TLDR
- U.S. agencies accused six Chinese AI firms, including Moonshot, DeepSeek and Alibaba, of large-scale model distillation.
- The firms allegedly extracted billions of tokens through millions of requests to GPT, Claude, Gemini and Grok.
- Agencies said multiple accounts and proxy services helped the companies avoid detection and platform usage limits.
- Moonshot’s Kimi K3 gained attention after competing strongly with leading U.S. AI models on coding benchmarks.
- Chinese AI and semiconductor stocks weakened as investors reassessed competition between U.S. and Chinese AI developers.
U.S. security agencies have accused six Chinese artificial intelligence companies, including DeepSeek, Alibaba and Moonshot, of using leading American AI models to improve their own systems. The NSA, FBI and CISA said the companies extracted billions of tokens through millions of requests since late 2024.
The agencies said the activity involved Claude, GPT, Gemini and Grok models. They said the firms used proxy services and multiple accounts to avoid detection and platform limits. U.S. officials described the activity as malicious knowledge distillation and said Chinese authorities likely knew about it. The agencies said the activity began in late 2024 and continued across multiple model providers through coordinated accounts and proxy networks globally.
Moonshot Draws Attention After Kimi K3 Release
Moonshot drew wider attention after releasing Kimi K3, an open-weight model that ranked strongly on a major coding benchmark in July. The model performed well against systems from Anthropic and OpenAI, raising fresh questions about the pace of Chinese AI development.
Kimi K3 also showed that Moonshot could build a competitive model without matching the spending levels of major U.S. AI companies. CoinDesk reported that the release challenged assumptions about model costs and whether leading AI systems would remain concentrated among large American developers.
Agencies Describe Large-Scale Distillation Activity
Distillation is a common AI training method. Developers send prompts to a stronger model, collect its answers, and use that data to improve another system. U.S. agencies said the Chinese companies used the method at a much larger scale.
According to the agencies, the firms sent millions of carefully designed requests to American models. The goal was to capture coding, math, and reasoning abilities. Officials said the companies used technical methods to avoid controls, usage limits, and detection across several platforms.
Markets React to AI Competition Concerns
Chinese AI stocks fell after the reports and renewed concern about competition in the sector. Z.ai dropped about 27%, while MiniMax fell around 16%. Semiconductor shares also weakened across Asia as investors reassessed the outlook for AI spending and model development.
The selling spread to other markets. Bitcoin and other major cryptocurrencies fell, while South Korea’s Kospi fell another 3.5% the next session. Investors also reviewed whether large U.S. spending on chips and data centers would always lead to better AI models than lower-cost rivals from China.
