Alibaba and Moonshot Are Quietly Charging for “Free” AI — and U.S. Firms Are Paying

Alibaba is preparing to require commercial revenue sharing from large-scale users of its next Qwen open-source AI model — a move that confirms Chinese AI labs are no longer content to give away frontier technology without a financial return.

The development, reported by Reuters citing two people familiar with Alibaba’s strategy, arrives as Moonshot — the Chinese startup behind the Kimi K3 model — has already embedded similar commercial terms into its licensing agreement. Together, the moves signal a coordinated, if separately executed, shift in how Chinese AI firms plan to monetise their most powerful models. For the broader industry, the question is no longer whether open-source AI is free. It is how much it will cost, and who will pay.

The Context

For most of the past two years, the defining narrative around Chinese AI has been disruption through openness. While OpenAI, Anthropic, and Alphabet’s Google maintained closed, proprietary model architectures, Chinese labs including Alibaba and Moonshot released open-weight models — making the underlying learned parameters freely downloadable — and won rapid global adoption as a result.

The strategy worked. Developers, cloud providers, and enterprises across the United States and Europe integrated Chinese models into production pipelines, often because the performance-to-cost ratio was compelling. Moonshot’s Kimi K3 demonstrated that China’s open-source AI strategy was gaining real traction, with the model achieving competitive benchmark scores while pricing input and output tokens at roughly a third of the cost of Anthropic’s equivalent Fable model, according to publicly listed prices for both.

That price advantage attracted exactly the customers Chinese labs needed: high-volume, commercially serious users in the United States and Europe. It is precisely those users who are now being asked to share a portion of what they earn.

The Move

Alibaba plans to implement commercial revenue-sharing terms for its forthcoming Qwen3.8-Max model as early as next week, according to the two anonymous sources cited by Reuters. The model is open-source and open-weight — its parameters will remain downloadable — but the licensing terms will include a clause requiring commercial-scale users to negotiate a revenue-sharing agreement with Alibaba.

The framework mirrors what Moonshot has already done with Kimi K3. Tucked into that model’s licensing terms is a provision requiring any party that offers the model as a commercial service and generates more than $20 million in annual revenue to enter into a commercial agreement with Moonshot. One source told Reuters that Moonshot is seeking up to a 30% revenue share under those agreements. The Alibaba rate is not yet finalised, as discussions with prospective partners are ongoing.

The commercial reality is already materialising. Chinasoft International, a Chinese IT services firm listed in Hong Kong, disclosed a revenue-sharing agreement with Moonshot in a regulatory filing last month, though the percentage was not revealed. DigitalOcean, the U.S. cloud computing provider, confirmed it has a commercial agreement with Moonshot while offering Kimi K3 and other Chinese models to its customers. “This is a tried and tested open-source ‘freemium’ model,” said Paddy Srinivasan, DigitalOcean’s CEO, in comments to Reuters.

What is striking about the timing is not that Alibaba and Moonshot are pursuing revenue — it is that they are doing so simultaneously and with structurally similar terms, despite operating as separate companies in a nominally competitive market. Whether this represents informal coordination, parallel strategic reasoning, or simply the same MBA playbook applied to the same problem, the effect is the same: the two most globally distributed Chinese open-source AI models are now converging on a licensing architecture that turns scale against the very users who made those models globally relevant. The more successful a U.S. company becomes by deploying a Chinese open-source model, the more obligated it becomes to the Chinese lab that built it.

The Stakeholders

Alibaba

For Alibaba, the Qwen model family has served a dual purpose: demonstrating technical credibility in AI and driving adoption of its cloud computing platform, Alibaba Cloud. To date, the company has charged developers for API access when models are hosted on its own infrastructure but permitted most open-source deployments in customers’ own data centres without payment. The planned revenue-sharing clause would extend Alibaba’s commercial reach beyond its own cloud, capturing value from the third-party deployments it previously subsidised. The move is consistent with a broader industry shift from AI hype to cost discipline, in which providers of all kinds are seeking durable revenue rather than pure growth metrics.

Moonshot

Moonshot occupies a more aggressive position. The startup is operating in a geopolitically charged environment: the White House has accused it of stealing technology from Anthropic, a claim Chinese officials have called unfounded. Despite that backdrop, U.S. cloud providers including DigitalOcean are offering Kimi K3 and confirming commercial arrangements with Moonshot — a fact that illustrates how market incentives can outpace policy concerns. Mark Zuckerberg has publicly argued that banning Chinese AI models would backfire, and the commercial traction Moonshot is achieving lends that argument concrete weight.

U.S. Cloud and Infrastructure Providers

For companies like DigitalOcean, Together AI, and Fireworks AI, Chinese open-source models represent a product and margin opportunity. Dan Fu, vice president of kernels at Together AI, explained the logic to Reuters: infrastructure providers make money by optimising how models run — better token efficiency, lower latency — rather than by owning the model weights themselves. “At the application layer, there’s value out there for how you use it, how you actually get the models and the tokens to do something useful,” Fu said. Revenue-sharing terms complicate that calculus but do not eliminate it, as long as the underlying models remain competitively priced and performant.

U.S. AI Labs

OpenAI, Anthropic, and Google remain closed-source, and their models remain more expensive. The freemium approach being adopted by Chinese labs puts pressure on them in a specific way: it offers enterprises a low-friction entry point to frontier AI, with costs that only appear at scale. Thinking Machines Lab, the San Francisco startup founded by former OpenAI CTO Mira Murati, released its first open-source model last month and is widely expected to release more powerful ones — a sign that U.S. labs are beginning to treat open-source as a competitive necessity rather than a philosophical choice. “I don’t see a fundamental barrier” to powerful open-source U.S. models, said Lin Qiao, CEO of Fireworks AI, speaking to Reuters.

How Chinese Open-Source AI Licensing Compares to Alternatives

The comparison is instructive. Meta’s Llama models — the most widely deployed Western open-weight alternative — impose a commercial agreement threshold based on user volume rather than revenue, and do not seek a revenue share. Chinese labs are applying a revenue-sharing mechanism that more directly extracts value from commercial success. As Microsoft pushes developers toward OpenAI’s proprietary models through GitHub Copilot, the Chinese open-source approach offers a structurally different and, for many developers, cheaper entry point — at least until revenue thresholds are crossed.

The security dimension also deserves attention. Reports that Chinese military researchers used OpenAI and Anthropic models to train defence AI have intensified scrutiny of cross-border AI technology flows. Revenue-sharing agreements between U.S. companies and Chinese AI labs will likely draw additional regulatory attention, particularly given the existing export control environment. Whether that scrutiny translates into policy action remains to be seen, but the commercial relationships now being formalised are precisely the kind that regulators and lawmakers have been monitoring.

The Implications That Matter

1. Open-source is no longer a synonym for free at commercial scale. Both Alibaba and Moonshot have now embedded revenue-sharing triggers into their licensing terms, establishing a precedent that will shape how every future Chinese open-weight model is evaluated by enterprise legal and procurement teams.
2. U.S. firms paying Chinese labs for AI access creates regulatory exposure. Formalised revenue-sharing agreements between American cloud providers and Chinese AI labs — during a period of active technology-transfer scrutiny — will attract attention from the White House, OFAC, and relevant Congressional committees, regardless of whether the underlying technology is commercially available.
3. The freemium playbook pressures Western closed-source labs on pricing. As Chinese models price API tokens at a fraction of OpenAI and Anthropic’s rates while offering comparable performance, the cost argument for proprietary Western AI weakens — particularly for cost-sensitive mid-market enterprise buyers who may not hit the revenue-share threshold.
4. Meta’s Llama licensing model may face competitive re-evaluation. If Alibaba and Moonshot successfully extract revenue shares from commercial deployments, Meta and future U.S. open-source entrants will face investor pressure to adopt similar mechanisms, potentially fragmenting what has been a relatively permissive open-source AI ecosystem.
5. The next frontier is enforcement, not licensing. Writing a revenue-sharing clause into an open-weight model license is straightforward; enforcing it across jurisdictions, particularly when model weights can be downloaded and redeployed without disclosure, is a materially harder problem. How Chinese labs handle non-compliance will determine whether this business model is durable or merely aspirational.