Liu Liehong, the administrator of China’s National Data Administration, the country’s main data regulator, unveiled the term at a State Council press conference in March, explaining that tokens were now “the settlement unit linking technological supply with commercial demand.”
The National Data Administration disclosed that China now processes 140 trillion tokens every day, up from just 100 billion at the start of 2024. Chinese AI models have now surpassed U.S. models on OpenRouter, a popular marketplace for AI models.
Investors have bought into the AI boom. IPOs in Hong Kong are at a five-year high thanks to a steady stream of Chinese AI and tech startups, including AI labs MiniMax and Zhipu AI, and chip designer Biren.
“We believe that China is the big winner in this tech war for a number of reasons: valuation, wider adoption of AI, an advantage in power generation,” Mohit Kumar, Jefferies’ global macro strategist, told Fortune in mid-March at the bank’s Asia Forum in Hong Kong.
China’s goal is now to build a “token economy,” backed by a proliferation of efficient, open-source models and a push into real-world AI applications. Yet like their U.S. peers, Chinese firms are grappling with expensive research costs and heavy capital expenditure pledges, while also fending off Washington’s export controls, designed to keep them one step behind in the chip race.
Big tech pivots
The AI boom rescued China’s big tech companies from years of regulatory purgatory.
Alibaba, the e-commerce giant, has invested in open-source models, which can be downloaded and modified freely by developers. That low barrier to entry has made its Qwen models a compelling option for startups unwilling to pay for proprietary models from OpenAI and Anthropic. Qwen has won over developers from Southeast Asia to the Middle East, and it’s also convinced Western users too: Meta’s most recent model, Muse Spark, is trained partly off of Qwen.
Unlike Alibaba, ByteDance has largely kept its AI models proprietary, instead leveraging its product design and consumer experience strengths to win users. The company’s chatbot, also called Doubao, is China’s most-used AI app, with 100 million daily active users over the Chinese New Year holiday in February.
Tencent, which operates the ubiquitous WeChat messaging platform, has been a step behind its rivals when it comes to AI. The company launched ClawBot in March, which appears as a contact within WeChat, allowing its over one billion monthly active users to connect directly with OpenClaw and execute tasks through the messaging interface.
Competition is fierce within China’s tech sector. Last week, Alibaba revealed its newest video generation model, Happy Horse, which performs better than the current leader, ByteDance’s SeeDance, according to some analyses.
And there’s still potential for another big tech company to shake things up. Xiaomi and Meituan, better known for smartphones and food delivery respectively, have launched their own large models.
Smaller startups
A new generation of Chinese AI startups are also winning converts in Silicon Valley.
When vibe-coding startup Cursor launched Composer 2, its latest coding service, eagle-eyed users discovered that the model had been built on Kimi K2.5, an open-source model from Beijing-based Moonshot AI. Cursor’s co-founder later acknowledged it was “a miss to not mention the Kimi base…from the start.”
Two other startups—Knowledge Atlas, better known as Z.ai, and MiniMax—have already listed in Hong Kong, giving some rare visibility into the economics of a frontier AI lab.
MiniMax reported $79 million in 2025 revenue, a 159% year-on-year jump, with 70% coming from overseas markets in an early signal of global appetite for Chinese foundation models. Yet it also posted an adjusted net loss of $250 million. Zhipu AI generated 724 million yuan ($104.8 million) in revenue, 132% higher than the year before, but its total losses ballooned to 4.7 billion yuan ($680 million), driven by R&D spending that jumped 45%.
Investors don’t seem to mind the massive losses. Zhipu’s shares are up more than 570% from its IPO price; MiniMax has risen more than 470%, at one point briefly exceeding the market cap of Baidu. Still, both stocks have swung wildly, rising and falling by double-digit percentages in single sessions.
Moonshot AI, backed by Alibaba and HongShan, is reportedly weighing a Hong Kong IPO, coming just a few months after a January funding round that valued the startup at $10 billion.
One startup that’s been notably quiet this year is DeepSeek, the Hangzhou-based lab that reset the whole AI conversation last year with its V3 and R1 models. Developers are eagerly awaiting the public release of V4, the latest version of its model.
Physical AI
China is also surging ahead in physical AI, backed by supply chains that can cheaply manufacture advanced technology.
Unitree Robotics, perhaps China’s most prominent humanoid robot startup, has filed for a 4.2 billion yuan ($610 million) IPO on Shanghai’s STAR Market. Unlike many of its robotics peers in China and overseas, Unitree doesn’t lose money, posting an adjusted net profit of roughly 600 million yuan ($87 million). Other major Chinese robotics startups include Agibot and UBTech.
Chinese companies are also pushing hard in automated driving. Pony AI launched Europe’s first commercial robotaxi service in Zagreb, Croatia in early April, in partnership with Uber and Croatian operator Verne. WeRide has also partnered with Uber to offer fully commercial robotaxis in Dubai.
Governments, consumers get on board
Chinese users are far more comfortable with AI than their Western counterparts. An Edelman survey from October found that 87% of Chinese respondents trust AI, against 32% in the U.S.
The country’s short drama industry is just one example of consumer comfort with AI. Video platforms launched roughly 470 new dramas every day in January, thanks to plummeting production costs. A short drama can now be generated with AI tools for around 100,000 yuan ($14,600), about ten percent of the conventional cost, with the production window shortened from 15–30 days to under five.
Chinese consumers are also embracing AI agents, with a series of major tech companies hosting workshops to walk potential users through the process of installing OpenClaw on their personal devices.
Local governments are amplifying the push, offering subsidies to “one-person companies,” solo entrepreneurs building AI agent businesses.
Beijing’s approach is more measured, both pushing AI as a strategic priority while also proactively moving to ward off some potential risks, such as by warning against security vulnerabilities in OpenClaw-based agents and proposing regulations for AI companion apps.
Yet the most significant policy advantage may not be directly connected to AI at all. China has aggressively expanded its power generation and transmission capacity in recent years. Goldman Sachs estimates that China will have approximately 400 gigawatts of spare power capacity by 2030, roughly three times projected global data center demand.
Constraints at home and abroad
Still, Chinese AI companies face numerous headwinds that constrain what they can do, particularly compared to the leading U.S. AI developers.
Due to U.S. export controls limiting the sale of the most advanced AI chips to China, domestic companies are forced to rely on domestically made chips, primarily from Huawei; overseas data centers; or on U.S. hardware sourced through grey markets. Chinese chips are getting better: on April 8, Alibaba unveiled a new data center run entirely on its own home-designed Zhenwu chips. Yet production yields and performance still remain far behind the U.S. chip supply chain.
China’s venture capital ecosystem is also thinner than Silicon Valley’s. Unease with Beijing’s tech regulation and U.S. regulatory pressure lead many global investors to avoid Chinese startups. Moonshot AI, at an $18 billion valuation, commands mostly China-based investors. Anthropic, by contrast, raised $30 billion in a Series G round in February 2026, at a $380 billion post-money valuation, backed by a global consortium of deep-pocketed institutional investors including GIC, Coatue, Founders Fund, and ICONIQ.
That funding pressure forced some founders to take radical action, with some going as far as skipping the Chinese market entirely. Manus AI, which launched a buzzy AI agent last year, reincorporated as a Singapore entity; Meta later acquired the agentic AI startup for roughly $2 billion in late 2025.
Beijing has taken a dim view of the deal. Two Manus co-founders, CEO Xiao Hong and chief scientist Ji Yichao, are now subject to an exit ban, according to the Financial Times.
The token economy
Yet the biggest unresolved question in Chinese AI is much the same as in the U.S.: How to turn tokens into profits.
Alibaba spent 123 billion yuan ($17 billion) on capital expenditure in 2025, which helped contribute to a 66% plunge in net income. Tencent hasn’t spent quite as much money, with capex of just 79 billion yuan ($11.6 billion). ByteDance, as a private company, faces less pressure from shareholders about profitability, but the Financial Timesreported late last year that the TikTok owner expects to spend $23 billion on AI infrastructure.
That’s still a lot smaller than what U.S. giants are spending. Alphabet spent $94 billion on capital expenditures last year; Meta spent $75 billion. Both companies plan to spend even more this year.
But monetization pressure may already be pushing some of China’s tech companies to rethink their strategy. Both Alibaba and Z.ai have released some of their most recent models in a closed format, at least at first. Both companies, as well as others like Baidu, are also hiking prices for their models and cloud services.
Going forward, China’s tech companies are going to put AI at the center of their business. Last month, Alibaba reorganized its entire AI operation into what it calls the “Alibaba Token Hub,” which consolidates five previously separate units, including Tongyi Laboratory (its foundational model research arm), Qwen, and an enterprise AI division called Wukong, under CEO Eddie Wu’s direct oversight.
“ATH is built around a single organising mission: create tokens, deliver tokens and apply tokens,” Wu said in a letter announcing the reorganization.
