U.S. export control on advanced semiconductors is starting to become counterproductive. While the intended goal is to strengthen national security by preventing China from obtaining cutting edge American-designed AI chips, the real effects are hurting leading American companies, particularly Nvidia, while helping China’s national champion, Huawei. 

The first round of export control, announced in October 2022 by the Department of Commerce, was reasonable enough – banning the selling of high-end GPUs like the Nvidia’s A100 and H100 to China. Being a nimble and adaptive technology company, Nvidia came up with two alternative models, the A800 and H800, that could comply with the regulations and still be sold to China. One year later, a second round of more restrictive and expansive export control is squeezing Nvidia’s entire China business, roughly one-fifth of its revenue, by not only banning the two alternatives, but also signaling further scrutiny of future similar alternatives, like the H20 model that Nvidia is currently developing.

Last month, when Commerce Secretary Gina Raimondo was asked about Nvidia’s alternative products at the Reagan National Defense Forum, she called the behavior “not productive” and would immediately ban any future alternative chip that enables China to “do AI” or has “AI special sauce.” To anyone with a basic understanding of how these GPUs work, “AI special sauce” sounds like meaningless gobbledygook. Is it referring to each GPU’s raw processing power? Or the networking that links multiple GPUs together to enable parallel processing among the chips? Or the High Bandwidth Memory (HBM) that is packaged right next to the GPUs to boost training performance on the type of large foundation AI models that power applications like ChatGPT?  

Nvidia’s China business holds important geopolitical value for the U.S. It is not just about one company making more money. It is about maintaining some access to a competitor’s vast market that would reveal clues and information to its current progress and future direction. In other words, knowing how many H800s or H20s are sold in China is perhaps the best way to estimate and predict how far behind (or ahead) China's AI capabilities really are. If one day Nvidia’s China business becomes a sacrificial lamb, that would only undermine America's own geopolitical positioning. 

Things are already trending in that direction.

Large Chinese technology companies, like Alibaba, Tencent, Baidu, and ByteDance, used to be aggressively stockpiling the H800 chips before they were banned – a proofpoint of Nvidia’s, and by extension America’s, dominance. Now, they are on the fence about doing the same for the H20 chips, because the performance may no longer be compelling enough, while the specter of future export ban makes any effort to stockpile not worthwhile. Instead, budget allocation and purchase orders are being spent on Huawei’s Ascend 910B chips, a product that is still two to three generations behind Nvidia’s best offering but can be reliably obtained. These are millions of dollars from China’s own technology companies that could have gone to Nvidia’s R&D, but are instead diverted to fund Huawei’s advancement. And instead of helping Nvidia compete better, we are holding up the punitive stick of more export control, while dragging its CEO, along with the CEOs of Intel and Micron, to testify in front of Congress for more political theater.

Huawei’s progress has already blindsided US officials once. Its newest smartphone model, the Mate 60 Pro, surprised everyone in Washington last August by sporting a homemade 7nm processor – a level of chip-making capability that US export control was supposed to prevent China from achieving. Hurting Nvidia and letting Huawei fill the void only adds more fuel to its engine, more opacity to our knowledge of China’s AI capabilities, and more unpleasant surprises ahead.  

“The Art of War” by Sun Tzu is widely studied by generals, policymakers, and business leaders everywhere. One of its most fundamental and widely-cited principles is: “Know yourself and know the enemy, and you can fight a hundred battles without defeat.” There is also a second part to this principle that is less known: “If you know yourself but not the enemy, for every victory gained you will also suffer a defeat.

America is still in the lead by many measures when it comes to AI specifically and technological innovation broadly. To maintain that lead versus China, ironically, it would serve us well to brush up on the wisdom of a Chinese military strategist from 2500 years ago, before we commit more strategic mistakes that shoot ourselves in the foot.



美商务部于2022年10月宣布的第一轮出口制裁还算合理 —— 禁止向中国销售高端GPU,如英伟达的A100和H100。作为一家灵活及快速迭代的科技公司,英伟达很快推出了两款替代产品,A800和H800,既可以合规又能销售给中国客户。一年后,第二轮更严格和广泛的出口管制政策正挤压着英伟达在中国的整个业务,大约占其收入的五分之一。制裁不仅包括这两种替代品,还有意表明对未来类似替代品的进一步审查和管制,如英伟达目前正在开发的H20芯片产品

上个月,当美商务部长吉娜·雷蒙多在里根国防论坛上被问及英伟达的替代品时,她称这种行为“无益”,并会立即制裁任何未来类似的替代芯片,如果这些芯片使中国能够 “发展人工智能” 或拥有 “人工智能秘诀”。对于任何对人工智能的技术基本常识或GPU的科技本质有些了解的人来说,“人工智能秘诀” 这种字眼听起来基本上是毫无意义的废话。是在指每款GPU的算力吗?还是连接多个GPU以实现芯片间并行处理的网络宽带容量?还是紧靠GPU旁边的高带宽内存(HBM),用来提升在基础AI模型(如ChatGPT等应用程序所使用的GPT模型)的训练性能?



像阿里、腾讯、百度和字节这种中国科技巨头,曾经在H800芯片被制裁之前果断而积极的囤积 —— 这也证明了英伟达乃至美国的AI科技中的领先地位。但现在,这些大厂对于H20芯片仅持观望态度,因为其性能可能已不足以值得囤积,何况总有未来再被出口禁令制裁的风险。相反,预算和订单正在拨给华为昇腾910B芯片。这款产品虽然落后于英伟达最好产品起码两三代,但可以可靠地买到。本可以流向英伟达研发的数百万美元,现在却被转向资助华为。美国不但没有在帮助英伟达,反而举起了更大的制裁“惩罚棍”,还拖拽其CEO以及英特尔和美光的CEO去在国会前作听证会,搞更多的政治表演。

华为在芯片方面的进步步伐已经是美国官员们的一个盲点。去年八月,华为最新的智能手机,Mate 60 Pro,出炉时在华盛顿引起了巨大的波动和惊讶,因为内核是国产自制的7纳米芯片 —— 本应该是美国出口管制可以阻止中国国产芯片制造能达到的能力。伤害英伟达,让华为填补空白,只会为其研发引擎添加更多燃料,增加美国对中国人工智能发展进度的盲点,带来更多不愉快的惊讶。

《孙子兵法》被各国人士广泛阅读和研究,从军队领导,到政策制定者和商界领袖。其中一条最基本且被广泛引用的原则是:“知彼知己,百战不殆。” 这段名言还有第二部分,却没有太多人引用:“不知彼而知己,一胜一负。”