中美两边都在卡AI模型出口，但权重早就挂在Hugging Face上了，这管制能执行吗？看看Coinbase和DeepSeek的实际依赖就知道了。
Axios报道特朗普政府考虑通过采购规则和实体清单限制美国使用中国AI模型，源于Kimi K3发布。Financial Times称中国正咨询阿里巴巴、字节跳动和智谱，拟收紧训练数据出口和禁止海外下载开源模型权重，但保留API访问。此举与习近平在世界AI大会上的开放表态形成对比。Coinbase已使用中国模型削减50%AI支出，DeepSeek在OpenRouter上token份额超过谷歌、Anthropic和OpenAI各自份额。问题在于已公开的权重难以追溯限制，下一代模型可能面临更严管控。
As Washington debates banning Chinese open-weight …
As Washington debates banning Chinese open-weight AI models, Beijing is now weighing export curbs of its own. The tit-for-tat is speeding up.
Axios reported over the weekend that the Trump administration is considering procurement rules, Entity List additions and security advisories to limit US use of Chinese models after Kimi K3's release. Today, the Financial Times says China is consulting Alibaba, ByteDance and Zhipu on tighter controls - limiting overseas transfers of training data and blocking foreign downloads of open model weights.
China would still allow overseas API access, but not weight downloads for customisation. This lands just days after Xi Jinping reaffirmed openness in AI at the World AI Conference, exposing a growing gap between public messaging and internal control talks.
Both sides are running into the same enforcement wall. Once weights are public on Hugging Face or GitHub, restrictions don't scale. Coinbase has already cut AI spend by 50% using Chinese models. DeepSeek now takes more OpenRouter token share than Google, Anthropic or OpenAI individually. The issue isn't access - it's how to manage the dependencies that are already baked in.
We're heading toward a bifurcated AI world where what you can download depends on where you sit. Current weights are already out there. The question is what happens with the next generation of models, and whether either government can actually enforce the controls they're discussing.