For the past two years, China's AI labs have promoted a simple proposition: release models that rival leading systems from Silicon Valley, then make the weights available so developers can build businesses around them with few restrictions.
That approach may be beginning to change.
Alibaba is preparing to ask the largest commercial users of the next version of its Qwen model to share a portion of the revenue they generate from it, according to two people familiar with the company's plans. The measure is expected to accompany the release of Qwen3.8-Max as early as next week, the people said, speaking on condition of anonymity because the plans have not been made public.
The model itself will remain open-weight. Developers will still be able to download it and fine-tune it, running the model on their own hardware-the defining feature that has distinguished many Chinese releases from the closed systems developed by OpenAI, Anthropic and Google. What could change is how commercial use is licensed once companies begin generating significant revenue from those models.
Alibaba would not be the first company to adopt such an approach.
Last month's release of Kimi K3 by Chinese AI startup Moonshot included a licensing clause requiring organizations that resell the model as a service and generate more than $20 million in annual revenue to negotiate a commercial agreement. According to the same two people, those agreements can involve revenue-sharing arrangements, in some cases reaching as much as 30%.
The arrangement is no longer purely theoretical. A Chinese IT services provider disclosed a revenue-sharing agreement with Moonshot in a regulatory filing last month, although it did not disclose the percentage involved. Moonshot did not respond to a request for comment.
Alibaba is expected to adopt a similar framework, the two people said, although the revenue-sharing rate has not been finalized and discussions remain ongoing. Until now, the company has primarily charged customers using Qwen through Alibaba Cloud, while allowing most open-weight releases to run without licensing fees on customers' own infrastructure.
The strategy reflects a model widely used in open-source software: make the core technology broadly available, while generating revenue from large commercial deployments and related services.
"This is a tried and tested open-source 'freemium' model," said Paddy Srinivasan, chief executive of DigitalOcean Holdings, one of several U.S. cloud providers offering access to Chinese AI models. He confirmed that DigitalOcean has a commercial agreement with Moonshot but declined to discuss its terms. Customers, he said, are paying for deployment help and access to future releases as much as for the model itself.
Dan Fu, vice president of kernels at Together AI, said much of the value comes from improving how efficiently models operate in production, including reducing token costs and adapting models for enterprise workloads.
The economics help explain the appeal. Based on published input and output token pricing, Kimi K3 is priced at roughly one-third the cost of Anthropic's Fable model, making a revenue-sharing agreement less significant for many commercial users.
The business rationale is straightforward. The broader geopolitical context is more complex.
The reported arrangements would direct revenue from U.S. companies to Chinese AI developers at a time when the White House has alleged that Moonshot used technology derived from Anthropic's models. Chinese officials have rejected the allegation. Regardless of how those claims are ultimately resolved, commercial relationships between U.S. businesses and Chinese AI developers continue to expand.
Meanwhile, China's lead in open-weight AI is facing increasing competition. Thinking Machines Lab, the San Francisco startup founded by former OpenAI Chief Technology Officer Mira Murati, released its first open model last month, with more advanced releases widely expected in the coming months.
"I don't see a fundamental barrier" to highly capable open-source AI models emerging from the United States, said Lin Qiao, chief executive and co-founder of Fireworks AI, who declined to discuss her company's commercial arrangements with Moonshot.
The specific revenue-sharing percentage is the number that matters.
A modest single-digit share would resemble a conventional commercial licensing fee for enterprise users. A substantially higher percentage, approaching the upper levels reportedly used in some Moonshot agreements, would represent a more meaningful shift in how open-weight AI models are monetized.
Comments