Chinese AI models are popular. But can they make money?

In the year since DeepSeek, a little-known artificial-intelligence lab from China, shocked the world with a whizzy new model, the country’s clout in AI has only grown. It is now the undisputed global leader in so-called open-weight models, which make freely available the numerical parameters learned during training (though not the underlying data or source code). On Hugging Face, a popular library for such open models, downloads of Chinese ones have now overtaken those from America (see chart 1).

It is not just DeepSeek—which reportedly plans to release its next whizzy model in a few weeks—that has made waves. In September Qwen, a family of models from Alibaba, a Chinese e-commerce giant, became the most popular on Hugging Face, dethroning Llama, a rival set from Meta. Just after Christmas the American social-media titan announced it would pay more than $2bn to acquire Manus, another Chinese AI lab that had relocated to Singapore (the Chinese government could yet block the deal). In early January MiniMax and Z.ai, two more Chinese model-makers, listed in Hong Kong. Their share prices have since rocketed.

China’s models are world-class. On LMArena, a ranking site, the proprietary models offered by America’s leading AI companies—OpenAI, Google, Anthropic and xAI—dominate the top spots in performance benchmarks. But open alternatives from Alibaba, Z.ai and other Chinese firms are not far behind.

The trouble has been making money from these open models. According to figures from Sensor Tower, a data provider, OpenAI’s ChatGPT made $1.7bn globally through Apple’s App Store in the 12 months to October, compared with a total of $500,000 for the ten most popular Chinese AI chatbots. Most of them remain free to use. Selling services to help businesses deploy open models and infrastructure to run them on are the main ways in which their developers profit. But that has not been easy. MiniMax made just $53m in sales in the first nine months of 2025, and lost $512m. Z.ai is also deeply in the red. China’s AI industry—and particularly its startups—looks headed for a shake-out.

The difficulties of China’s model-makers stem in part from their ferociously competitive yet relatively small domestic market. DeepSeek’s break-out success last year triggered a “culling” as weaker model-makers retreated, notes Kyle Chan of the Brookings Institution, a think-tank. Still, the market remains intensely crowded. According to Bloomberg Intelligence, a research group, there were more than 500 Chinese AI models on offer in September, up from only 14 two years earlier.

Revenue from Chinese consumers, who enjoy an overwhelming range of options, is negligible. For most developers, the enterprise market has not been particularly lucrative either. Chinese companies are notoriously stingy when it comes to spending on the software they use to operate; in total they fork over around $50bn, less than a tenth of what American companies do.

That explains why Chinese AI companies have been looking abroad for growth. One path lies through the global south. Zixuan Li, who leads the overseas business of Z.ai, says the company has been gaining business customers mostly in countries such as Brazil, Malaysia and Singapore, which “have no conflict” with China. In November Singapore’s national AI programme announced that its Sea-Lion model had switched from Llama to Qwen, running on Alibaba’s cloud infrastructure. Goldman Sachs, an investment bank, estimates that big Chinese cloud-computing providers—among them Alibaba, Baidu, ByteDance and Tencent—will invest $70bn this year in data-centre infrastructure, with a particular focus on Asia, Latin America and the Middle East.

Many of these markets, however, are also small. The bigger prize by far is the West, where Chinese models have also been gaining a following. Brian Chesky, chief executive of Airbnb, has noted that the lodging platform is among the companies now using Qwen.

Most of the adoption in the West, however, has been by penny-pinching startups, rather than deep-pocketed corporations. Chinese models are attractive for smaller companies because they are free and engineered to be less computationally intensive than many Western models. They can also be readily adapted for specific purposes. Developers have created more than 180,000 “derivative” versions of the 400 open Qwen models Alibaba has released. Lin Qiao, boss of Fireworks AI, a platform for running open models based in San Francisco, notes that startups want to experiment quickly and cheaply. Open models work “really well” for this.

Bigger businesses, however, remain skittish. According to a survey by Menlo Ventures, a Silicon Valley venture-capital firm, Chinese models account for around 10% of open-model usage among American enterprises (see chart 2). Most rely instead on models from Meta or Mistral, a French developer. One executive at a large American business cites concerns over data privacy, and the risk that a Chinese model could be banned in the West, forcing systems to be adapted. Even when large firms do use Chinese code, they tend to run it on their own infrastructure rather than accessing it via the provider’s cloud. That keeps the risk low, but it also keeps revenue out of the model-maker’s pockets.

Doing business in America may only get harder for Chinese AI companies. Last year Z.ai was added to America’s “entity list”, restricting its access to the country’s technology. Several state governments have prohibited the use of DeepSeek on official computer networks. The saga of TikTok, a wildly popular Chinese-owned short-video app that must now sell its American operation to a local consortium, highlights the treacherous political terrain faced by Chinese platforms.

Europe may prove somewhat more open to business, particularly if its relations with America sour further. As it falls further behind in the race to develop cutting-edge AI, its governments may decide there is more to be gained from encouraging adoption of the technology. Embracing open Chinese models would help with that. Even in Europe, however, Chinese AI providers may struggle to make money. Convincing companies to experiment with free models is one thing; persuading them to use Chinese-owned data centres is quite another.

All this will not pose an existential problem for tech giants such as Alibaba and Tencent. By embedding AI into other services, such as e-commerce and social media, they can make these better and tighten their grip over users at home. They have cash to spare, and are investing far less than America’s tech giants in data centres. For China’s many AI startups, however, difficult times lie ahead. ■