When DeepSeek will break onto the world stage in January 2025, it seemed to come out of nowhere. However, the Huge Language Model was just one of thousands of generative AI tools that have been released in China since 2023, and each of them has a public archive.
The country’s top internet regulator, the Cyberspace Administration of China (CAC), requires any company launching an artificial intelligence tool with “public opinion properties or social mobilization capabilities” to first list it in a public database: the Algorithm Registry. In their declaration, developers must demonstrate how their products avoid 31 categories of risk, ranging from age and gender discrimination to psychological harm to “violation of fundamental socialist values.”
Applicants submit their applications to the local CAC (say, the Shanghai CAC for companies registered in Shanghai), which forwards the applications to the central CAC for final approval. Only then is the tool publicly entered into the algorithm register. Matt Sheehan, a researcher at the Carnegie Endowment for International Peace, notes that while the European Union is working to adopt a single, comprehensive law on artificial intelligence, China’s approach to regulation is more ad hoc, focusing on specific algorithms and developing iterative standards. (There is no comparable registration system or centralized regulatory agency in the US.)
Over time, CAC has inadvertently created the most detailed map of the domestic AI ecosystem in the world.
*Data current as of April 2025 includes both ‘Generative AI’ and ‘Deep Synthesis’ algorithms
Open the August 2024 CAC update and you’ll see DeepSeek listed as entry 152, a single row in a neatly packed table. Scroll through the table and you’ll find the Homestays AI and the Management AI designs patents. One helps obstetricians-gynecologists in a maternity ward in Shanghai; another helps manage the state’s power grids. Kendra Schaefer and her colleagues at Trivium China, a Beijing-based consulting firm, compile the CAC updates into a comprehensive database, augmented by their own research.
A broad look at the boom
Nearly 80 percent of China’s generative AI registrations are concentrated in and around key technology hubs – Beijing, Shenzhen, Shanghai and Hangzhou. Each city has its strengths: Beijing elite universities, national laboratories, and political clout give it an advantage for large-scale innovation; Shenzhen (in Guangdong) is home to a dense hardware supply chain and a huge engineering talent pool; Shanghai, located close to multinational corporations, is at the forefront of commercialization; and Hangzhou (in Zhejiang) is powered by Alibaba’s e-commerce empire.
But innovation is spreading far beyond the coasts. Chongqing positions itself as an AI production and logistics hub; and thanks to ponderous state investment, Hefei in Anhui province has gained the nickname “China’s speech valley” due to its cluster of speech recognition companies, including iFlyTek. Submissions also come from less obvious regions such as Guizhou, China’s “Big Data Valley,” where massive data centers power Huawei’s Pangu model, and Inner Mongolia, where state-owned enterprises are integrating artificial intelligence in mining and agriculture.
*Data current as of April 2025
In the Trivium dataset, state-related entries – from state-owned enterprises to government-backed research institutes – account for 22 percent of submissions. Many state-linked companies are working with Huge Tech to build artificial intelligence: PetroChina, for example joined forces with Huawei and iFlyTek to develop oil and gas applications; State network I used DeepSeek building a model optimizing power grids.
Foreign companies constitute only 0.5 percent. reports. Ikea, for example, has an bright shopping algorithm that generates product recommendations. Yum China, the parent company that operates Kentucky Fried Chicken in China, has listed a model that generates menus and promotional materials.
Zeroing in on the competition
*Data current as of April 2025
More than half of the entries in the algorithm registry concern what Schaefer calls cross-sector technologies. These include basic models, “general purpose” text generators and a wide range of multimedia tools – voice changers, 3D renderers, image creation tools. “No one wants to be in a situation where they rely on a competitor’s technology,” Schaefer says. Unlike the United States, where OpenAI, Anthropic and Google DeepMind dominate the market, China’s competition in building basic AI remains diverse and contentious. However, these models are costly to build and the market is starting to consolidate. “Chinese Six”AI Tigers”-Moonshot, Minimax, Zhipu, Baichuan, 0.1AI and Stepfun are all backed by Alibaba or Tencent. ByteDance’s Doubao has overtaken DeepSeek as the most popular chatbot in China, but its place at the top is uncertain.
Niche natives
While giants fight for chatbot supremacy, startups are strenuous at work in every sector imaginable.
Squirrel AI Squirrel
Founder Derek Li says his 12-year-old company is ahead of the curve in information technology. “They put wheels on the horse,” he says, adapting artificial intelligence to existing, antiquated software. Squirrel claims to diagnose knowledge gaps, measure progress and adapt lessons in real time.
When China banned for-profit tutoring in 2021, the company’s revenue dropped overnight. This rotated to license its platform to franchisees who also sold the company’s AI-equipped tablets. The Squirrel network covers over 3,000 centers across China, serving 1.2 million students. Now the company plans to expand to the United States.
Li, who he withdrew his sons from a private school in Shanghai so they can learn at home on the Squirrel platform, claims that “in the future, teachers will not teach knowledge.” Instead, he says, “they will become data analysts, understanding reports of students’ learning and abilities, and psychologists, understanding emotions and shaping their personalities.”
Kanshe AI language
AI Kanshe (translated to “AI Sees Tongue”) is a conventional Chinese medicine startup that analyzes health through images of the tongue, hand and face. The company was founded by Li Wenhua, a former employee of Yaoshi Bang, one of China’s first online pharmaceutical platforms. Li, a long-time student of tongue and hand diagnostics, wanted to combine the diagnostic methods of conventional Chinese medicine with newfangled machine vision. The company serves both consumers and healthcare professionals in clinics, pharmacies and some hospitals by offering tools to support diagnosis and decision-making. Its model is trained on over 100,000 annotated images of tongues, hands and faces.
Zhongtan Puhui Cloud Technology Zhongtan Puhui Cloud Technology
Founded in 2024 by Wu Song, a former Wall Street quantitative trader, Zhongtan Puhui Cloud Technology develops artificial intelligence-based tools for carbon accounting. Wu argues that the ecological transition still relies on burdensome human labor that can be automated.
Zhongtan Puhui creates AI agents that handle a range of carbon accounting tasks, including carbon footprint tracking and emissions audits. Its clients include China Minmetals Group and DHL, as well as diminutive and medium-sized exporters from the Yangtze River Delta.