If superintelligence isn’t imminent, the Trump administration may be right to loosen advanced chip export controls

When President Trump announced in December 2025 that his administration would approve export licenses for Nvidia’s H200 chip and other advanced chips, China hawks from the Biden administration exploded in outrage. “This decision is nuts,” former National Security Adviser Jake Sullivan told The New York Times. “We are literally handing away our advantage. China’s leaders can’t believe their luck.” 

But the reality is more nuanced. Effective export controls must balance the specific advantage the U.S. aims to preserve against the potential damage to other interests, such as global market dominance. As James Lewis of the Center for Strategic and International Studies documented, well-intentioned but poorly conceived export control efforts have often backfired and impeded efforts to maintain a technological advantage.  

Sullivan and his co-author, Tal Feldman, explained in a January 2026 Foreign Affairs article that a lot of decisions about chip export control depend in large part on what you believe about the imminence of superintelligence. “Will AI progress accelerate toward superintelligence,” they asked, “or plateau for an extended period?” In a world where superintelligence is achievable, the technology would be hard to imitate quickly. And given that China is racing at full speed, the authors argue the U.S. should tighten export controls—including every layer in the chip supply chain “to the limits of enforceability.”  

That might be the world conceived of by the previous administration when it tightened chip export controls in 2022. According to Chris Miller, author of “Chip War,” “the real cause” of the 2022 controls was the National Security Council’s conclusion that advanced chips “would drive progress in artificial intelligence.” At that time, a significant body of expert and industry opinion held that progress in AI was only a matter of increasing the data, parameters, and computer time used to train large language models (LLMs). According to Miller, it was this concern over “scaling laws” that drove the administration “to ratchet upward the chip controls on China, targeting the country’s access to AI-capable GPU chips.” 

When OpenAI released ChatGPT in November 2022, the AI conversation quickly turned to superintelligence. AI progress seemed so swift that artificial superintelligence, where computer models would substantially surpass humans in all cognitive tasks, appeared within reach in just a few years.  

If superintelligence was right around the corner—but only with the highest-end chips that only the U.S. possessed—then a high-end chip blockade might have made sense. If the U.S. could delay China’s chip development for just a few years, U.S. firms would get to superintelligence first, thereby giving the U.S. a permanent and decisive economic, technological, and military advantage. This thinking led to the 2023 chip control updates and the January 2025 diffusion rule, which the Trump administration repealed in May 2025.

But the path to superintelligence seems a lot less clear today than it did then. AI scaling seems to have hit a wall, producing modest improvements in select model benchmarks with no dramatic breakthroughs. All signs point to future AI progress coming from useful applications, demonstrated by Anthropic’s recent improvements to its coding tool, Claude Code, and its new legal services and  financial services capabilities built on Claude Cowork. Chinese companies have produced AI models such as Zhipu’s open-source GLM-5 with similar coding capabilities relying entirely on domestic Chinese chips.

Continued attempts to achieve superintelligent models through scaling parameters, data, and compute seem less promising. Indeed, the big language models themselves seem increasingly indistinguishable from each other in practice, leading industry analysts like Paul Kedrosky and Mary Meeker to say they have become commodities, differentiated only by price and ease of use for constructing useful services.  

But if that is so, then continued export controls on high-end U.S. chips make less sense now than they might have years ago. They would cut U.S. suppliers out of the top-end of the world’s biggest chip market and allow increasingly capable Chinese chips to take their place. Their effectiveness seems limited, in any case, since even with controls in place, Chinese firms such as DeepSeek and Alibaba have developed highly capable open-source AI models that gained international popularity in 2025. Zhipu’s latest model built with Huawei’s Ascend AI chips also enjoy a dramatic price advantage over comparable U.S. models.

The more relaxed U.S. approach that President Trump announced in December 2025 would grant export licenses to China for Nvidia’s H200 chip, which is highly capable but has significantly lower processing performance compared to Nvidia’s Blackwell B300. China approved some imports of the newly available U.S. chips that include conditions and limits designed to continue its push toward technological self-reliance. The U.S. has crafted conditions for case-by-case chip approval including a volume restriction to limit exports to China to 50% of the total amount sold to U.S. customers and know-your-customer rules to prevent military use. The U.S. also collects 25% of the sales prices as an import fee, which some commentators regard as an illegal fee-for-license arrangement. No transfer has taken place yet as security reviews on the U.S. side continue, and Chinese import regulations remain pending.  

Wrangling over these implementation details might delay the flow of high-end chips to China, but it seems clear that U.S. chip export controls will be loosened going forward. Without the superintelligence justification for the toughest possible export controls, this more relaxed posture seems to be the right direction.