13 March 2025
Next month’s Beijing half-marathon will include a separate race run simultaneously by robots. This is China’s AI race at its most literal. But what does the Chinese AI industry look like? Can we peak behind the “red curtain” to find out?
On 06 March, the launch of Manus – the “world’s first fully autonomous AI agent” – sent shockwaves through the global tech industry. Manus uses a multi-agent architecture with a central agent (the “executor”). The executor breaks down tasks into manageable sub-tasks, then assigns each sub-task to the appropriate specialist hub agent, before coordinating the overall response.
Manus is built on Anthropic’s Claude 3.5 Sonnet and versions of Alibaba’s Qwen models. Its architecture is reminiscent of the most recent Chinese AI phenomenon, DeepSeek-R1, which used “mixture of experts” to analyse queries and direct them to specialised parts of its large language model (LLM).
"Manus is not just another chatbot, nor is it merely an improved search engine dressed in futuristic branding. It is the world’s first fully autonomous AI agent, a system that doesn’t just assist humans — it replaces them. "
Forbes, 08 March 2025
But this post isn’t about Manus. It’s about the national industry that has created two of the biggest AI upsets in less than two months: China.
Putting aside its AI breakthroughs, mounting tensions between the US and its traditional allies have caused more countries to turn towards China as a potential trading partner or even strategic ally.
Just two weeks ago, Jeffrey Sachs told the European Parliament that China “is your competitor, not your enemy.” But how has China suddenly become a dominant global AI competitor?
In this post we’ll dive into:
Key Chinese players and tech advancements
China’s AI policy
The “federated” AI supply chain
Minerals and embargoes
Regulation and censorship
Open-source culture
1. Key players and tech advancements
In contrast to the US AI market, which is dominated by its “Magnificent Seven,” China’s AI industry is extraordinarily diverse with over 4,300 AI companies. Yet it too has dominant players, known as the "Six Tigers." But unlike their well-established US competitors, China’s “Tigers” are all start-ups that have achieved "unicorn" valuations. And DeepSeek doesn't even make the list...
China’s Six AI Tigers
Zhipu AI
Founded in 2019 by two professors from Tsinghua University, Zhipu is one of China’s earliest Generative AI start-ups. Based in Beijing, it develops foundation models and applications that include its chatbot ChatGLM and AI video generator Ying. According to Quartz: “Zhipu raised more than one billion yuan — about $140 million — in a financing round earlier this month that included Alibaba, Tencent, and some state-backed entities.”
Moonshot AI
Moonshot AI also began at Tsinghua University, where it was founded in 2023 by researcher and alum of Carnegie Mellon, Yang Zhilin. Also based in Beijing, Moonshot created the Kimi AI chatbot, which has almost 13 million active monthly users. According to Quartz: “The company, which is valued at $3.3 billion, is backed by some of China’s largest tech firms, including Alibaba and Tencent.”
MiniMax
MiniMax was founded in 2021 and is most famous for its AI chabot Talkie. The chatbot was launched as an app called Glow in 2022, but was rebranded as Xingye in China and Talkie in international markets where available. Specialising in consumer-facing AI, MiniMax is based in Shanghai. It has also developed a text-to-video generator called Hailuo AI. According to Quartz, Alibaba led MiniMax’s $600 million funding round last March, which led to a $2.5 billion valuation.
Baichuan Intelligence
Founded in March 2023, Baichuan is based in Beijing and launched two open-source LLMs in 2023: Baichuan-7B and Baichuan-13B. Although only available in China, the models were tested on multi-language datasets for “general knowledge, mathematics, coding, language translation, law, and medicine.” According to Quartz, Baichuan raised five billion yuan ($687.6 million) in July, during a funding round that valued the company at more than 20 billion yuan. Alibaba, Tencent, and several state-backed funds were among the investors.
StepFun
Based in Shanghai, StepFun has released 11 foundation models, including visual, audio, and multimodal AI systems. It was founded in 2023 by Jiang Daxin, a former SVP at Microsoft. StepFun’s Step-2 language model has 1 trillion parameters, and is ranked highly alongside models from DeepSeek, Alibaba, and OpenAI on LLM benchmarking site LiveBench. According to Quartz, Fortera Capital, a state-owned private equity firm, helped Stepfun raise “hundreds of millions of dollars” in Series B funding last December.
01.AI was founded by Kai-Fu Lee, a former employee of Apple, Microsoft and Google, in 2023. Based in Beijing, it has launched two AI models: Yi-Lightning and Yi-Large. Both are open-source, and are among the top-ranked large language models in the world for language, reasoning, and comprehension. Similar to DeepSeek-R1, the Yi-Lightning model has remarkably efficient training costs. On LinkedIn, Lee said that Yi-Lightning was trained on 2,000 of Nvidia’s H100 chips for one month. Yi-Large can reportedly engage in human-like conversations in both English and Chinese.
Alongside these six “tigers,” China is also home to tech giants Alibaba, Tencent, ByteDance and Huawei.
Within days of the launch of DeepSeek-R1, Alibaba released an updated version of its Qwen 2.5 AI model, Qwen 2.5-Max, which it claims surpasses R1’s capabilities. ByteDance (parent company of TikTok) released its Doubao-1.5-pro model on 22 January, and Tencent (WhatsApp owner) continues to operate its Hunyuan model. Tencent claims that Hunyuan can perform as well as Meta’s Llama 3.1 but requires a tenth of the computing power.
With models that appear far more efficient than their US competitors, and developments including the world’s first autonomous AI system in Manus, how has China produced such a broad spectrum of AI companies and groundbreaking innovations?
2. Government AI policy
Last week, China held its National People’s Congress (NPC) in Beijing on 05 March. The NPC is the highest authority of the People’s Republic of China, and its sole branch of government. All state entities, including the Supreme People’s Court and State Council, are subject to the power of the NPC.
The main event of the NPC was the launch of the Work Plan 2025, China’s equivalent of the US State of the Union Address or UK's King’s Speech. Reading the Work Plan, Prime Minister Li Qiang announced that China is deepening its commitment to driving economic growth through AI and technology.
The Work Plan pledged to increase government spending on AI as part of reaching (or maintaining) China’s 5% economic growth rate.
The core AI strategy of the CCCP (Communist Party of China) is the ‘AI Plus’ Initiative, launched by the Politburo in 2024. The initiative includes major subsidies for biomanufacturing, quantum technology, embodied AI (robots and prosthetics) and 6G networks. China intends to invest 10 trillion yuan ($1.4 trillion) in the next 15 years to gain global dominance in advanced technology.
The “state venture capital guidance fund” was announced on 10 March to “focus on cutting-edge fields such as artificial intelligence, quantum technology and hydrogen energy storage,” according to Zheng Shanjie, Chairman of the National Development and Reform Commission. CNN reports that the fund “is expected to attract nearly 1 trillion yuan ($138 billion) in capital over 20 years from local governments and the private sector.”
But this is not China’s first major AI investment.
In January, the government created a 60 billion yuan AI investment fund, following an additional wave of US export controls on superprocessor chips. In 2023, the Bank of China agreed to invest 1 trillion yuan ($137 billion) in AI over the next 5 years. Total capital spending for China’s tech industry, including from private sector investment, has risen by 61% in the last 12 months, according to Goldman Sachs.
“Beijing plans to invest 10tn Chinese yuan ($1.4tn; £1tn) in the next 15 years as it competes with Washington to gain the edge in advanced tech.”
‘From chatbots to intelligent toys: How AI is booming in China,’ BBC News, 10 March 2025
China’s core AI strategy was launched in 2017, with the goal of becoming the world’s leading AI innovation hub by 2030. By 2020, China was the world number one for published research papers and patents relating to AI. In 2021, the US National Security Commission on AI reported that “US dominance is under threat” in AI: language seen in President Trump’s most recent Executive Order on Removing Barriers to American Leadership in Artificial Intelligence.
China also announced a huge increase in funding for education in STEM subjects, in particular AI, at the NPC last week. This includes a focus on increasing STEM places at universities and providing training for children as young as six in schools. According to the BBC, in 2020 more than 3.5 million of China’s students graduated with degrees in STEM: a number likely to increase in the coming years.
Yet China still faces significant economic turmoil from US tariffs and export controls on superprocessor chips. After the NPC last week, China’s tech CEOs are being discouraged from travelling to the US. However, the country hopes that US withdrawal from geopolitics will create opportunities to fill the vacuum, and that its approach to harnessing AI will provide long-term economic growth.
But China is not seeking to emulate the US industry. Its AI sector has many key differences to the US model...
3. The federated AI supply chain
One of the most striking features of the Chinese AI industry is that there is no equivalent of a single tech centre like Silicon Valley. Instead, China has fostered a “federated supply chain” with different geographic regions dominating each aspect of the AI life cycle.
Beijing is the focal point for core algorithmic development and R&D activities.
Shanghai dominates AI application development, particularly in manufacturing, healthcare and finance.
Shenzhen focuses on the upstream AI market, particularly consumer-focused technology.
DeepSeek is based in Hangzhou, close to financial centre and AI app development hub Shanghai.
As countries around the world, including Britain, rush to emulate the single tech hub approach of the US, it is striking that China’s AI industry is geographically-dispersed into zones that specialise in particular aspects of the AI life cycle.
4. Minerals and embargoes
The geography of China plays a further crucial role in its AI industry. Having such a large land mass, it is rich in many of the rare earth minerals crucial for microchip production.
According to Statista, it currently has the world’s largest rare earth mineral reserves by a huge margin: 44 million metric tons, compared to the US’s 1.8 million (the second country to China is Vietnam, with 22 million).
However, the ability to mine and process rare earth minerals into microchips has proven more challenging. China still struggles to produce the most advanced superprocessors, but is rapidly developing its internal chip production, relying on its integrated supply chains.
"In recent years, China has made significant strides in establishing a comprehensive semiconductor supply chain. Although Chinese enterprises may still lag behind the global vanguard in certain aspects, they have made substantial progress in the overall capabilities across the entire semiconductor supply chain."
Global Times, 09 March 2025
Increasing US export controls on advanced chips to China have been discussed in previous posts, but show no sign of abating as the US adds general trade tariffs on Chinese exports. However, DeepSeek-R1’s development suggests that Chinese innovation may be thriving in a more constrained environment, as computer scientists battle to overcome technological limitations with more efficient mathematics.
In spite of this, tensions between China and the world’s largest superprocessor producer, Taiwan, continue. Last December the largest Chinese fleet in a decade rolled into the Taiwan seas, sparking fears that Taiwan’s disputed independence may be at risk. China’s policy, reiterated in the Work Report 2025, is for "peaceful unification" with Taiwan.
Taiwan’s largest superprocessor producer, TSMC, has recently proposed a joint venture with the US’s Nvidia and Broadcom, according to Reuters. Such a step would be a bold move away from Beijing and towards the US.
However, with China’s mainland chip production expanding, and its abundant supply of minerals ready for extraction and production, it may be that Taiwan will remain free from direct occupation in the near future.
5. Regulation and censorship
In addition to hardware restrictions, the Chinese AI industry has another major constraint on its operational freedom: regulation and censorship.
China is perhaps the most heavily-regulated technology market in the world, with over 40 policy and regulatory initiatives in effect. Two of its most significant regulations are its AI Safety Governance Framework of September 2024, and the Interim Measures for the Management of Generative Artificial Intelligence Services of 2023.
AI safety and governance
The AI Safety Governance Framework was released by the National Technical Committee 260 on Cybersecurity on 09 September 2024. The Framework addresses the ethical, social and safety implications of AI, focusing on a human-centred approach and "principles of developing AI for good.”
As the contents list in the image shows, the Framework includes principles for AI safety governance with an implementation framework, a classification system for AI safety risks and required technological mitigations, and a set of comprehensive AI governance measures with specific safety guidelines for various actors in the supply chain.
The Framework was created to implement China’s Global AI Governance Initiative. In June 2024, the Initiative was the basis for China proposing a UN General Assembly Resolution on Enhancing International Cooperation on Capacity-building of Artificial Intelligence, A/78/L.86. The US supported the Resolution.
In these respects, China echoes many of Europe’s proposals for regulating AI, including those found in the EU AI Act. However, the second aspect of China's regulatory approach is more uniquely Chinese…
Censorship
China’s Interim Measures for the Management of Generative Artificial Intelligence Services, issued on 10 July 2023 by seven departments including the Cyberspace Administration of China, focus on state control over Generative AI (GenAI) and its outputs.
According to Bird and Bird: “The Generative AI Measures apply to the use of generative AI technology that provides services to the “public” within the territories of China.” They do not apply to any development not intended to be provided to the public, echoing the scope of the EU’s AI Act (Art 2).
Where the Measures diverge from the EU, however, is in content control. GenAI models must show “core socialist values” in their output (Art 4(1)), any use of AI to “subvert national unity” is banned (Art 4(1)), and systems must undergo additional government security checks if they may have “public opinion attributes or social mobilization capabilities” (Art 17).
The Measures also require developers to register their algorithms with the government. However, China may not be entirely unique in this approach. Some of the Biden Administration’s actions under Executive Order 14110 required US AI companies to inform the Department of Commerce if they intended to supply foundation models to foreign customers. President Trump has since revoked this Order.
The most stringent censorship aspects of China’s Measures on GenAI are uniquely Chinese. Yet there is evidence that users outside China who download distilled versions of models like DeepSeek-R1 are able to train the model with non-CCCP approved content on their local devices.
Anti-trust and competition
In another contrast with the US, China still retains a strong anti-monopolistic stance against big tech, frequently investigating and acting against any perceived market dominance.
Since 2020, China’s “tech crackdown” has included the State Administration for Market Regulation (SAMR) issuing the Anti-Monopoly Guidelines for the Platform Economy. The Guidelines provided the first legal definition of digital platforms in China, and began a crackdown on anti-competitive agreements, abuse of dominant positions, and monopolistic behaviour. This was followed by revisions to China’s Anti-Monopoly Law in 2022, which increased the fines for severe violations.
High-profile antitrust cases were brought against China’s biggest online platforms during the crackdown, including an unprecedented 18.228 billion yuan fine (around $2.8 billion, or 4% of China revenues in 2019) for Alibaba in 2021. This was the first time the tech giant had recorded a loss since becoming a public company in 2014.
"The tech crackdown ostensibly came in the wake of their unchecked growth and to address monopolistic behavior. Some decisive actions include the SAMR’s imposition of substantial fines on two of China’s corporate giants, Alibaba and Meituan, totaling US$2.8 billion and US$530 million, respectively."
Dezan, Shira and Associates, China Briefing, February 2023
Whilst China retains its distrust of monopolistic tech companies – a trait it shares with the EU – the crackdown has slowed. Now the focus is on the AI Plus Initiative, aimed at driving investment in the entire AI supply chain, including the manufacture of superprocessor chips.
"…in an international digital policy landscape increasingly characterised by digital sovereignty agendas, the issue of monopoly in digital platform economies is a small but potentially overlooked and undervalued point of convergence between the EU and China."
Wang, Yi & Gray, Joanne, ‘China’s evolving stance against tech monopolies: A moment of international alignment in an era of digital sovereignty,’ Media International Australia (2022)
As Europe increasingly looks away from the US and towards other global powers, it seems there is some – if not complete – commonality between the AI regulatory frameworks of the EU and China. How the two regulatory systems evolve in the coming years, however, remains to be seen.
6. Open-source culture
Perhaps as a result of mandatory government reporting and a stringent anti-monopolistic environment, China's AI industry has a strong open-source culture. Many AI models in China are released open-source, including 01.AI’s Yi series, Minimax-01, and over 100 new models from Alibaba Cloud (supporting 29 languages and high application interoperability).
This approach contrasts sharply with the US AI industry, where most models remain closed-source and proprietary (although Meta’s Llama Series are a notable exception).
"While the supposedly free-market US has often prioritised proprietary models, China has built a thriving AI ecosystem by leveraging open-source technology, fostering collaboration between government-backed research institutions and major tech firms."
The Conversation, 12 February 2025
But perhaps China’s open-source culture is beginning to change the US approach. This week, OpenAI proposed a greater developer offering to allow individuals to use its products in a more open-source environment. The move came in direct response to the rise of DeepSeek and release of other Chinese open-source models.
The potential impact of a more open development environment, combined with the availability of smaller distilled models at far lower price points, is huge. Millions more developers, academics and researchers are now able to access some of the world's most advanced AI models, giving them an opportunity to develop the technology further.
It is astonishing to think that a country known for its secrecy and opacity could have ushered in a more transparent era for tech development.
Conclusion
China currently faces a number of demographic, social and economic challenges; whether the growth of its AI industry will solve these issues remains to be seen. Its ambition, however, is undeniable. So too is its effect on the global AI industry.
More open-source development, a greater focus on efficiency and energy savings, and a drive to produce models with fewer advanced microchips: these trends are all the result of China’s recent AI advances.
But perhaps the biggest impact of China's AI race will be seen in future generations.
The BBC reports that more and more children in China use AI toys on a daily basis to learn about language, coding and even how to play chess. Toys with embedded AI are available for as little as $40 in toy stores, built by companies like SenseRobot, which has sold more than 100,000 chess robots in China and recently signed a deal with US supermarket chain Costco.
The rise in AI education and its integration into the everyday lives of children could create the first "AI native" generation in China. How they shape the future of AI will impact us all.
Next month we’ll see robots running a parallel race at the Beijing half-marathon. It’s easy to dismiss this as a publicity stunt or gestural geopolitics, but perhaps we should all look a little more closely at the “embodied AI” on display. They may just be the next big AI advance to come out of China… and onto the shelves of our supermarkets.