美国拟封堵中国租用海外AI算力,中国已建平行供应链

Nano Banana Pro,提示词由 THE DECODER 提供
中国企业一直通过在第三国合法租用 Nvidia 算力来绕开美国的出口禁令。在 Nvidia CEO 黄仁勋辞职之后,美国正在堵上这一漏洞。一项调查还揭示,中国如何通过一整条围绕模型蒸馏建立起来的产业链,缓冲了持续数年的芯片禁运。
多年来,美国一直试图通过严格的出口管制切断中国进口先进 AI 处理器的渠道。中国企业的应对方式是利用一个完全合法的漏洞——通过基于云的远程访问,从获准接收美国硬件的外国数据中心租用其所需的算力。
美国商务部工业与安全局(BIS)现在希望以一项临时规则关闭这一变通做法。该规则使远程访问的相关规定与现有的实体出口禁令保持一致。据彭博社报道,新加坡、马来西亚、泰国、印度尼西亚和巴西的数据中心运营商,若要从 Nvidia 等美国制造商处购买先进 AI 芯片,将需要获得美国政府许可证。申请将逐案审批。
美国总部企业的子公司可豁免许可证要求,向仅运行在线零售商或社交媒体平台商业推荐系统的数据中心发货同样豁免。这些规则一经在《联邦公报》上发布即生效。消息公布后,Nvidia 股价一度下跌 2.9%,费城半导体指数跌幅最高达 5.5%。
非法芯片转运与黄仁勋的倒台
不过,真正的政治导火索是涉嫌非法走私实体芯片。这项新规出台的几天前,Nvidia CEO Jensen Huang 突然辞职,他此前承认与 Alice Huang 有婚外情,后者曾是总部位于新加坡的 Megaspeed International 的经理。美国当局怀疑现居中国的 Alice Huang 从事间谍活动。在据报被禁的 Nvidia 芯片出现在一处与她有关联的中国数据中心后,美方正就 Megaspeed 可能违反出口管制展开调查。Megaspeed 否认存在任何不当行为,美国总统 Donald Trump 周一表示,Jensen Huang 并未危害国家安全。
但这起丑闻仍在华盛顿引发了一场围绕芯片制裁执法不力的公开权力斗争。据 Axios 报道,美国商务部长 Howard Lutnick 推迟了一项由 Trump 授权的针对非法芯片转运的措施,促使财政部长 Scott Bessent 公开指责其不作为。商务部在公告中罕见直白地表示,该规则草案由财政部和国家安全委员会起草,并按其指示制定。
另一篇路透社独家调查则揭示了中国的开发者如何应对多年来顶级国产硬件的短缺。通过梳理 300 多篇中国论文、专利、招聘信息和政府文件,路透社首次展示出中国对模型蒸馏作为权宜之计的依赖程度之深。这项数年前作为学术变通做法起步的技术,如今已发展成一个分工高度发达的产业部门。
这项技术并不会让 NVIDIA 这类强大芯片变得不再必要,但它大幅降低了开发者自身的算力负载,因为向美国教师模型发起一次昂贵的查询,就能替代反复的计算。中国企业如今自行开发软件,自动化地从美国领先模型中提取答案、代码和题目解答。随后由专门的人员处理这些素材,用于训练更小、成本低得多的本地学生模型。
研究论文显示出这种做法已经深入到何种程度。早在 2024 年 11 月,上海交通大学的研究人员就为阿里巴巴的 Qwen2.5-72B 构建了 10,750 条训练样本。他们使用了 OpenAI o1 的约 5,000 条回复以及 GPT-4o mini 的翻译结果,尽管他们当时自己就警告称,对外部创新的依赖正在逼近。阿里巴巴在 2025 年 4 月描述了其系统性地使用 GPT-4 和 GPT-4o 获取外语训练数据。2026 年 2 月,上海交通大学与外卖公司美团使用 GPT-4o 训练用于协调编程智能体的更小模型。
来自北京的政策指令正在推动这一增长。政府希望在中国制造业中部署更小、成本更低的人工智能系统。官方政府文件并未直接指示任何人复制美国模型。不过,八个部委在 1 月发布的“AI+制造”方案以及上海和深圳的举措,明确鼓励蒸馏及其他降低算力需求的技术。实际上,这种由国家推动的效率追求,正促使中国开发者大规模地把最强大的美国模型当作数据来源。
据路透社报道,美国官员早在7月就在成都与中国监管机构进行了交涉,因为华盛顿将大规模蒸馏视为中国AI快速追赶的主要原因。中国业内人士也指出,他们在算法和模型设计方面取得了自身进展。美国财政部长贝森特如今威胁称,如果隐蔽的蒸馏攻击越界构成知识产权盗窃,将实施制裁并列入实体清单。中国开发者因规则不明,最近要求北京工信部就使用外国模型给出明确指引。
没有炒作的 AI 新闻——由人工甄选
可完整访问 THE DECODER 上的每一篇文章
参与评论与社区讨论
每周通过邮件收取 AI 新闻回顾
每年6次:“AI Radar”——深入剖析最重要的 AI 议题
每日 AI 新闻,始终紧跟最新动态
我们完整的十年存档
由一支拥有 10 年以上 AI 领域经验的团队报道
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output)
Nano Banana Pro prompted by THE DECODER
Chinese companies have been skirting US export bans by legally renting Nvidia computing power in third countries. After Nvidia CEO Jensen Huang's resignation, the US is closing that loophole. An investigation also reveals how China has cushioned the years-long chip embargo through an entire industry built around model distillation.
For years, the US has tried to cut China off from importing advanced AI processors through strict export controls. Chinese companies responded by exploiting an entirely legal loophole, renting the computing power they needed through cloud-based remote access from foreign data centers that were allowed to receive US hardware.
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.
AI News Without the Hype – Curated by Humans
Full access to every article on THE DECODER
Join the comments and community discussions
A weekly AI news recap via mail
6x/year: "AI Radar" — deep dives on the AI topics that matter most
Daily AI news, always up to date
Our full ten-year archive
Covered by a team with 10+ years in AI
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(The US wants to block China from renting AI power abroad but China already built a parallel supply chain from US AI output),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。