In the race for AI dominance with China, the United States’s one undeniable advantage is our access to the world’s most advanced AI hardware. We are home to the world leader in AI chip design, Nvidia, over half of hyperscale cloud capacity, and nearly 75 percent of the world’s AI supercomputing resources.
This lead, however, is tenuous at best. With a global build-out of new infrastructure underway, most of the aggregate computing capacity needed to power the AI revolution into the 2030s and beyond has yet to be built. China might soon race ahead of America.
Finally, a reason to check your email.
Sign up for our free newsletter today.
In this light, approving the export of Nvidia’s H20 chip to China, as the Trump administration did in August, could not be more misguided. The H20 is a cut-down chip that Nvidia created specifically for China following the export controls introduced by the Commerce Department in October 2023. With fast memory and Nvidia’s polished software systems, the H20 is formidable for deploying AI models to users.
Inference—the process of running, rather than training, AI models—has become increasingly relevant amid the emergence of reasoning models and agents, which solve longer tasks and generate reasoning data for training. Nvidia has at least 1.3 million H20s destined for Chinese buyers; if exported in full, they would represent the single largest block, a clear plurality, of inference silicon across all of China.
The Trump administration paused H20 exports to China in April but reversed course after Nvidia CEO Jensen Huang lobbied President Trump and the White House directly. The H20s have yet to ship, though, giving the administration a window of opportunity to reconsider its position. The damage could potentially be minimized by limiting Nvidia’s exports to the subset of H20s already fully assembled, or by giving U.S. buyers a right of first refusal, as proposed by Senator Jim Banks in the GAIN AI Act.
Instead, President Trump is reportedly considering approving additional sales to China of Nvidia’s new B30A chip, which boasts 12 times the performance power of the H20. Running two B30A side-by-side would let Chinese companies approximate Nvidia’s best chip, the B300—ending America’s hardware advantage altogether.
China is trying aggressively to close its compute gap with the U.S. Earlier this year, the state-owned Bank of China pledged 1 trillion yuan (roughly $140 billion USD) to finance new AI supercomputing centers for companies such as Baidu, ByteDance, and Alibaba. It’s also pursuing an aggressive buildout of energy infrastructure, with China adding 429 gigawatts of new generation capacity last year alone.
Leadership in AI requires dominating some key inputs, including data, talent, energy, and compute. Given that it is less energy constrained, China is poised to pull ahead of the U.S., but for export controls on advanced AI chips and semiconductor manufacturing equipment.
China’s state-backed chip foundry, SMIC, is severely constrained on the production side; the Commerce Department estimates that it can produce only 200,000 of Huawei’s top AI chip, the Ascend 910C—and with relatively low quality. Chinese companies may be surging production, but even if SMIC’s output doubled over the next year, China’s chip production capacity will still pale in comparison to the 10 million-plus Nvidia chips TSMC (Nvidia’s manufacturing partner) plans to produce in 2026.
As frontier AI systems grow more powerful, the balance of geopolitical power will be increasingly proxied by the global distribution of AI compute. Consider that the length of tasks AIs can perform autonomously is doubling roughly every four to seven months. If this trend holds, by the end of next year the best AI agents will be capable of reliably performing many tasks that would normally take human experts a full eight-hour workday. The same chips used today to run chatbots and simple coding agents will soon be running highly autonomous, generalist AIs with the potential to supercharge scientific R&D and economic growth. In other words, the value derived from advanced AI hardware is set to skyrocket.
Critics of chip-export controls argue that they surrender market share to China’s domestic AI chip designer, Huawei. Yet Huawei’s chip performance and SMIC’s production capacity remain far behind the curve and not in position to satisfy China’s domestic demand any time soon, much less to compete in international markets. Chinese AI company DeepSeek was even forced to delay its latest model after struggling to get Huawei’s training chips to work.
As it stands, Chinese companies are making every effort to skirt U.S. export controls through sophisticated smuggling operations that divert controlled hardware through a web of shell companies and third-country intermediaries. With millions of controlled chips exported every year, the Bureau of Industry and Security (BIS)—the agency within the Commerce Department tasked with enforcing export controls—has struggled to keep pace, not least because of its limited resources. BIS has just 12 export control officers around the world, including only two for China and one for all of Southeast Asia. These are America’s front-line agents for doing “end-use checks” to ensure foreign buyers aren’t diverting chips to unauthorized end-users—but their ability to inspect Chinese firms is limited, given a unique agreement that caps their window for post-shipment verifications to a mere 180 days rather than the standard five years.
A recent analysis from Erich Grunewald and Tim Fist estimates that as much as 40 percent of China’s AI training capacity in 2024 was gained via smuggling (the median estimate is 10 percent). Shockingly, the value of the hardware seized in a single busted smuggling operation typically surpasses BIS’s entire annual enforcement budget, leading the Trump administration to request that BIS’s budget be increased 77 percent to enable long-overdue modernizations. Grunewald and Fist still anticipate that smuggling will become China’s primary means of acquiring AI compute in the years ahead.
Rather than relaxing export controls and forfeiting the AI race altogether, the U.S. should be doubling down on enforcement to ensure that our lead remains secure. That means holding the line on existing chip controls, strengthening controls on the specialized components that SMIC needs to catch up on manufacturing, and adopting scalable anti-smuggling mechanisms, such as on-chip location verification.
America’s AI lead is tenuous. Along with accelerating development at home, the administration should be tightening restrictions on China’s access to advanced AI hardware. Now is not the time to back down.
Photo: AN MING / Feature China/Future Publishing via Getty Images