Moonshot AI Released Kimi K3 Using Banned U.S. Chips and Illicit Data, Charging $3 per Million Tokens
Updated
Updated · Council on Foreign Relations · Jul 31
Moonshot AI Released Kimi K3 Using Banned U.S. Chips and Illicit Data, Charging $3 per Million Tokens
3 articles · Updated · Council on Foreign Relations · Jul 31
Summary
July 16 marked Moonshot AI’s release of Kimi K3, with model weights posted publicly on Monday, making it one of China’s strongest open-weight AI models.
White House allegations say K3 was trained with banned U.S. AI chips in Thailand and data illicitly extracted from leading U.S. labs through large-scale distillation attacks.
K3 still appears roughly 6 to 8 months behind top U.S. systems and cannot run on ordinary laptops, leaving it dependent on advanced cloud chips that are mostly made by U.S. firms.
Moonshot charges $3 per million tokens for cloud access, a price the report says exceeds some versions of flagship models from OpenAI and Anthropic despite K3 being free to download.
The report argues those tactics heighten IP-theft and security concerns and could push Washington toward cutting Chinese AI models off from U.S. businesses, cloud services and chips.
Can export controls truly stop China's AI rise when covert distillation lets rivals clone frontier models for pennies?
If open-weight AI can autonomously hack systems, how long until a stripped-down foreign model triggers a global cyber crisis?
Could the urgent push to block foreign AI inadvertently destroy the open-source innovation ecosystem that tech giants rely on?
Kimi K3’s 2.8 Trillion Parameters: The AI Model at the Center of U.S.-China Chip Smuggling and IP Theft Allegations
Overview
In July 2026, Moonshot AI launched Kimi K3, a massive open-weight AI model that quickly narrowed the performance gap with Western rivals and caused a surge in user demand, even forcing the company to pause new subscriptions. This breakthrough drew intense scrutiny from US officials, who accused Moonshot AI of bypassing export controls by remotely accessing restricted Nvidia chips in Thailand and of stealing intellectual property from Anthropic’s Claude Fable 5. While experts doubted the technical feasibility of such rapid distillation, the controversy sparked a global debate over open-weight AI, with Nvidia and Google Cloud defending openness and the Chinese government rejecting US allegations as unfounded.