美国应该向中国出售 Hopper 芯⽚吗？

Should the US Sell Hopper Chips to China? 美国应该向中国出售Hopper芯⽚吗？ Assessing the impacts of exporting the H200 or H100 AI chips EMERGING TECHNOLOGY By,,, andSaif KhanTao BurgaTim FistGeorgia Adamson 执⾏摘要 Executive summary Last month, the Trump administration decided not to export a version of NVIDIA’sflagship Blackwell AI chips (such as the rumored “B30A” chip) to the People’sRepublic of China (henceforth “China”), with President Trumpstating“We will notlet anybody have [the most advanced AI chips] other than the United States.”Now, the administration is expected toconvenea high-level meeting todecidewhether to authorize the export of NVIDIA’s H200 chips to China. 上个⽉，特朗普政府决定不向中华⼈⺠共和国（以下简称“中国”）出⼝英伟达的旗舰Blackwell AI芯⽚的⼀个版本（例如传闻中的“B30A”芯⽚），特朗普总统表示“我们不会让任何⼈拥有（除美国以外的）最先进的AI芯⽚。”现在，政府预计将召集⼀次⾼级别会议，以决定是否授权将英伟达的H200芯⽚出⼝到中国。 Released in March 2024, the H200 is NVIDIA’s best AI chip from the previous “Hopper” generation and is an upgraded version of the H100 chip. Much likecutting-edge Blackwell chips, Hoppers use TSMC’s 4nm manufacturing processtechnology. Despite being a previous-generation AI chip, the H200 and the H100will likely be highly useful for frontier AI workloads throughout 2026: H200于2024年3⽉发布，是NVIDIA之前“霍普”世代中最好的AI芯⽚，是H100芯⽚的升级版。与最先进的Blackwell芯⽚类似，霍普采⽤台积电的4nm制程技术。尽管是上⼀代AI芯⽚，H200和H100在2026年期间很可能仍对应前沿AI⼯作负载⾮常有⽤： As of December 2025, 18 of the 20 most powerfulpublicly documentedGPU clusters in the world primarily used Hopper chips, including all of thetop 7. Hoppers currently represent just over half as much installed AIcompute as Blackwells. 截⾄2025年12⽉，全球公开记录的20个最强⼤GPU集群中有18个主要使⽤霍普芯⽚，其中包括排名前7的全部集群。霍普⽬前代表的已安装AI计算量略多于Blackwell的⼀半。 Previous generations of AI chips have remained in use for frontier modeltraining for roughly four years. If this trend holds for Hoppers, they will likelybe used for frontier training throughout 2026 — especially the H200, whichwas released in late 2024. 以往⼏代AI芯⽚在前沿模型训练中的使⽤寿命⼤约为四年。如果这⼀趋势对霍普仍然适⽤，它们很可能在整个2026年继续⽤于前沿训练——尤其是于2024年底发布的H200。 Permitting export of meaningful volumes of advanced Hopper chips to Chinawould have six key implications: 允许向中国出⼝⼤量先进霍普芯⽚将带来六个关键影响： 1.The decision would be a substantial departure from the Trumpadministration’scurrent export control strategy,which seeks to denypowerful AI compute to strategic rivals. The H200 would be almost 6x aspowerful as the H20 — a chip that requires an export license to China andhas been approved for export in only limited quantities. 这⼀决定将⼤幅偏离特朗普政府当前旨在剥夺战略对⼿强⼤AI算⼒的出⼝管制策略。H200的性能⼏乎是H20的6倍——⽽H20是⼀款对华出⼝需申请许可且仅在有限数量下获批出⼝的芯⽚。 2.China would have access to chips that outperform any chip its companiescan domestically produce, and at much higher quantities.Huawei is notplanning to produce an AI chip matching the H200 until Q4 2027 at theearliest.Even if this timeline holds, China’ssevere chip manufacturingbottlenecksmean that it will not be able to produce these chips at scale,reaching only 1–4% of US production in 2025 and 1–2% in 2026. 中国将能获得性能超过其企业本⼟可⽣产的任何芯⽚的产品，⽽且数量远⾼于本⼟产量。华为最早也要到2027年第四季度才计划⽣产可匹配H200的AI芯⽚。即便这⼀时间表成⽴，中国在芯⽚制造⽅⾯的严重瓶颈意味着其⽆法实现⼤规模⽣产，预计在2025年仅能达到美国产量的1–4%，在2026年为1–2%。 3.Chinese AI labs would be able to build AI supercomputers that achieveperformance similar to top US AI supercomputers, albeit at a cost premiumof roughly 50% for training and 1-5x for inference, depending on workload.Blackwells have been marketed as delivering 30x the inferenceperformance of the H200. However, this multiplier does not account forprice differences and reflects best-case assumptions for Blackwell-basedclusters and worst-case assumptions for Hopper-based clusters. In anapples-to-apples analysis, we estimate near parity between Blackwells andH200s for many inference workloads; up to a 5x advantage for Blackwellson the type of inference workloads for which they are best suited; and a 1.5x advantage for Blackwell-based clusters in training. This means that withaccess to Hoppers, Chinese labs could build AI training supercomputers ascapable as American ones at 50% extra cost— a premium that the ChineseCommunist Party (CCP) would likely at least partly subsidize. The Blackwelladvantages could increase if AI developers can exploit its advanced “low-precision” features,which some might do in 2026, but there is not yetevidence that any have done so at scale. 中国的⼈⼯智能实验室将能够构建在性能上接近美国顶级⼈⼯智能超算的系统，尽管在训练上成本⼤约⾼出50%，在推理上则视具体情况⾼出1-5倍。workload.Blackwells被宣传为其推理性能是H200的30倍。然⽽，这⼀倍数并未考虑价格差异，⽽且基于对Blackwell架构集群的最好情形假设和对Hopper架构集群的最坏情形假设。进⾏同类⽐较分析时，我们估计在许多推理⼯作负载上Blackwell与H200接近于同等；在最适合Blackwell的那类推理⼯作负载上，Blackwell最多可占有5倍优势；⽽在训练上，基于Blackwell的集群可有1.5倍的优势。这意味着如果能获得Hopper，中国实验室可以以额外50%成本建造出与美国同等能⼒的⼈⼯智能训练超算——这⼀溢价很可能会被中国共产党（CCP）⾄少部分补贴。如果⼈⼯智能开发者能够利⽤Blackwell的先进“低精度”特性，Blackwell的优势可能会增加，有些开发者可能会在2026年这样做，但⽬前尚⽆证据表明有⼈已经在⼤规模上这样做。 4.These exports would counterfactually add to China’s total supply ofadvanced AI compute, and would likely do little to slow China’sindigenization efforts.China’s chip manufacturing bottlenecks meanHuawei’s peak production will still