Trump Administration Gets H200 Chip Sales to China Right and Wrong

The Trump administration recently announced that it would allow the sale of certain advanced chips to China, including Nvidia’s H200, which the U.S. government had previously restricted. The Trump administration is correct to allow sales to China as these exports will divert funds from Chinese competitors to U.S. companies, bankroll R&D into next-generation chips, and ensure leading Chinese AI firms continue to rely on U.S. technology. However, its plans to impose a 25 percent fee on such sales undermine this goal.

Some policymakers have denounced the policy change. For example, Sen. Elizabeth Warren (D-MA) argued that selling these chips “risks turbocharging China’s bid for technological and military dominance and undermining U.S. economic and national security.”

However, these critics are wrong for a number of reasons.

First, efforts to cut off U.S. chipmakers from the Chinese market only accelerate China’s progress in AI chips. When U.S. firms like Nvidia and AMD cannot sell into China, Chinese chipmakers have greater market opportunities as they no longer face competition from U.S. suppliers. Moreover, the absence of U.S. chips spurs Chinese innovation to fill in the gaps for Chinese buyers who would prefer to buy American technology but can no longer access it.

Rather than slowing China's ascent, U.S. government restrictions on American chips hasten it. China could not have designed a better policy if it tried. Indeed, after the Trump administration finally reversed its policy and greenlit sales of the less capable H20 chip in August, the Chinese government decided in September to ban its firms from buying Nvidia products. China would rather have its companies buy domestic chips, even if they lag the current state of the art, because they are primarily interested in supporting their domestic champions. In short, the American ban had been achieving exactly what China wanted.

Second, the H200 is not Nvidia’s most advanced chip on the market, meaning the United States will not, as some Democrat senators argue, “squander America’s primary advantage in the AI race.” The H200 is part of the Hopper architecture, first released in 2022. It presented a major advancement at the time, but the company has since moved on. Its current set of chips, such as the B100 and B200, is based on the Blackwell architecture and delivers another significant performance increase. Nvidia plans to release its next set of new chips, using its new Rubin architecture, in 2026. These chips will include higher performance memory and a significant increase in performance for some types of calculations. In short, exporting the H200 would mean that China will have access to good U.S. chips, but not America’s best.

Third, China does not need H200 chips to compete with the United States in AI. China already has a significant amount of computing power across its data centers, with about one-quarter of global capacity compared to 44 percent for the United States. Indeed, it presently has a surplus of compute capacity in the country, after a surge in data center construction in recent years. Moreover, both Microsoft CEO Satya Nadella and Nvidia CEO Jensen Huang have stated that the biggest bottleneck for AI is not access to chips but access to power. Indeed, China may have an advantage in this regard as, unlike the United States, it has invested in a massive buildout of power generation and transmission capacity that is expected to exceed its data center demands.

Fourth, many critics make too much of the direct head-to-head comparisons of Huawei’s chips versus Nvidia’s chips. There is no doubt that the Ascend 910C, Huawei’s main offering, is an inferior chip when looking at the specs. But Huawei’s goal is not to compete by building chips that can outperform Nvidia’s on paper. Instead, it is focusing on building supernodes—highly optimized AI systems with innovations in networking, optics, and software. Huawei is competing with Nvidia by offering its CloudMatrix supernode that integrates 384 Ascend 910C chips to compete with Nvidia’s GB200 NVL72 system, which contains 72 Blackwell GPUs. And Huawei’s recently announced Atlas 950 supernode, to be released in 2026, will integrate 8,192 chips, providing 6.7 times more processing power than Nvidia’s NVL144 system, which will also be released next year. Finally, Huawei plans to continue stacking these supernodes to create a supercluster with more than 500,000 chips. While Huawei’s strategy is not guaranteed to win in the marketplace, it is highly focused on optimizing its system design for parallel processing and efficiency, exactly the metrics demanded by customers running AI applications. Nvidia faces real competition.

Finally, keeping Chinese firms on U.S. chips has major implications for the broader AI ecosystem. Nvidia chips use CUDA, a proprietary parallel computing platform for developers. It is the industry standard, with a broad ecosystem of libraries, optimization tools, and software development kits that developers can use to create applications. It also provides native support for PyTorch (developed by Meta), TensorFlow (developed by Google), and JAX (developed by Google and Nvidia), popular machine learning frameworks. In contrast, Huawei’s chips use an alternative platform, CANN, which natively supports MindSpore (developed by Huawei) as its software framework for machine learning. Switching from one platform to another can introduce costly delays, so once developers change, they may not want to switch back.

There is strong demand for H200 chips among Chinese companies, so there is a window of opportunity to keep China on the hook and counter Chinese government efforts to restrict companies from using American chips. The administration should focus on ensuring market access for U.S. chipmakers as part of its trade negotiations with China, rather than demanding a cut of revenue from chip exports to China. Given that China is pouring tens of billions more to support its chipmaking industry, the United States should not impose unnecessary fees on U.S. chipmakers. Indeed, the goal should be to make U.S. chipmakers the most successful in the world, so that they can reinvest the revenue from sales to China in R&D to stay ahead and compete globally.

The Trump administration’s decision to allow sales of the H200 is the right move to make. Policymakers who are critical of the decision should look more closely at the details, as limiting these sales would be much worse for U.S. economic and national security interests.