中国最强AI模型由禁售NVIDIA芯片训练

一项最新研究称,中国企业开发的领先开放权重模型,是在美国政府禁止向其最大地缘政治对手出口的 NVIDIA 芯片上完成训练的。
中国前沿实验室显然普遍能够获取 Blackwell GPU,这凸显出这家领先芯片设计商所面临问题的严重程度。NVIDIA CEO 黄仁勋本周离开公司,此前他承认与 Megaspeed 高管 Alice Huang 有婚外情,而 Megaspeed 是一家被指涉将 NVIDIA 芯片走私到中国的公司。Alice Huang 还被指为中国情报机构的特工。
尽管黄仁勋与一名客户以及两名员工的关系引发了最多的公众关注,但 NVIDIA 可能因违反出口管制而承担的公司责任,很可能也是其决定与黄仁勋分道扬镳的因素之一。
美国的出口管制旨在保护本国在 AI 能力上相对于中国的领先地位,但黄仁勋长期以来一直寻求打开这一市场。Blackwell GPU 被禁止出口至中国或中资企业,而特朗普政府允许 NVIDIA 出口性能较弱的 H200 芯片,条件是货物须经由迂回的运输路线,从而让美国可以主张征收 25% 的关税。
Will DePue 曾是 OpenAI 研究员,现领导非营利研究组织 Unemployment Labs。他分析了 DeepSeek、Moonshot AI、MiniMax 和蚂蚁集团公开发布的大语言模型的权重——即其独特结构。权重仍会反映出训练这些模型所用芯片的配置,DePue 找到了有力证据,表明最受关注的模型——包括 Moonshot 的 Kimi K3 版本和 DeepSeek 的 V4 模型——带有在 Blackwell 芯片上训练留下的指纹。
白宫科技顾问 Michael Kratsios 今年早些时候援引美国情报来源称,Moonshot 使用了被禁止出口的 Nvidia 芯片。非营利研究机构 Epoch AI 估计,2024 年和 2025 年约有 200 万枚先进芯片通过刻意隐藏最终目的地的方案流入中国。
这并不意味着这些模型完全是在 Nvidia 芯片上训练的,也不意味着必然违反了出口管制法律;这些芯片可能位于境外数据中心——事实证明,这是中国企业获取算力的常用途径,并在马来西亚带动了一场产业革命。Epoch 的数据估计,2024 年和 2025 年有 27.5 万枚 Nvidia AI 加速器进口到马来西亚。
然而,出口禁令表面上的目标正是阻止这种情况:中国 AI 实验室用最强大的芯片进行训练。对华强硬派长期以来一直呼吁对“谁在使用芯片”实施更严格的管控,而不仅仅是管控芯片归谁所有或安装在哪里。
“如果我们关心美国的技术优势,那么在我看来,你不允许中国公司在中国大陆或香港、澳门获取芯片,却又说,实际上,你只要跑到马来西亚就可以为所欲为,这毫无道理,”乔治城大学安全与新兴技术中心研究员 Sam Bresnick 今年夏天对 TechCrunch 说。
由 Alice Huang 领导的 Megaspeed 公司多年来通过其子公司 Speedmatrix 处于这一贸易的中心,并一直是美国和新加坡政府调查的对象。周五,Bloomberg 还报道称,一家与 Alice Huang 有关联的公司参与了向上海某数据中心走私被禁 Nvidia 芯片的活动。
今年5月,做空研究机构 Culper 发布报告,认为 Nvidia 存在“中国问题”,因为其收入的很大一部分由中国需求驱动,或来自远程数据中心,或来自违反美国法律的走私。报告详述了 Megaspeed 与 Nvidia 的关系,并声称 Alice Huang 旗下的公司背后有被禁止获取美国芯片的中国企业支持。
讽刺的是,做空机构 Culper 押注 Nvidia 下跌,并非出于对美国政府出手打击的预期,而是聚焦于中国政府对禁令的回应——重新全力投入本土芯片产业。
如果中国成功,中国对 Nvidia 芯片的数十亿美元需求可能会枯竭。据报道,约有 60 万颗 H200 芯片囤积在香港,等待运往中国内地。尽管中国企业可能想要这些芯片,但观察人士对于它们为何迟迟未获进口许可并无共识。
斯坦福大学专注中美科技政策的研究学者 Graham Webster 表示,他不认同扣住这些芯片会推动本土生产的说法,因为芯片需求实在太大了。关于此次滞留的其他推测包括:为一项可能涉及进口大量美国商品的潜在贸易协议做准备,或是对芯片中可能存在的监控技术的担忧。
目前,美国企业在追踪客户如何使用 AI 芯片等受管制商品方面所承担的义务有限。不过,如果政府调查人员证实芯片走私确曾发生、且黄仁勋对此知情,那么黄仁勋与 Alice Huang 之间在个人和职业上的纠葛将给 Nvidia 带来大得多的麻烦。Nvidia 和黄仁勋本人都为 Megaspeed 的数据中心作了担保,称 Nvidia 官员曾实地走访这些数据中心,以确认其合法性。
美国政府对待 Nvidia 和 Megaspeed 的方式也正受到审视。据报,负责监管芯片出口许可相关机构的商务部长 Howard Lutnick 可能丢掉职位,此前有报道称,他手下的调查人员未能查出可疑的芯片买家,并且与 Nvidia 官员关系过于密切。
当你通过我们文章中的链接进行购买时,我们可能会赚取少量佣金。这不会影响我们的编辑独立性。
Tim Fernholz 是一名记者,报道科技、金融和公共政策领域。他长期跟踪报道私营航天产业的崛起,著有《Rocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race》。此前,他在全球商业新闻网站 Quartz 担任高级记者十余年,职业生涯始于华盛顿特区的政治记者。你可以通过发送邮件至 [email protected] 联系 Tim 或核实其联系请求,也可通过 Signal 向 tim_fernholz.21 发送加密消息。
正在加载下一篇文章
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(Export-banned Nvidia chips are behind China’s most powerful AI models — uh oh, Jensen!)
The leading open-weight models developed by Chinese companies were trained on Nvidia chips that the U.S. government restricted from shipping to its largest geopolitical rival, a new research effort claims.
The apparently widespread access to Blackwell GPUs among Chinese frontier labs shows the magnitude of the problem facing the leading chip designer. Nvidia CEO Jensen Huang left the company this week after admitting to his affair with Alice Huang, executive of Megaspeed, a company implicated in smuggling Nvidia chips into China. Alice Huang has also been alleged to be an agent of China’s intelligence service.
While Jensen Huang’s relationships with a customer and two employees have drawn the most public interest, Nvidia’s potential corporate liability for export control violations likely factored into its decision to part ways with Huang.
US export controls are intended to protect the country’s leadership in AI capabilities against China, but Huang has long sought to open up that market. Blackwell GPUs are banned from export to China or Chinese-owned firms, while the Trump administration allows Nvidia to export the H200, a less powerful chip, provided shipments take a circuitous shipping route that allows the US to claim a 25% tariff.
Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs, analyzed the weights — the unique structure — of large language models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group. The weights still reflect the configurations of the chips used to train them, and DePue was able to find compelling evidence that the most closely watched models, including Moonshot’s Kimi’s K3 releases and DeepSeek’s V4 models, bear fingerprints from training on Blackwell chips.
White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned Nvidia chips, citing US intelligence sources. Epoch AI, a nonprofit research organization, estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination.
That doesn’t mean that these models were trained entirely on Nvidia chips, or that export control laws were necessarily violated; the chips could have been in an offshore data center, which has proven a popular way for Chinese firms to access compute and driven an industrial revolution in Malaysia. Epoch’s data estimates that 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025.
However, the ostensible goal of the export bans is to prevent exactly that: Chinese AI labs training on the most powerful chips. China hawks have long called for stronger controls on who uses chips, not just who owns them or where they are installed.
“If we care about the U.S. tech edge, it doesn’t make any sense to me that you would not allow Chinese companies to access chips in mainland China or Hong Kong and Macau, but then say, actually, if you just go over to Malaysia, you can do whatever you want,” Sam Bresnick, a researcher at the Georgetown Center for Security and Emerging Technology, told TechCrunch this summer.
Megaspeed, the company led by Alice Huang, has been at the center of that trade for several years through its subsidiary Speedmatrix, and has been the subject of investigations by both the U.S. and Singaporean government. Friday, Bloomberg also reported that a company with links to Alice Huang was involved in smuggling banned Nvidia chips to a data center in Shanghai.
In May, the short-selling research firm Culper published a report arguing that Nvidia had a “China problem” because a significant share of its revenue was driven by Chinese demand, either through remote data centers or smuggling in violation of U.S. law. It detailed Megaspeed’s relationship with Nvidia and alleged that Alice Huang’s companies are backed by Chinese firms that are barred from accessing U.S. chips.
Ironically, short-seller Culper’s bet against Nvidia wasn’t driven by the expectation of a U.S. government crackdown. Instead, it was focused on the Chinese government’s response to the bans, which was to recommit to its domestic chip industry.
If China succeeds, billions of dollars in Chinese demand for Nvidia chips could dry up. Some 600,000 H200 chips are reportedly stockpiled in Hong Kong, awaiting shipment into China. While Chinese companies might want the chips, observers have no consensus on why they’ve haven’t been cleared for import.
Graham Webster, a research scholar at Stanford who focuses on U.S.-China technology policy, said he doesn’t buy the idea that holding the chips back would drive domestic production, since demand for chips is so large. Other theories for the hold-up include preparation for a potential trade deal that would involve importing a large volume of U.S. goods, or concerns about potential surveillance technology in the chips.
For now, U.S. companies have limited obligations to track what their customers do with controlled goods like AI chips. However, Jensen Huang’s personal and professional entanglements with Alice Huang create a much bigger problem for Nvidia if government investigators confirm that chip smuggling took place and that he was aware of it. Nvidia and Huang have both vouched for Megaspeed’s data centers, saying Nvidia officials had visited them to confirm they were legitimate.
The U.S. government’s treatment of Nvidia and Megaspeed is also under scrutiny. Secretary of Commerce Howard Lutnick, who oversees the agencies that licensing chip exports, could reportedly lose his job after reports that his investigators failed to uncover suspicious chip buyers and were too cozy with Nvidia officials.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Tim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author of Rocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race. Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C. You can contact or verify outreach from Tim by emailing [email protected] or via an encrypted message to tim_fernholz.21 on Signal.
Loading the next article
原文为英文,正文由自动抓取并保留原文。
原文发布于 TechCrunch AI(Export-banned Nvidia chips are behind China’s most powerful AI models — uh oh, Jensen!),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。