Last reviewed
May 2, 2026
Sources
25 citations
Review status
Source-backed
Revision
v1 ยท 6,265 words
Add missing citations, update stale details, or suggest a clearer explanation.
Artificial intelligence in China refers to the research, development, deployment, and governance of artificial intelligence within the People's Republic of China. By the mid-2020s China was the world's second largest national AI ecosystem after the United States, measured by published research output, private investment, the number of large language models in production, and the size of the domestic AI hardware industry. The Chinese government has treated AI as a national strategic priority since at least 2017, when the State Council issued the New Generation Artificial Intelligence Development Plan, and has matched the rhetoric with industrial policy, procurement, regulation, and large public funds.
The Chinese AI sector spans the major internet platforms (Baidu, Alibaba, Tencent, ByteDance), telecom and hardware giants (Huawei, Xiaomi), a cohort of well funded large language model startups (the so called "AI tigers" including DeepSeek, Moonshot AI, MiniMax, Zhipu AI, 01.AI, Baichuan), specialist computer vision firms (SenseTime, Megvii, Yitu, CloudWalk), domestic chip designers (Cambricon, Biren, Moore Threads, Huawei's HiSilicon), and major academic centres at Tsinghua University, Peking University, the Chinese Academy of Sciences, and the Beijing Academy of Artificial Intelligence (BAAI). The release of DeepSeek's R1 reasoning model in January 2025 was widely seen as the first time a Chinese open weights model directly challenged the leading United States foundation models on cost and capability.
The ecosystem operates under tight constraints. Since October 2022 the United States has progressively restricted exports of advanced AI chips and the equipment needed to make them, cutting Chinese firms off from the Nvidia H100 and most successor parts. Chinese labs have responded with a combination of stockpiling, smuggling, software optimisation, and a domestic accelerator program led by Huawei's Ascend line. Domestically, generative AI is regulated under a series of measures issued by the Cyberspace Administration of China (CAC) covering deep synthesis content, recommendation algorithms, and large language model services, all of which require pre release security reviews and registration with the algorithm registry.
AI featured in Chinese national planning before 2017, but mainly as a sub component of broader information technology and robotics initiatives. The State Council's Made in China 2025 plan, issued in May 2015, named ten priority sectors including "new generation information technology" and "high end CNC machine tools and robotics," and committed the state to raising domestic content and self sufficiency in core components. The 13th Five Year Plan (2016 to 2020) repeated AI and machine learning as research priorities. The 2016 "Internet Plus AI Three Year Implementation Plan" jointly issued by the National Development and Reform Commission, Ministry of Science and Technology, Ministry of Industry and Information Technology, and the Cyberspace Administration set initial targets for AI applications in manufacturing and consumer services.
The central document is the State Council's New Generation Artificial Intelligence Development Plan, issued on July 8, 2017 (Guofa No. 35). The plan set a three step timetable: by 2020, the AI industry should be "in step" with the world leaders; by 2025, China should achieve major breakthroughs in basic AI theory and lead in some technologies; by 2030, China should be the world's primary AI innovation centre, with a core AI industry valued at 1 trillion yuan and a related industries scale of 10 trillion yuan. The plan named six key tasks, including basic theory, key technologies (vision, speech, natural language, swarm intelligence), and infrastructure, and listed thirteen national open innovation platforms tied to specific companies (autonomous driving was assigned to Baidu, smart cities to Alibaba, medical imaging to Tencent, voice to iFlytek, and so on).
A July 2017 translation by Graham Webster, Rogier Creemers, Paul Triolo, and Elsa Kania, published by New America, became the most widely cited English version of the plan and shaped how Western policy circles read Chinese AI ambitions. The plan also explicitly identified military civil fusion (junmin ronghe) as a core principle, formalising a channel between civilian AI research and the defence sector.
China was among the first jurisdictions to issue binding rules on generative AI. Three regulations stand out:
The algorithm registry, hosted by the CAC, lists registered services and their providers. By the end of 2024 the public list included several hundred entries covering the major Chinese LLMs and recommender systems. A separate "large model filing" track was introduced in 2023 specifically for generative AI services and grants permission to make services available to the general public; by mid 2024 more than 180 large models had passed this filing process.
In September 2024 China's National Technical Committee 260 on Cybersecurity, in coordination with the CAC, released version 1.0 of the AI Safety Governance Framework. The framework is voluntary in form but signals the official Chinese position on AI risk taxonomy. It groups risks into endogenous safety risks (issues with the models themselves, such as hallucination, bias, and adversarial robustness) and application safety risks (misuse for fraud, disinformation, cyberattacks, or critical infrastructure failure), and lists corresponding technical and governance countermeasures. China also signed the Bletchley Declaration at the November 2023 UK AI Safety Summit and the Seoul Declaration at the May 2024 follow up summit.
The 14th Five Year Plan (2021 to 2025), adopted at the National People's Congress in March 2021, raised AI to the level of one of seven "frontier sciences and technologies" alongside quantum information, integrated circuits, brain science, genetics and biotechnology, clinical medicine and health, and deep space, deep sea, and polar exploration. Beijing, Shanghai, Shenzhen, Hangzhou, Hefei, and Chengdu issued local AI development plans with their own targets and subsidies, and many established AI industrial parks. Shanghai's Zhangjiang AI Island and Beijing's Zhongguancun district hosted concentrations of AI startups and research centres.
Chinese AI activity is heavily concentrated in a small number of companies and research institutes. The table below lists the most prominent organisations as of 2025.
| Organisation | Type | HQ | Founded | Flagship AI assets |
|---|---|---|---|---|
| Baidu | Internet platform | Beijing | 2000 | ERNIE LLM family, Apollo autonomous driving, Kunlun AI chips |
| Alibaba | E commerce, cloud | Hangzhou | 1999 | Qwen LLM family, Alibaba Cloud, T head Yitian server CPU |
| Tencent | Internet, gaming | Shenzhen | 1998 | Hunyuan LLM, Tencent Cloud, WeChat AI features |
| ByteDance | Internet, short video | Beijing | 2012 | Doubao LLM, Seed (formerly Skylark) research lab, Volcano Engine |
| Huawei | Telecom, hardware | Shenzhen | 1987 | Pangu LLM, Ascend AI accelerators, MindSpore framework |
| Xiaomi | Consumer electronics | Beijing | 2010 | MiMo LLM, MiLM, in house automotive AI |
| iFlytek | Speech, education | Hefei | 1999 | Spark (Xinghuo) LLM, voice and translation |
| SenseTime | Computer vision | Hong Kong, Shanghai | 2014 | SenseNova LLM, computer vision platforms |
| Megvii | Computer vision | Beijing | 2011 | Face++ vision platform |
| Yitu Technology | Computer vision | Shanghai | 2012 | Vision AI for healthcare and security |
| CloudWalk | Computer vision | Guangzhou | 2015 | Vision AI, public sector deployments |
| DeepSeek | LLM startup | Hangzhou | 2023 | DeepSeek V2, V3, R1 |
| Moonshot AI | LLM startup | Beijing | 2023 | Kimi assistant, Kimi K1.5 |
| MiniMax | LLM startup | Shanghai | 2021 | abab and MiniMax M1 models, Talkie |
| Zhipu AI | LLM startup | Beijing | 2019 | GLM and ChatGLM model family |
| 01.AI | LLM startup | Beijing | 2023 | Yi model family |
| Baichuan AI | LLM startup | Beijing | 2023 | Baichuan model family |
| BAAI | Public research | Beijing | 2018 | Wudao series, FlagOpen open source stack |
The "AI tiger" label was popularised in Chinese tech media in 2023 to describe the cohort of well funded LLM startups: Zhipu, Baichuan, Moonshot, MiniMax, 01.AI, and (later) Stepfun. DeepSeek, founded by High Flyer hedge fund founder Liang Wenfeng, sat outside this group at first but became the most internationally visible Chinese AI lab in 2025.
Baidu was the earliest of the Chinese internet majors to invest seriously in deep learning. It hired Andrew Ng to lead its Silicon Valley AI lab from 2014 to 2017, established the Institute of Deep Learning, and released its first version of the ERNIE (Enhanced Representation through Knowledge Integration) language model in 2019. ERNIE Bot (Wenxin Yiyan in Chinese) launched on March 16, 2023, making Baidu the first Chinese company to publicly release a ChatGPT style assistant. Baidu's Apollo autonomous driving program, launched in April 2017, operates a fleet of robotaxis under the Apollo Go brand in cities including Wuhan, Beijing, and Chongqing. Baidu also designs the Kunlun AI accelerator, manufactured initially by Samsung on 14 nm and later by SMIC on more advanced nodes.
Alibaba's DAMO Academy, established in 2017, became the parent of much of the company's AI research. The Qwen (Tongyi Qianwen) family of large language models launched in April 2023 and rapidly became the most widely used open weights model series of Chinese origin. Alibaba released a series of versions including Qwen, Qwen 1.5, Qwen 2, Qwen 2.5, and Qwen 3, with sizes ranging from 0.5 billion to several hundred billion parameters and dense and mixture of experts variants. Qwen 2.5 Coder topped open source coding leaderboards in late 2024. Qwen models are released under the Apache 2.0 or Tongyi licence and are available on Hugging Face and ModelScope, Alibaba's own model hub. Alibaba Cloud is the largest Chinese cloud provider and the primary serving infrastructure for Qwen.
Tencent's Hunyuan family includes text, image, and 3D generation models. Hunyuan Large, released in November 2024, is a 389 billion parameter mixture of experts model with 52 billion active parameters. Hunyuan Video, an open source text to video model released in December 2024, was widely benchmarked against Sora and Kling. Tencent uses Hunyuan internally to power features in WeChat, QQ, and Tencent Meeting.
ByteDance, the parent of Douyin and TikTok, launched its Doubao consumer chatbot in August 2023 and quickly became the largest Chinese consumer assistant by monthly active users, surpassing 70 million MAU by late 2024 according to Aicpb. ByteDance's research lab, formerly Skylark and now branded as ByteDance Seed, develops the Doubao model series. Volcano Engine is the company's enterprise cloud arm and offers Doubao through API.
Huawei plays a dual role as both a model and chip maker. The Pangu model series, first announced in 2021, includes Pangu Alpha (Chinese language LLM), Pangu CV, Pangu Sigma (a sparse model), and a Pangu Weather model that the company published in Nature in July 2023, claiming superior medium range forecasting performance versus the European Centre for Medium Range Weather Forecasts. Huawei's Ascend AI accelerators (310, 910, 910B, 910C) anchor Chinese efforts to substitute for Nvidia GPUs and are paired with the open source MindSpore deep learning framework, an alternative to PyTorch.
Zhipu AI, spun out of Tsinghua University's Knowledge Engineering Lab in 2019, develops the GLM and ChatGLM models, including GLM 4 (released January 2024) and the multimodal CogVLM. Zhipu raised significant funding from Alibaba, Tencent, Saudi Aramco's Prosperity7 Ventures, and others. The United States Department of Commerce added Zhipu and several subsidiaries to the Entity List in January 2025, citing concerns over support for Chinese military modernisation.
Moonshot AI, founded by Yang Zhilin in March 2023, became known for the long context Kimi chat assistant, which marketed a 200,000 character context window in October 2023 and expanded to 2 million characters in 2024. Moonshot raised about 1 billion USD from Alibaba and others in early 2024. The Kimi K1.5 model, released January 2025, focused on reasoning and was benchmarked against OpenAI's o1 series.
MiniMax, founded by Yan Junjie in late 2021, released the abab model series and the MiniMax M1 reasoning model in 2025. The company also operates Talkie, a global character chat app.
01.AI, founded by Kai Fu Lee in March 2023, released the Yi model family. Yi 34B, released November 2023, was briefly the highest scoring open weights non Chinese language model on the Hugging Face Open LLM Leaderboard. The company restructured its pre training operations into a joint venture with Alibaba in late 2024 amid the cost pressures of frontier training.
Baichuan AI, founded by former Sogou CEO Wang Xiaochuan in April 2023, released the Baichuan series and focused increasingly on medical applications.
DeepSeek is the AI lab spun out of High Flyer, a quantitative hedge fund founded by Liang Wenfeng in 2015. High Flyer purchased thousands of Nvidia A100 GPUs in 2021 to 2022, before the United States export controls cut off direct access, giving DeepSeek an unusually large compute base for an independent lab. DeepSeek released DeepSeek LLM in late 2023, DeepSeek V2 in May 2024 with a novel multi head latent attention architecture, DeepSeek V3 in December 2024 (a 671 billion parameter mixture of experts model with 37 billion active parameters that the company reported was trained for about 5.6 million USD in compute), and the DeepSeek R1 reasoning model in January 2025. R1's release, with weights and a permissive licence, was widely covered in international media in late January 2025 and was associated with a sharp single day decline in Nvidia's share price on January 27, 2025.
Chinese AI hardware splits into three layers: GPU and AI accelerator design, fabrication, and high bandwidth memory.
Semiconductor Manufacturing International Corporation (SMIC) is the largest Chinese contract chip maker. SMIC achieved volume 14 nm production in 2019 and demonstrated 7 nm class production in 2023, when the Huawei Mate 60 Pro shipped with the Kirin 9000S system on chip manufactured by SMIC using deep ultraviolet lithography rather than the extreme ultraviolet lithography that TSMC uses for equivalent nodes. ASML has been blocked from selling extreme ultraviolet equipment to China since 2019 and from selling its most advanced deep ultraviolet machines (NXT 2000i and successors) since 2024 under combined US and Dutch controls. Other Chinese foundries include Hua Hong Semiconductor and Yangtze Memory Technologies; Yangtze Memory was added to the Entity List in December 2022.
High bandwidth memory is a binding constraint on Chinese AI accelerators. As of 2025 the world's HBM supply was effectively controlled by SK Hynix, Samsung, and Micron. Chinese efforts to indigenise HBM are led by ChangXin Memory Technologies (CXMT) and Wuhan Xinxin (XMC, a Yangtze Memory subsidiary), but Chinese HBM remained one or two generations behind the international leaders.
The dominant Chinese deep learning frameworks are Baidu's PaddlePaddle (open sourced 2016), Huawei's MindSpore (open sourced 2020), and Alibaba's MNN. Chinese researchers also use international frameworks: PyTorch is the standard at the major LLM labs.
The leading Chinese university AI programs are at Tsinghua University, Peking University, the Chinese Academy of Sciences (especially the Institute of Automation and the Institute of Computing Technology), Shanghai Jiao Tong University, the University of Science and Technology of China, Zhejiang University, Fudan University, Nanjing University, Westlake University (founded 2018 in Hangzhou), and the Hong Kong University of Science and Technology. Tsinghua's THUDM lab developed the GLM and CogVLM models that anchor Zhipu's product line, and many of the founders of the LLM startups are Tsinghua alumni: Yang Zhilin (Moonshot), Zhang Yaqin and Tang Jie (Zhipu), Wang Xiaochuan (Baichuan), and others.
The Beijing Academy of Artificial Intelligence (BAAI), established in November 2018 with backing from the Beijing municipal government, was the first non corporate Chinese AI institute to train a frontier scale model: the Wudao 1.0 (March 2021) and Wudao 2.0 (June 2021) models. Wudao 2.0 was reported to have 1.75 trillion parameters, although the architecture was a sparse mixture of experts and the model was never made widely available outside BAAI's collaborators.
The Stanford AI Index reports show consistent Chinese leadership on quantitative measures of AI research. The 2025 AI Index, citing OpenAlex data, reported that China produced the largest share of AI publications in 2023 (23.2 percent of the global total), more than the United States (15.2 percent) and the European Union combined; on highly cited papers and patent filings China was also the leader. On talent flow, however, the same reports show that the United States retained the largest share of AI researchers working at frontier labs and the largest gross inflow of senior AI researchers, with China a net exporter of AI doctoral graduates.
The Center for Security and Emerging Technology (CSET) at Georgetown has tracked Chinese AI talent since 2019. CSET researchers including Remco Zwetsloot and Dahlia Peterson have documented that a large share of senior researchers at top Chinese AI labs received their doctorates in the United States, and that this pattern began to shift after 2020 as more Chinese students completed their degrees domestically.
The ChinAI Newsletter, written since 2018 by Jeffrey Ding, then a doctoral student at Oxford and now an assistant professor at George Washington University, has translated Chinese language AI policy and industry writing into English and is widely cited as a primary source on Chinese AI debates.
Chinese labs have produced a continuous stream of large language models since 2021. The list below covers the most prominent post 2023 releases.
| Model | Developer | Released | Parameters (active) | Notes |
|---|---|---|---|---|
| Wudao 2.0 | BAAI | June 2021 | 1.75T (sparse MoE) | Early Chinese frontier scale claim, not widely available |
| ERNIE Bot | Baidu | March 2023 | undisclosed | First Chinese ChatGPT analogue; ERNIE 4.0 released October 2023 |
| ChatGLM 6B | Zhipu / Tsinghua | March 2023 | 6B dense | Open weights, popular fine tuning base |
| Spark (Xinghuo) | iFlytek | May 2023 | undisclosed | Speech focused assistant |
| Qwen | Alibaba | April 2023 | 7B to 72B dense | Open weights under Tongyi licence |
| Qwen 2 / 2.5 | Alibaba | June 2024 / Sep 2024 | 0.5B to 72B + MoE | Open weights, strong on coding |
| Qwen 3 | Alibaba | 2025 | up to ~235B MoE | Open weights, multilingual |
| Yi 34B | 01.AI | November 2023 | 34B dense | Open weights |
| Baichuan 2 | Baichuan AI | September 2023 | 7B and 13B dense | Open weights |
| Hunyuan | Tencent | September 2023 | undisclosed | Powers WeChat features |
| Hunyuan Large | Tencent | November 2024 | 389B (52B active) MoE | Open weights |
| Pangu Sigma | Huawei | 2023 | 1.085T sparse | Internal use, paper only |
| Doubao | ByteDance | August 2023 | undisclosed | Largest Chinese consumer chatbot by MAU |
| Kimi (Moonshot v1) | Moonshot AI | October 2023 | undisclosed | 200K then 2M character context |
| Kimi K1.5 | Moonshot AI | January 2025 | undisclosed | Reasoning model |
| MiniMax abab 6.5 | MiniMax | April 2024 | undisclosed | 245k context |
| MiniMax M1 | MiniMax | June 2025 | 456B (45.9B active) MoE | Open weights, lightning attention |
| GLM 4 | Zhipu | January 2024 | undisclosed | Closed; ChatGLM 4 9B open weights |
| DeepSeek V2 | DeepSeek | May 2024 | 236B (21B active) MoE | Open weights, multi head latent attention |
| DeepSeek V3 | DeepSeek | December 2024 | 671B (37B active) MoE | Open weights, reported ~5.6M USD training run |
| DeepSeek R1 | DeepSeek | January 2025 | 671B (37B active) MoE | Open weights reasoning model |
| MiMo | Xiaomi | April 2025 | 7B dense | Open weights, reasoning oriented |
Most Chinese open weights models are distributed via Hugging Face and Alibaba's ModelScope under permissive licences. The DeepSeek V3 and R1 weights are released under the MIT licence, which made them attractive to international researchers and downstream developers.
On standard benchmarks (MMLU, GSM8K, MATH, HumanEval, SWE Bench, GPQA), the strongest Chinese models in 2024 to 2025 ranged from comparable to within a few months of the leading United States closed models, depending on the benchmark. Independent evaluations by Epoch AI and the LMSYS Chatbot Arena placed Qwen 2.5, DeepSeek V3, and DeepSeek R1 in the top tier of open weights models internationally.
China's AI regulation operates within the broader internet governance regime built around the 2017 Cybersecurity Law, the 2021 Data Security Law, and the 2021 Personal Information Protection Law. Generative AI services that are made available to the Chinese public must:
The security assessment includes review of training data sources and content moderation. Models must avoid generating content that violates the existing internet content rules, which include categories such as content that endangers national security, undermines national unity, or harms social order. In practice the Chinese consumer LLMs decline to discuss a defined set of politically sensitive topics, and the algorithm registry provides a record of the providers and their stated training data and use cases.
Chinese rules treat foundation model providers and downstream service providers differently. The 2023 Interim Measures explicitly target services made available to the Chinese public; pure model research, internal enterprise use, and overseas deployment are subject to lighter requirements. This distinction has allowed Chinese labs to maintain active publication and open weights release programs while still complying with the consumer service rules.
In March 2025 the Cyberspace Administration of China issued the Measures for Labelling AI Generated Content (effective September 1, 2025), requiring both visible labels and metadata watermarking on AI generated text, image, audio, and video output.
The National People's Congress Standing Committee included an Artificial Intelligence Law on its 2024 legislative agenda, but as of mid 2025 no comprehensive AI law had been adopted; observers expected a draft to be published for consultation before the end of the 14th Five Year Plan in 2025.
The most significant external constraint on Chinese AI is United States export control of advanced semiconductors and the equipment to make them. The Bureau of Industry and Security at the Department of Commerce (export controls) has issued a series of measures:
The practical effect on Chinese AI compute has been substantial. Chinese cloud providers were unable to legally purchase H100, H200, B100, or B200 GPUs after October 2023. Nvidia's H20, a chip designed to fall just below the United States thresholds, became the dominant Nvidia part shipped to China in 2024; its export was further restricted in April 2025. Chinese firms responded with: (1) accelerated procurement of domestic accelerators, especially Huawei Ascend; (2) algorithmic and systems optimisation, exemplified by DeepSeek's reported low cost training runs; (3) rental of overseas compute, including some grey market access to United States cloud regions; and (4) reported smuggling, documented in Reuters and Wall Street Journal reporting through 2024 and 2025.
The Entity List, maintained by BIS, has expanded to cover most of the prominent Chinese AI hardware companies and several model developers. Additions include SenseTime (October 2019), iFlytek (October 2019), Megvii (October 2019), CloudWalk (May 2020), Yitu (May 2020), Cambricon (December 2022), Moore Threads and Biren (October 2023), and Zhipu and several subsidiaries (January 2025).
Secondary effects of the controls include: the bifurcation of the global AI software stack (CUDA dominant outside China, an emerging Ascend and CANN stack inside China); the growth of Chinese model exports to third countries through open weights distribution; and a domestic policy push to accelerate self sufficiency in semiconductor manufacturing, organised around the China Integrated Circuit Industry Investment Fund (the so called "Big Fund"), which raised its third tranche of 344 billion yuan (about 47 billion USD) in May 2024.
Private investment in Chinese AI grew rapidly from 2018 to 2021, contracted in 2022 and 2023 with the broader Chinese tech downturn, and rebounded in 2024 alongside the global generative AI boom. According to the Stanford AI Index 2025 report, China attracted about 7.6 billion USD of private AI investment in 2024, second to the United States at 109 billion USD. Earlier years were closer: China was about 13 billion USD in 2021 versus 53 billion in the United States. The CB Insights and PitchBook trackers show similar gaps with different absolute numbers. Industry analysts attribute the relative slowdown in Chinese private AI funding after 2021 to a combination of regulatory tightening on internet platforms (the 2020 to 2022 platform rectification campaign), domestic capital market headwinds, and the chilling effect of United States outbound investment screening on cross border capital flows.
Government guidance funds and state owned investors play a larger role in Chinese AI investment than in the United States. The most prominent are the China Internet Investment Fund (operated by the Cyberspace Administration), the Big Fund (Integrated Circuit Industry Investment Fund), and various municipal funds in Beijing, Shanghai, Shenzhen, and Hefei. State backed investors took stakes in 1 percent of Zhipu, Baichuan, Moonshot, MiniMax, and 01.AI in 2024 and 2025 in exchange for special management share rights, a structure first reported by the Financial Times.
Mergers and acquisitions of independent Chinese AI labs accelerated in 2024 and 2025 as the cost of frontier training rose. The 01.AI pre training joint venture with Alibaba (announced December 2024 and finalised in early 2025), the Baichuan partnership with Tencent, and several smaller transactions consolidated frontier scale training inside the major cloud providers.
The release of DeepSeek V3 on December 26, 2024 and DeepSeek R1 on January 20, 2025 marked the first widely recognised moment when an independent Chinese lab matched the leading United States closed models on a major capability dimension while making weights and methodology public. DeepSeek V3 reported a training compute cost of 2.788 million GPU hours on Nvidia H800 chips, equivalent to about 5.576 million USD at the lab's stated 2 USD per GPU hour rate. The accompanying technical report described an architecture combining mixture of experts with multi head latent attention, FP8 mixed precision training, and a custom load balancing strategy. The R1 reasoning model used reinforcement learning from rule based rewards on top of V3 to elicit chain of thought style reasoning.
The global response in late January 2025 was unusually visible. The DeepSeek consumer app reached the top of the Apple App Store free charts in the United States on January 27, 2025. Nvidia's share price fell about 17 percent that day, wiping nearly 600 billion USD of market capitalisation. Coverage in the Financial Times, Reuters, Wall Street Journal, and Bloomberg framed R1 as a turning point in the assumed United States lead, although researchers including Epoch AI cautioned that the reported training cost excluded earlier infrastructure investment, that DeepSeek's compute base of perhaps 50,000 H800 and earlier H100 and A100 GPUs was substantial in absolute terms, and that the gap on inference cost and many benchmarks reopened with later United States releases.
DeepSeek's broader contribution to the Chinese AI sector was twofold. It demonstrated that an independent lab could compete at the frontier without the resources of a major internet platform, encouraging a wave of new Chinese AI startups in 2025. It also entrenched the Chinese sector's preference for open weights distribution, a pattern that distinguishes the Chinese frontier from the largely closed United States frontier and that continues with Qwen 3, Hunyuan, and others.
Chinese AI capacity has been deployed in domains beyond consumer chatbots. Computer vision and facial recognition systems from SenseTime, Megvii, Yitu, CloudWalk, Hikvision, and Dahua are used widely by Chinese police and municipal governments. Independent research, including the work of the Australian Strategic Policy Institute, Human Rights Watch, and academics such as Darren Byler, has documented the use of these systems in mass surveillance programs in Xinjiang. The United States Entity List additions of these companies in 2019 and 2020 cited human rights grounds.
Autonomous driving is a major Chinese AI application. Baidu Apollo, Pony.ai, WeRide, AutoX, Didi Autonomous Driving, and Huawei's intelligent driving solutions all operate testing fleets, and several have launched commercial robotaxi services in restricted operational design domains. Chinese electric vehicle makers (BYD, Nio, XPeng, Li Auto, Zeekr) deploy advanced driver assistance systems with increasingly large neural networks; XPeng's XNGP and Huawei's ADS 3.0 use end to end learned planners derived from Tesla style architectures.
In robotics, Unitree, EngineAI, Robot Era, Booster Robotics, AgiBot, and other firms produce humanoid robots and quadrupeds. Unitree's H1 humanoid (announced August 2023) and G1 humanoid (announced May 2024) are sold to research and industrial customers. Chinese central and provincial governments announced humanoid robot industrial development plans in late 2023 and 2024 setting targets for 2025 and 2027.
In drug discovery, the AlphaFold family from Google DeepMind has been complemented by Chinese efforts at companies including XtalPi and Insilico Medicine, with public sector work at the Beijing Genomics Institute and the Chinese Academy of Sciences.
Chinese AI development is embedded in a larger competitive relationship between China and the United States. The 2017 development plan was widely interpreted in Washington as evidence of strategic intent, and prompted a series of United States measures including the 2018 Foreign Investment Risk Review Modernization Act, the establishment of the Joint AI Center at the Department of Defense, the 2020 American AI Initiative executive order, the export controls described above, and the October 30, 2023 Executive Order on AI by the Biden administration. The Trump administration's January 23, 2025 Executive Order on Removing Barriers to American Leadership in Artificial Intelligence revoked parts of the 2023 order while continuing the export control framework.
At the multilateral level China signed the 2023 Bletchley Declaration, the 2024 Seoul Declaration, and supported the 2024 United Nations General Assembly resolution on AI. China has also pursued bilateral AI dialogues with the United States (formal track 1 talks were held in Geneva in May 2024 and the autumn of 2024) and the European Union, alongside its own outreach through the Belt and Road Initiative and the Global AI Governance Initiative announced by Xi Jinping in October 2023.
The Chinese position in international AI governance forums has emphasised state sovereignty over data and algorithms, opposition to United States led restrictions on technology transfer, and support for inclusive development for the Global South. United States officials and several Western governments have argued that Chinese AI is enabling state surveillance and military modernisation. Both framings shape policy decisions, and the underlying technical trajectory continues largely independent of either rhetorical position.