Moonshot seeks up to 30% of Kimi K3 sales on Azure, AWS and Google Cloud

Moonshot AI is negotiating with Microsoft, Amazon, and Google to earn up to 30% of revenue from Kimi K3 services sold through Azure, AWS, and Google Cloud, according to Reuters. The talks, which may end without agreements, aim to monetize the 2.8-trillion-parameter open-weight model through enterprise cloud distribution, testing whether open-weight labs can achieve proprietary-model economics.

Moonshot seeks up to 30% of Kimi K3 sales on Azure, AWS and Google Cloud Reuters reports that Yang Zhilin's Beijing lab is seeking a share of K3-related service revenue as it turns a 2.8-trillion-parameter open-weight model into an enterprise cloud offering. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Reuters https://www.reuters.com/business/retail-consumer/chinas-moonshot-talks-with-microsoft-amazon-google-over-k3-revenue-sharing-2026-08-26/ Why it matters Moonshot is testing whether an open-weight lab can keep model access broad while earning proprietary-model economics through US cloud distribution. The answer will shape how other independent AI labs fund increasingly expensive releases. Moonshot AI https://www.moonshot.ai/about?ref=runtimewire is negotiating with Microsoft, Amazon and Google for access to the distribution channels that its open-weight strategy cannot provide on its own. Moonshot wants as much as 30% of the revenue generated by Kimi K3 /models/moonshotai/kimi-k3 services sold through Microsoft Azure, Amazon Web Services and Google Cloud, according to three people familiar with the preliminary discussions who spoke to Reuters https://www.reuters.com/business/retail-consumer/chinas-moonshot-talks-with-microsoft-amazon-google-over-k3-revenue-sharing-2026-08-26/?ref=runtimewire . The negotiations include unresolved questions over data access, the revenue split and how the parties would audit token usage. They may still end without agreements. The talks put a commercial structure around Yang's central bet: release a large model's weights to accelerate adoption, then collect revenue where most enterprises will actually run it. K3 can be downloaded and modified, though few customers have the infrastructure or appetite to operate a model containing 2.8 trillion parameters. The cloud providers already have the enterprise contracts, billing systems and computing capacity needed to convert developer interest into recurring usage. Open weights still need a sales channel Moonshot launched Kimi K3 on July 16 and released its weights on July 27 /article/kimi-k3-yang-zhilin-long-context-open-weight-agents . Its technical report https://arxiv.org/abs/2607.24653?ref=runtimewire describes a mixture-of-experts system with 2.8 trillion total parameters, 104 billion activated parameters and a one-million-token context window. K3 is designed for coding, reasoning, multimodal work and agentic tasks that stretch across long sequences. arxiv.org https://arxiv.org/abs/2607.24653?ref=runtimewire Those specifications explain why downloadable weights do not eliminate the need for a commercial host. Running K3 requires substantially more infrastructure than deploying a small local model, even though its sparse architecture activates only part of the system for each token. Azure, AWS and Google Cloud would give Moonshot a path into enterprises that want K3's economics or capabilities without building a specialized inference stack. Moonshot's proposed share is a ceiling rather than an agreed rate. Reuters reported that the request is consistent with terms Moonshot has outlined for other large customers, and that Moonshot has already reached similar arrangements with smaller cloud platforms. Chinasoft International disclosed a revenue-sharing and product collaboration with Moonshot in July, without publishing the financial terms. A deal with any of the three US providers would also test how much value remains with a model developer after a cloud platform supplies the hardware, customer relationship, security controls and billing. Token auditing is particularly important under that structure. Moonshot would need enough visibility to verify usage and calculate its payment, while cloud customers may resist giving a Chinese model provider broad access to operational data. Yang's research thesis reaches the billing layer Yang arrived at this problem through language-model research rather than cloud sales. He studied computer science at Tsinghua University and completed a doctorate at Carnegie Mellon University in 2019. Before returning to China, he worked with research groups at Google Brain and Meta AI, according to a July profile by Channel News Asia https://www.channelnewsasia.com/east-asia/china-deepseek-zhipu-ai-moonshot-liang-wenfeng-yang-zhilin-tang-jie-6269866?ref=runtimewire . channelnewsasia.com https://www.channelnewsasia.com/east-asia/china-deepseek-zhipu-ai-moonshot-liang-wenfeng-yang-zhilin-tang-jie-6269866?ref=runtimewire Yang co-authored Transformer-XL, which addressed the fixed context limits of early Transformer systems, and XLNet, a pretraining method that built on that work. Long context remained a defining idea when Yang founded Moonshot in early 2023. Moonshot's first Kimi assistant emphasized the amount of text it could process, and K3 extends that thesis to a one-million-token context window. Moonshot says its technical group includes researchers behind Transformer-XL, RoPE, Group Normalization and ShuffleNet. arxiv.org https://arxiv.org/abs/1901.02860?utm source=openai&ref=runtimewire The proposed cloud agreements would give Yang a way to monetize that research without withdrawing K3's weights. Moonshot can use broad availability to attract developers while charging through managed inference, subscriptions and APIs. That structure also lets the cloud providers add another model to their catalogs without funding its training from the beginning. Moonshot's own K3 report is more measured than some of the comparisons surrounding the launch. Its authors say the model still trails the strongest proprietary systems in their evaluation, even as they claim leading results against other open and closed models in several categories. Reuters reported https://www.reuters.com/business/retail-consumer/chinas-moonshot-talks-with-microsoft-amazon-google-over-k3-revenue-sharing-2026-08-26/?ref=runtimewire that separate evaluations from Arena.ai and Artificial Analysis placed K3 strongly in web development and complex multi-step tasks. Benchmarks can establish technical interest; they do not establish demand at enterprise scale. A well-funded push for distribution Moonshot raised more than $2 billion in May, Reuters reported. TechCrunch reported https://techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets/?ref=runtimewire that Long-Z Investments, the venture arm of Meituan, led a roughly $2 billion financing at a reported $20 billion valuation, with Tsinghua Capital, China Mobile and CPE Yuanfeng participating. Alibaba, Tencent, HongShan, ZhenFund, IDG Capital and 5Y Capital have also been reported as Moonshot backers. techcrunch.com https://techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets/?utm source=openai&ref=runtimewire That capital financed the expensive part of Yang's strategy: training increasingly large models and securing enough computing capacity to serve them. Cloud distribution addresses the next constraint. Moonshot needs paying usage large enough to support continued model development, especially if it pursues a Hong Kong listing as Reuters has reported. K3's release already exposed how quickly demand can outrun capacity. Moonshot paused new subscriptions shortly after the July launch and said its GPUs were under strain. A managed-cloud strategy spreads inference across partners that have spent years building global capacity, although the proposed revenue share would leave Moonshot dependent on platforms that also distribute competing models. The US relationship comes with a political veto The negotiations are unfolding while US officials are considering tighter restrictions on Chinese AI developers. Treasury Secretary Scott Bessent said in July that Moonshot could be added to a trade blacklist. US officials have also accused Moonshot of using Anthropic's Fable model and acquiring restricted Nvidia chips https://www.reuters.com/world/us/us-accuses-chinas-moonshot-stealing-anthropics-fable-latest-ai-model-2026-06-30/?ref=runtimewire , according to Reuters. Moonshot rejected the distillation allegation, telling China's National Business Daily that K3's gains came from changes to its underlying architecture. As RuntimeWire reported in July /article/bessent-chinese-ai-sanctions-kimi-k3-distillation , Bessent's warning came before Moonshot released K3's weights. That uncertainty is part of the commercial negotiation. Microsoft, Amazon and Google would gain a model that has attracted developer attention and competes on inference cost. They would also assume compliance, reputational and continuity risks if Washington restricts Moonshot after customers have built systems around K3. For Yang, the cloud talks mark a shift from proving that Moonshot can build a frontier-scale open-weight model to proving that Moonshot can collect revenue from its use abroad. A signed agreement would give K3 a route into US enterprises without forcing Moonshot to close the model. The reported 30% request shows Yang intends to keep a meaningful share of the economics, even when an American cloud provider owns the customer and pays the infrastructure bill.