Microsoft reportedly evaluating Moonshot's Kimi K3 for Copilot to cut $600M in inference costs

Microsoft is reportedly evaluating Moonshot AI's Kimi K3 for integration into its Copilot service on Azure, according to industry reports. If adopted, the Chinese open-weight model could displace existing OpenAI and Anthropic models in parts of Copilot's stack, with estimates suggesting savings of up to $600 million in AI inference costs. Kimi K3 is a 2.8-trillion-parameter multimodal mixture-of-experts model, released July 16, that recently topped a key coding benchmark.

The evaluation is striking given Microsoft's deep, multibillion-dollar entanglement with OpenAI — the prospect of Microsoft even testing a Chinese open-weight alternative for its flagship AI product underscores how far the open-weights surge has shifted the economics. For Microsoft, the appeal is straightforward: Copilot serves enormous inference volume, and an open-weight model it can host on its own Azure infrastructure removes per-token licensing costs and gives it control over deployment.

This fits squarely with the week's dominant theme — Chinese open models (Kimi K3, Qwen 3.8, DeepSeek V4, GLM-5.2) pressuring the economics that justified US labs' pricing. It also parallels Hugging Face's pivot to GLM-5.2 and reflects enterprises increasingly treating frontier capability as commoditized.

Caveats are substantial: this is reported evaluation, not adoption, and Microsoft faces obvious geopolitical, security, and contractual (OpenAI-relationship) complications in shipping a Chinese model in Copilot. The $600M figure is an estimate. Readers should watch whether Microsoft confirms any pilot, how it would address data-security concerns around a Chinese-origin model, and whether OpenAI responds to the competitive signal.