Moonshot Wants Azure, AWS, and Google Cloud to Host Kimi K3 — and Up to 30% of the Revenue
China's Moonshot AI is negotiating revenue-sharing deals that would put Kimi K3 on all three major US clouds. It would be the first major pact of its kind between a Chinese lab and American hyperscalers.
Moonshot AI is in talks with Microsoft, Amazon, and Google to host its Kimi K3 model on Azure, AWS, and Google Cloud — with Moonshot seeking up to a 30% cut of the revenue those services generate, according to a Reuters exclusive published today.
No Chinese AI lab has ever struck a major revenue-sharing deal with a US hyperscaler. Plenty of open-weight Chinese models already run on American clouds — DeepSeek and Qwen variants are available through Bedrock and Vertex — but those are permissively licensed weights the clouds serve without paying the model’s creator. What Moonshot is negotiating is different: a commercial arrangement where the lab gets paid per token, the way Anthropic and OpenAI do through their cloud partnerships.
The leverage comes from demand. Kimi K3 has become one of the most-used open models among US developers on price-performance, and the clouds are seeing real consumption they currently monetize without sharing anything upstream. Moonshot already proved the template in July with a revenue-sharing agreement with Chinasoft International, a Chinese IT services firm. Now it wants the same structure from the three biggest platforms in the world.
The unresolved issues are exactly the ones you would expect: how the revenue split works, what data access Moonshot gets, and how token usage gets audited across clouds that treat consumption metrics as competitive secrets. Reuters’ sources stress the talks are early and may not produce a deal.
The geopolitics are the real story. Washington has spent three years walling off China from advanced AI chips, yet Chinese models keep landing on American infrastructure because they are cheap and good. A signed deal would formalize a strange loop: US companies paying a Chinese lab for a model that was trained under US chip sanctions, running on US datacenters. Expect scrutiny from the same lawmakers who pushed the export controls — a formal revenue relationship is far more visible than passively hosting open weights.
For developers, the practical upside is straightforward: first-party Kimi K3 endpoints on the cloud you already use, with SLAs and enterprise compliance instead of self-hosted weights or third-party inference providers. If even one of the three deals closes, the others follow fast — no hyperscaler will concede a popular model to a competitor’s catalog.
Watch whether Microsoft moves first. Azure has been the most aggressive about stocking third-party models, and it has the most to prove after building its catalog almost entirely around OpenAI.
Sources
Related reading
- AI Models Moonshot AI Releases Kimi K3, the Largest Open-Weight Model Ever at 2.8 Trillion Parameters
- AI Models DeepSeek Nears a $74 Billion Valuation as Its ~$7B Pre-IPO Round Closes — With CATL and China's National AI Fund Aboard
- AI Models Alibaba's Qwen3.8-Max Beats GPT-5.6 Sol and Claude Fable 5 on Key Benchmarks