The Trump administration announced last week that it would allow China to purchase advanced artificial intelligence chips from the tech giant Nvidia. The decision, a reversal of the administration’s past restriction on chip sales, is a profound mistake.

特朗普政府上周宣布，允许科技巨头英伟达将先进人工智能芯片卖给中国。这一决定推翻了此前的芯片销售禁令，是一个严重的错误。

It cedes the United States’ greatest point of leverage in A.I.: control of the global computing power supply chain. Mr. Trump claims to champion American tech, releasing an “A.I. action plan” this week. But permitting these chip sales threatens American dominance in A.I., undermines U.S. tech companies and risks our national security — all in favor of one chipmaker’s near-term profits.

它放弃了美国在人工智能领域最大的筹码：对全球算力供应链的控制权。特朗普虽然声称捍卫美国科技，本周还发布了一份“人工智能行动计划”，但允许此类芯片销售将威胁到美国在人工智能领域的主导地位，伤害本土科技企业，并危及我们的国家安全——所有这些都仅仅是为了一个芯片制造商的短期利润。

Computer chips are the lifeblood of powerful A.I. systems. A.I. companies compete desperately to buy this hardware; many spend the vast majority of their funding on chips. (I do advisory work for some of these companies, including Anthropic.) American chipmakers, including Nvidia, sell more than 10 million of these chips annually. China relies on these chips to aid both its military and its A.I. companies, but can make only about 200,000 of its own per year, the Trump administration said last month.

计算机芯片是强大人工智能系统的命脉。人工智能公司不惜一切代价购买这些硬件设备；许多公司将绝大部分资金用于芯片的采购。（本人为包括Anthropic在内的其中一些企业提供顾问服务）包括英伟达在内的美国芯片制造商每年销售超过1000万片此类芯片。特朗普政府上个月表示，中国依赖这些芯片来辅助其军事和人工智能公司，但自主年产量仅约20万片。

In other words, Nvidia’s chips will give China’s A.I. ecosystem, and its government, just what it needs to surpass the United States in the most critical arenas. A.I. technology could soon transform military operations, potentially enabling better hacking and sophisticated drone warfare. Ample evidence suggests that Chinese military suppliers prefer Nvidia chips and use A.I. systems trained on U.S. chips. The stakes are not hard to grasp: We should not allow American troops and intelligence officers to be targeted by Chinese A.I. trained on Nvidia chips.

换言之，英伟达芯片将为中国的人工智能生态及其政府提供在关键领域超越美国所需的核心资源。人工智能技术或将迅速改变军事行动，有可能让黑客攻击变得更为高效，以及助力复杂的无人机作战。有充足的证据表明，中国的军工企业青睐英伟达芯片，并使用在美国芯片上训练的人工智能系统。利害关系不言而喻：我们不应让美国军队及情报人员成为基于英伟达芯片训练的中国人工智能系统的打击目标。

The first Trump administration understood these risks. Mr. Trump placed export controls on some chip-making equipment in order to hobble China’s A.I. chip industry, while the Biden administration — in which I served as the White House special adviser on A.I. — went further, restricting the sale of additional equipment and of Nvidia’s flagship chip, called the H100. Both administrations accused China of using advanced A.I. to modernize its military and to commit human rights abuses.

首届特朗普政府深谙此风险。他曾对部分芯片制造设备实施出口管制，以遏制中国人工智能芯片产业的发展，到了我担任白宫人工智能特别顾问的拜登政府则更进一步，限制了更多设备以及英伟达旗舰芯片H100的销售。两届政府均指责中国利用先进人工智能推进军事现代化及侵犯人权。

Last year, in response to these restrictions, Nvidia created a chip called the H20, which is designed to elude U.S. export controls so it can be sold in China. The controls focused on processing power, so Nvidia gave the H20 chip mediocre processing but large amounts of high-bandwidth memory. That memory allows the H20 to outperform the H100 when running, as opposed to training, A.I. systems — a process called inference. Inference is becoming increasingly important: One analysis suggests that it will make up more than 70 percent of A.I. needs by 2026. This memory, and other components of the H20, could have been used to make H100 chips for American companies.

去年，英伟达为规避管制，专为中国市场开发了一款名为H20的芯片。管制措施主要针对算力进行限制，故H20虽计算性能平庸，却配备了大容量高带宽内存，使其在AI系统推理（而非训练）环节反超H100。推理正变得越来越重要：一项分析表明，到2026年，它将占到人工智能需求的70%以上。这些内存及H20其他组件本可以用来为美国企业制造H100芯片。

Before reversing itself, the Trump administration rightly blocked H20 sales to China in April. A White House official said that there was “bipartisan and broad concern” about how China would use the chips.

政策逆转前，特朗普政府在4月正确地禁止了H20芯片的对华销售。一位白宫官员表示，“两党普遍担忧”中国使用这些芯片的意图。

The controls appeared to be working. DeepSeek, China’s most impressive A.I. company, reportedly failed to follow through on a previous breakthrough because it was cut off from the H20. This weakness was no surprise: DeepSeek’s chief executive had repeatedly admitted that restrictions on the export of U.S. chips were the greatest impediment to his company’s future.

这些控制措施似乎正在见效。据报道，中国最具实力的人工智能企业DeepSeek据传因无法获取H20芯片导致重大技术突破受阻。这一颓势早有预兆：该公司首席执行官多次承认，美国芯片出口限制是该公司未来发展的最大障碍。

But Nvidia’s chief executive, Jensen Huang, lobbied hard for a reversal. To do so, he hyped China’s Huawei chips, claiming they were on par with Nvidia’s and that Huawei could produce them in competitive quantities. He urged the Trump administration to let Nvidia re-enter the Chinese market to stop Huawei from growing more powerful. China’s state media trumpeted Mr. Huang’s comments, as did Trump administration officials in justifying their reversal.

但英伟达首席执行官黄仁勋极力游说政策转向。为此，他夸大华为芯片的性能，声称它们与英伟达的芯片不相上下，并且具备量产能力。他敦促特朗普政府允许英伟达重返中国市场，以遏制华为的壮大。中国官方媒体大肆宣扬黄仁勋的言论，特朗普政府官员也以此为其政策逆转辩护。

But the claim that China’s chips rival the United States’ is false. Huawei simply has not shown that it can significantly increase A.I. chip manufacturing. Despite Chinese investments of hundreds of billions of dollars in chip manufacturing since 2014, U.S. export controls on chip-making equipment have held back Huawei’s production capacity. The estimated number of chips Huawei can manufacture this year would hardly be enough to fill a single cutting-edge data center.

但声称中国芯片能与美国芯片匹敌的说法是错误的。华为根本没有表现出显著提高了人工智能芯片制造的能力。尽管中国自2014年以来在芯片制造上投入了数千亿美元，但美国对芯片制造设备的出口管制阻碍了华为的生产能力。据估计，华为今年的芯片产量甚至不够一个尖端数据中心的需求。

Each of those Huawei chips also performs worse than advanced U.S. chips. Huawei chips account for only about 3 percent of global supercomputing power. If Huawei’s chips were as good as Nvidia’s, there wouldn’t be the overwhelming demand in China for the H20.

华为芯片的性能亦全面落后于美国的先进芯片。华为芯片仅占全球超级计算能力的约3%。如果华为的芯片和英伟达的一样好，中国就不会对H20有如此巨大的需求了。

The H20 is a very capable chip. It performs substantially better than Huawei’s best chip at inference. Opening the floodgates for H20s to flow into China will revitalize Chinese companies like DeepSeek as they try to supplant U.S. firms in the global market. Some other chipmakers may wonder why Nvidia gets to profit while their wares are subject to U.S. export controls.

H20是一款功能强大的芯片。它在推理方面的表现远优于华为最好的芯片。为H20开闸放行将重振DeepSeek等中国公司，它们正试图在全球市场取代美国公司。其他一些芯片制造商可能会质疑，为什么英伟达可豁免管制牟利，而它们的产品却要受制于美国的出口管制措施。

Worst of all, the H20 decision could fracture the hard-won bipartisan consensus on the need for American A.I. dominance over China. Mr. Trump’s decision has drawn bipartisan opposition, but the A.I. industry senses weakness in his China policy. Mr. Huang, who visited China to laud its A.I. companies just after he lobbied Mr. Trump, told a Chinese audience, “I hope to get more advanced chips into China than the H20.”

最糟糕的是，允许销售H20的决定可能会破坏美国在人工智能领域保持对中国的主导地位这一来之不易的两党共识。特朗普的决定已引发了两党反对，而人工智能行业却从他的对华政策中嗅到了软弱。黄仁勋在游说特朗普后不久就访问了中国，赞扬中国的人工智能公司，并告诉中国的听众：“我希望有更多比H20更先进的芯片进入中国。”

As the world’s first $4 trillion company, Nvidia doesn’t need the Chinese market to thrive. The demand for A.I. chips from U.S. companies is enormous: Firms are poised to spend hundreds of billions of dollars in 2025 alone on A.I. technology and data centers. Amid the A.I. boom, Nvidia’s stock price has increased by a factor of 10 over the last two and a half years, even as U.S. chip restrictions tightened.

作为全球首家市值4万亿美元的企业，英伟达无需依赖中国市场也能蓬勃发展。美国企业对人工智能芯片的需求是巨大的：仅2025年，在人工智能技术和数据中心上的投资预计就达数千亿美元。在人工智能繁荣浪潮中，即便美国加强了芯片管制，英伟达股价过去两年半仍暴涨十倍。

You’d be forgiven for assuming that the United States got something from China in exchange for giving away its advanced technology. Not from what I can tell. The day after the H20 reversal, China ratcheted up its own export controls on critical minerals and battery technologies — areas where it has an advantage. It was a fitting and lamentable coda to a decision that benefited one company at the expense of America’s A.I. leadership and its national security.

如果你以为美国是以先进技术换取中国在其他方面的让步，这样想情有可原。但据我所知，事实并非如此。就在H20政策逆转次日，中国就加紧了对关键矿产和电池技术的出口管制——中国在这些领域具有优势。这无疑为这项决策画上了一个可以预见的可悲句号：它让一家企业获益，却牺牲了美国在人工智能领域的领导地位和国家安全。