Alibaba's next Qwen AI will demand a cut of your revenue — but only if you're big

Alibaba is preparing to draw a line in the sand with its next Qwen open-weight model: if you’re making real money off it, the company wants its share. Reuters reports that the Chinese tech giant will require larger commercial users who offer the model as a service to sign a revenue-sharing agreement. The exact rate hasn’t been locked down yet, but the shift marks a clear departure from the Apache 2.0 license that governed the current Qwen3 family, which generally let companies deploy the models in their own data centers without paying a licensing fee.

This isn’t happening in a vacuum. Alibaba appears to be following a trail blazed by rival Moonshot, which dropped similar terms with its Kimi K3 model last month. Moonshot’s license triggers a separate commercial agreement when a company’s combined revenue crosses $20 million in any 12-month period. One source told Reuters that Moonshot can demand up to 30% of the revenue involved in those deals—a startling figure for something often marketed as “open.” DigitalOcean CEO Paddy Srinivasan, whose company offers Kimi K3, calls it an open-source “freemium” model, which is a polite way of saying the meter starts running once your business gets traction.

The move exposes a growing tension in the AI world between the ideology of open-source and the brute economics of serving massive models. Terms like “open-weight” are already under scrutiny because downloadable parameters don’t automatically mean unrestricted commercial use, as the Open Source Initiative’s formal definition makes clear. Alibaba’s plan effectively creates a two-tier system: tinkerers and startups skate for free, but any company building a real business on Qwen’s back will have to negotiate. It’s a pragmatic acknowledgment that giving away a model with nearly 2.4 trillion parameters isn’t charity—it’s a land grab that eventually needs to produce a return.

I’m not surprised to see this model spreading. The infrastructure costs alone for running a model at this scale are punishing. Moonshot had to temporarily halt new sign-ups for Kimi K3 in July because usage was crushing its GPU capacity, and few users are expected to self-host a model that activates around 100 billion parameters per query. When the cost of serving a single model can choke a provider’s entire infrastructure, the era of purely altruistic open-weight releases was never going to last. The question now is whether developers will tolerate a 30% haircut on their AI revenue, or if this will push more commercial users back toward the closed, predictable pricing of OpenAI and Anthropic.