Chinese artificial intelligence developers could start charging cloud platforms commercial licensing fees to host their open-weight models, according to the head of Asia internet research at Goldman Sachs, as firms look to capture more revenue from their soaring global use. The assessment comes as Chinese AI models – such as Moonshot AI’s Kimi K3 and Zhipu AI’s GLM-5.2 – have reached performance levels just a fraction behind top US rivals. However, because Chinese developers typically release their models under open-source terms, foreign platforms and developers can freely download, modify and host the software’s core “weights” – the underlying parameters that encode their intelligence. Ronald Keung, Hong Kong-based analyst at the US investment bank, said Chinese model builders could generate more revenue by requiring third-party providers to buy commercial licences to serve these models – a process known as inference – on their own infrastructure. “Domestic growth in adoption is very fast, and we think there’s considerable adoption among small and medium enterprises in the global market, and even larger [companies] are starting to consider using [Chinese models],” Keung said. An open-source revenue gap Most Chinese models are currently distributed under permissive software licences, such as the MIT License, which originated from the Massachusetts Institute of Technology, allowing anyone to use, modify and host the models for free – even for commercial purposes. This means that rapid global adoption of Chinese AI is not directly generating revenue for the original creators. Instead, cloud providers, ranging from Alibaba Cloud and Microsoft Azure to specialised platforms like Together AI, host the models and capture the spending from users running them on those clouds. On OpenRouter, a popular online marketplace for US enterprise AI users, Chinese models recently captured a record 60 per cent of usage. For Zhipu’s GLM-5.2 alone, 34 inference providers are listed, most of them based in the US, while Zhipu’s own official channel is just one of them. Keung said a domestic shortage of computing power in China had effectively forced Chinese start-ups to let foreign third parties host their models. While Chinese AI model companies have seen rapid revenue growth this year – Zhipu has seen its annual recurring revenue (ARR) grow four times since March to US$1 billion in July, while tech outlet The Information reported that DeepSeek’s has reached US$500 million – they are still far behind their US counterparts Anthropic and OpenAI, with ARRs of US$74.1 billion and US$41.3 billion, respectively. “Even if [Chinese models] have very strong demand, they often run into not being able to satisfy that demand themselves,” Keung said. Allowing third parties to serve their open models helps Chinese firms build global branding without bearing the computational costs themselves. Chinese models have reached the level now, particularly in coding, where they can self-improve more sustainably Ronald Keung, Goldman Sachs Rethinking free-use policies To capture a bigger piece of the financial pie, some Chinese start-ups are beginning to rethink their free-use policies for third parties. In April, Shanghai-based MiniMax released its model M2.7 which, for the first time, required prior written permission from the company for commercial use. Following backlash from open-source enthusiasts, MiniMax softened its rules to allow some commercial use without prior authorisation, while enterprises with more than US$20 million in annual revenue are still required to obtain a commercial licence. On Monday, Moonshot AI introduced similar requirements with the same US$20 million threshold as it released the weights of its K3 model, marking a shift away from permissive licensing for the Beijing-based start-up. Neither Moonshot nor MiniMax have disclosed their licence pricing or customer count. The two companies did not respond to requests for comment. For MiniMax, the company’s experience underscores the risks of restricting access. Since shifting its licensing strategy, MiniMax’s Hong Kong-listed shares have fallen more than 80 per cent from their peak in March. By contrast, rival Zhipu AI – which kept its flagship models free and open – has seen its stock jump over 100 per cent in the same period. Technology giants Alibaba Group Holding, which owns the South China Morning Post, and Tencent Holdings have also allowed more permissive usage of their latest flagship open-weight models in recent weeks. High margins vs adoption risks A major uncertainty, however, was how quickly this shift might happen, Keung said. The challenge for a company in shifting towards commercial licences is that enterprise users and hyperscalers can easily switch to using the host of other Chinese open models with more permissive licences. Still, Keung said, commercial licensing offers a lucrative, long-term alternative revenue stream for Chinese open model developers. Commercial licences would be “highly margin accretive”, Keung said, because licensing fees deliver direct software revenue without requiring Chinese AI firms to pay for the expensive hardware needed to run the models. Goldman Sachs forecasts a 12-fold increase in Chinese AI model companies’ combined annual recurring revenue over the next four years, growing from US$10 billion at the end of 2026 to US$125 billion in 2030, with gross margins for model inference expected to rise to as much as 40 per cent. Still, recent injections of fresh capital would give Chinese companies more freedom to maintain permissive open-source licences while growing their user base, according to Keung. The Hangzhou-based start-up DeepSeek completed a 50 billion yuan (US$7.4 billion) funding round last month, while Zhipu – known internationally as Z.ai – and MiniMax raised US$4 billion and US$2 billion, respectively, through stock and bond sales earlier this month. “I would say that commercial licences may be a way to generate more revenue, but if you have lots of money on your balance sheet for the time being, then maybe it’s not the first thing these model companies may be focused on,” Keung said. “But I think the MiniMax example shows that it is feasible.” Geopolitical and technical hurdles Beyond business strategy, Chinese AI firms faced regulatory uncertainties, Keung said. In the US, the Trump administration has signalled plans to crack down on “distillation” – the practice of using outputs from advanced models to train and improve other AI systems. A Goldman Sachs report published earlier this month noted that, while distillation had helped accelerate Chinese progress this year, Chinese firms were likely to become less reliant on distilling American AI models. As more developers globally use Chinese models, particularly for writing software code, the platforms gain valuable real-world data that feeds into continuous self-improvement. “Chinese models have reached the level now, particularly in coding, where they can self-improve more sustainably,” Keung said.