Sources: Moonshot AI is in early talks over revenue-sharing agreements with Microsoft, Amazon, and Google to host Kimi K3, and is seeking up to a 30% share
Context & Ripple Effects
Moonshot’s commercial strategy has been moving alongside its capital and compute needs: it was reported to be seeking more Nvidia Blackwell chips for Kimi K4 after training K3 on Nvidia hardware, while also pursuing a sharply higher fundraising valuation. A hosting agreement would turn K3 distribution into a potential source of ongoing economics rather than solely a model-development expense.
The reported terms parallel Alibaba’s proposed revenue share for heavy users of its next Qwen model, suggesting Chinese model makers are testing ways to monetize broad access while cloud platforms supply the infrastructure. The talks remain unconfirmed.
First-order effects
- If agreements are reached, Microsoft, Amazon and Google would be able to host Kimi K3 under terms that allocate Moonshot as much as 30% of resulting revenue.
- Moonshot would gain prospective cloud distribution for K3 while retaining a direct claim on commercial revenue, complementing its push for more Blackwell capacity for Kimi K4.
Second-order effects
- The three hyperscalers would need to weigh Kimi K3’s distribution value against a revenue split that reduces the portion retained for hosting and related services.
- Alibaba’s reported Qwen approach and Moonshot’s negotiations put revenue-sharing terms, rather than access alone, at the center of competition to host commercially used models.
Third-order effects
- If this model spreads, AI infrastructure platforms may increasingly act as revenue-sharing distributors for third-party models, not merely sellers of compute.
- Model makers with valuable workloads could gain leverage to fund training through downstream usage economics, while cloud providers differentiate on the commercial terms attached to model access.
The trend: AI model monetization is shifting toward shared application revenue between model developers and the cloud platforms that distribute their workloads.