Chinese AI companies are systematically extracting capabilities from U.S. frontier models through distillation, gaining an advantage in open-weight local AI that U.S. firms cannot legally replicate.

Chinese AI labs DeepSeek, Moonshot, and MiniMax generated more than 16 million exchanges with Anthropic's Claude model through roughly 24,000 fraudulent accounts to extract reasoning, coding, and tool-use capabilities, with Anthropic describing this as systematic distillation occurring within weeks of the model's release.

The share of queries that local AI models can accurately answer rose from 23 percent in 2023 to 71 percent in 2025, enabling users to run open-weight models directly on laptops or smartphones without internet connection, with answers processed by the device's own chip rather than remote data centers.

Chinese firms are distilling and compressing U.S. frontier models to run on cheap hardware and shipping results globally, including back to the United States, while U.S. firms are constrained by terms of service prohibiting use of model outputs to train competing systems, creating an asymmetry where companies bearing the expense of frontier research are subsidizing competitor product lines.