China’s AI scale spikes
A specialist report says China’s AI systems used about 140 trillion tokens a day in March, driven by a mix of domestic models, Huawei’s Ascend chips and coordinated state strategy. That scale suggests China is increasingly vertically integrating models, chips and industrial policy to push heavy demand for AI compute and infrastructure. (digitalinasia.com)
China’s AI boom is getting easier to measure. In March, China’s National Data Administration said the country’s AI systems were processing more than 140 trillion tokens a day. That is up from 100 billion a day at the start of 2024 and 100 trillion by the end of 2025. The jump is so large that it stops sounding like a benchmark and starts sounding like infrastructure. (cntechpost.com) That number matters because tokens are not an abstract lab metric anymore. They are the unit that turns AI into a billable service. Chinese officials are talking about them that way in public. Liu Liehong, who heads the National Data Administration, described tokens as a settlement unit that links technical supply to commercial demand. He also said some model companies made more revenue in the first 20 days of 2026 than in all of 2025. That is what commercialization looks like when it arrives all at once. (cntechpost.com) The surprising part is not just the scale inside China. It is that Chinese models are also showing up in global usage data. Digital in Asia’s report points to OpenRouter, a large model API aggregation platform, where Chinese models accounted for 61% of token consumption among the top ten models in February 2026. During the week of February 16 to 22, those models processed 5.16 trillion tokens on the platform, versus 2.7 trillion for U.S. models. Four of the five most-used models were Chinese. That does not prove China leads on every frontier benchmark. It does show that cheap, capable models are getting used heavily in the wild. (digitalinasia.com) Usage at that level does not come from one breakout lab. The report’s main point is that the Western habit of reducing Chinese AI to DeepSeek misses the real story. China now has a wider model field, with companies such as Alibaba, DeepSeek, MiniMax, Moonshot, and Zhipu pushing different strategies at once. DeepSeek still matters because it helped reset assumptions about cost. Digital in Asia says its R1 model used a mixture-of-experts design that activated only a fraction of total parameters per inference, cutting compute needs sharply, and that its V4 model launched in February 2026. But the bigger shift is that there are now enough domestic models, serving enough real customers, to keep demand rising even without a single dominant winner. (digitalinasia.com) That demand runs straight into hardware. U.S. export controls were supposed to slow China by restricting access to Nvidia’s most advanced AI chips. Instead, they seem to have accelerated a different kind of buildout. Huawei’s Ascend line has become the spine of a domestic alternative stack. Huawei’s own enterprise material describes Ascend not as a chip but as a full computing platform, spanning cards, servers, clusters, and software for training and inference across edge, cloud, and data center deployments. That is the important detail. China is not only trying to replace one processor with another. It is trying to replace the whole system around it. (e.huawei.com) State policy is helping force the pieces together. China Daily reported that 26 government departments were coordinated to support pilot efforts around data sharing and access, and that the country had built more than 100,000 high-quality datasets by the end of 2025. The National Data Administration is also drafting policy for a national integrated data market and a unified data property-rights registration system. This is dry material. It is also the machinery that turns AI from a collection of companies into industrial policy. Models need data. Data needs rules. Rules shape where the servers get built. (cntechpost.com) Huawei sits in the middle of that machine. Its 2025 annual report says Huawei Cloud had expanded to 34 regions and 101 availability zones by the end of the year, while the company said it would keep investing in computing, cloud, and AI and continue building the ecosystem around Ascend. Even reports noting weaker cloud growth make the same underlying point: Ascend is no side project. It is one of the company’s strategic pillars. In China’s AI push, that matters because the country does not need a perfect Nvidia clone. It needs enough domestic chips, enough domestic cloud, and enough domestic models to keep 140 trillion tokens a day moving. (huawei.com)
