters more than foundational capabilities.
The State as Catalyst, Not Creator
SOEs account for approximately 22% of registered tools, but their approach differs significantly from private tech companies. SOEs typically partner with established tech firms rather than developing AI internally. Common patterns include: State Grid partnering with Alibaba for power grid optimization, PetroChina collaborating with tech vendors for oil and gas applications, and China Mobile being the exception—developing genuine in-house AI capabilities. SOEs provide funding, domain-specific data, and large-scale testing environments, while tech partners supply AI expertise and model development.
The state functions as a crucial adoption catalyst through multiple mechanisms: procurement mandates requiring SOEs to integrate AI, providing access to sensitive sector-specific datasets, funding research through institutions like the Chinese Academy of Sciences, and creating protected markets in critical infrastructure where only domestic AI is permitted. For AI teams in China, this means SOE partnerships offer volume, data access, and government support, but require navigating bureaucratic processes and aligning with national priorities.
The Diffusion Thesis: Hard Evidence from the Registry
My previous analysis suggested the true geo-strategic contest in AI lies not in creating elite models but in their widespread economic diffusion. This data from China’s mandatory generative AI registry moves this discussion from hypothesis to hard evidence. The sheer volume—already compounding by roughly 60 new approvals every week—confirms a blistering pace of adoption across the economy, offering a granular view of how AI is being woven into the country’s industrial and consumer fabric.
Diffusion is messy by design: hundreds of rival foundation models jockey for traction, then feed into focused pilots run by utilities, telcos and exporters. That public-private handshake—outlined earlier—converts the registry’s apparent chaos into sector-specific breakthroughs without the state having to build the models itself.
The result is a landscape where AI is less an abstract pursuit of general intelligence and more a tool for tangible, vertical applications. The registry reveals a surge in sector-specific solutions, from AI-powered “smart cockpits” in automobiles to digital tutors in education and diagnostic aids in healthcare. This confirms the initial thesis: while America may lead in frontier research, China’s state-influenced, hyper-competitive, and application-focused ecosystem is accelerating the practical integration of AI into everyday economic life, creating a formidable, if messy, engine for diffusion.
From Scramble to Consolidation: What’s Next for Chinese AI?
Watch for:
- evolution of regulations, particularly around safety assessment and auditability;
- hardware constraints and their impact on model development;
- convergence of AI with China’s manufacturing strengths (AI-enabled devices);
- provincial GPU-supply subsidies that are quietly shifting where fine-tuning and inference work get done.
The Chinese generative AI landscape offers both significant opportunities and unique challenges. Success requires understanding the interplay between state direction, market competition, and technical innovation—while navigating a regulatory environment unlike any other major AI market. The transparency provided by the CAC registry, combined with the rapid pace of development, makes this a critical market for global AI practitioners to understand and potentially engage with.
Speak at Ray Summit 2025
One of my favorite conferences takes place November 3-5 in San Francisco!
- This year’s conference spotlights the critical layers of AI development: open source infrastructure, multimodal data, post-training optimization, and scalable ML platforms, highlighted by a full track dedicated to vLLM.
- This is the definitive gathering for the community of builders, operators, and innovators shaping the future of AI infrastructure.
- Ray is being used by everyone from DeepSeek to x.AI to Thinking Machines to Tencent.
Submit a Talk
The post “Massive Scrum” of Models: New Data on China’s AI Gold Rush appeared first on Gradient Flow.
