AIsaAIsaStudio English 打开 AIsa Studio

美国拟限制中国境外租用 AI 算力,Reuters 揭示中国依赖蒸馏美国模型的平行供应链

原标题:The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output

美国商务部工业与安全局发布临时规则,将远程访问算力纳入出口管制,要求新加坡、马来西亚、泰国、印尼和巴西的数据中心运营方购买 Nvidia 等美国先进 AI 芯片需获得政府许可,美国公司子公司及仅运行商业推荐系统的数据中心可获豁免。消息公布后 Nvidia 股价一度下跌 2.9%,费城半导体指数最多跌 5.5%。另据 Reuters 调查,中国开发者通过蒸馏美国模型大规模生产训练数据来弥补本土算力缺口,官方政策也在推动这类降算力技术,OpenAI 和 Anthropic 长期指控此类行为违反服务条款,Anthropic 称现已成功拦截大多数蒸馏攻击。

为什么值得看

报道把美国远程算力新规与 Reuters 对中国模型蒸馏产业链的调查放在一起,读者可以同时看到管制动作和被管制方的替代路径。

阅读原文 ↗

本页的摘要、推荐理由与解读出自 AIsa Studio 的 AI 风向智能体,原文版权归原作者所有。

原文

The Bureau of Industry and Security (BIS) at the US Commerce Department now wants to shut down that workaround with an interim rule. It aligns remote-access rules with existing physical export bans. Data center operators in Singapore, Malaysia, Thailand, Indonesia, and Brazil will need a US government license to buy advanced AI chips from US manufacturers like Nvidia, Bloomberg reports. Applications will be reviewed on a case-by-case basis.

Subsidiaries of US-headquartered companies are exempt from the license requirement, as are shipments to data centers running only commercial recommendation systems for online retailers or social media platforms. The rules take effect once published in the Federal Register. Nvidia's stock fell as much as 2.9 percent after the announcement, while the Philadelphia Semiconductor Index dropped up to 5.5 percent.

Illegal chip diversions and Jensen Huang's downfall

The political trigger, though, was suspected illegal smuggling of physical chips. The new rule comes just days after the abrupt resignation of Nvidia CEO Jensen Huang, who had admitted to an affair with Alice Huang, a former manager at Singapore-based Megaspeed International. US authorities suspect Alice Huang, who now lives in China, of espionage. They are investigating Megaspeed for possible export-control violations after banned Nvidia chips reportedly turned up in a Chinese data center linked to her. Megaspeed denies any wrongdoing, and President Donald Trump said Monday that Jensen Huang did not compromise national security.

The scandal still set off an open power struggle in Washington over lax enforcement of chip sanctions. Commerce Secretary Howard Lutnick had delayed a Trump-authorized measure against illegal chip diversions, according to Axios, prompting Treasury Secretary Scott Bessent to publicly accuse him of inaction. The Commerce Department's announcement now states with unusual bluntness that the draft rule was written by and at the direction of the Treasury Department and the National Security Council.

Reuters investigation maps China's distillation industry

A separate exclusive Reuters investigation shows how Chinese developers cope with the years-long shortage of top-tier domestic hardware. By reviewing more than 300 Chinese papers, patents, job postings, and government documents, Reuters shows for the first time how heavily China relies on model distillation as a stopgap. What started years ago as an academic workaround has grown into an industrial sector with a highly developed division of labor.

The technique does not make powerful chips like Nvidia's unnecessary, but it drastically reduces developers' own compute load because a single expensive query to a US teacher model replaces repeated computation. Chinese firms now build their own software to automate the extraction of answers, code, and problem solutions from leading US models. Specialized workers then process the material to train smaller, far cheaper local student models.

Research papers show how deep this practice already runs. As early as November 2024, researchers at Shanghai Jiao Tong University created 10,750 training examples for Alibaba's Qwen2.5-72B. They drew on about 5,000 responses from OpenAI's o1 and translations from GPT-4o mini, though they themselves warned at the time of a looming dependence on outside innovation. Alibaba described its systematic use of GPT-4 and GPT-4o for foreign-language training data in April 2025. In February 2026, Jiao Tong University and food-delivery company Meituan used GPT-4o to train smaller models for coordinating coding agents.

Policy directives from Beijing are driving this growth. The government wants to deploy smaller, low-cost AI systems across China's manufacturing sector. Official government documents do not directly instruct anyone to copy American models. Still, an "AI+Manufacturing" package published by eight ministries in January and initiatives in Shanghai and Shenzhen explicitly promote distillation and other techniques that reduce computing needs. In practice, this state-driven push for efficiency drives Chinese developers to tap the most powerful US models as a data source at scale.

Ahead of the APEC summit in Shenzhen, this industrial-scale practice is becoming a geopolitical flashpoint. Distillation is not illegal, but it violates the terms of service of US developers like OpenAI and Anthropic. Both companies have long leveled accusations against Chinese AI labs, documenting the massive volumes of extracted data in detailed reports and deploying technical countermeasures. Anthropic recently said it now successfully blocks most such distillation attacks. These accusations were also covered in our Frontier Radar #4 on Chinese models.

US officials also confronted Chinese regulators in Chengdu as early as July because Washington sees mass distillation as the main reason for China's rapid AI catch-up, according to Reuters. Chinese insiders also point to their own advances in algorithms and model design. Treasury Secretary Bessent is now threatening sanctions and Entity List designations if covert distillation attacks cross the line into intellectual property theft. Chinese developers, uncertain about the rules, recently asked Beijing's Ministry of Industry for clear guidelines on using foreign models.

原文版权归 The Decoder 及原作者所有,本站转载仅供阅读。如需删除,请联系 support@aisa.one。

事件综述

美国将海外数据中心远程租用算力纳入对华出口管制

2026年11月16日,Ars Technica 报道特朗普政府准备打击允许中国远程访问受限高端 AI 芯片的海外数据中心,美国商务部称此举旨在封堵漏洞——此前中国、俄罗斯、伊朗等国可通过租用印尼、马来西亚、新加坡、泰国和巴西数据中心的芯片绕过出口管制,用于训练和运行可能支持军事与情报活动、威胁美国国家安全的先进 AI 模型。11月17日,The Decoder 报道美国商务部工业与安全局发布临时规则,将远程访问算力纳入出口管制,要求新加坡、马来西亚、泰国、印尼和巴西的数据中心运营方购买 Nvidia 等美国先进 AI 芯片需获得政府许可;美国公司子公司以及仅运行商业推荐系统的数据中心可获豁免。消息公布后,Nvidia 股价一度下跌 2.9%,费城半导体指数最多下跌 5.5%。另据 The Decoder 提及的 Reuters 调查,中国开发者通过蒸馏美国模型大规模生产训练数据以弥补本土算力缺口,官方政策也在推动这类降算力技术;OpenAI 和 Anthropic 长期指控此类行为违反服务条款,Anthropic 称现已成功拦截大多数蒸馏攻击。目前事件进展为规则已发布但报道未说明实施细节或中方回应。

其他来源的报道 1