China's latest AI model development sparks US national security debate over Arizona-fabricated chips (TSMC/Intel), intensifying export control scrutiny on semiconductor supply chains.
More than 1,300 employees and executives at OpenAI, Anthropic, Google, Meta, and other AI companies signed a letter last week urging the U.S. government to help build tools to deliberately slow AI development. That same week, Chinese AI companies accelerated.
On July 27, Beijing startup Moonshot AI released the full weights of Kimi K3, a 2.8-trillion-parameter model available for free download. Independent benchmarks place it within striking distance of the best American systems, though running it requires expensive, commercial-grade computing power. Arizona has a direct interest in this competition: some of the advanced Nvidia chips at the center of the dispute are manufactured at TSMC's facility in north Phoenix.
Lily McElwee, president and CEO of the Phoenix Committee on Foreign Relations and an expert on U.S.-China relations, said the release laid bare nearly every unresolved question in the rivalry. China's embrace of open-weight models—systems whose parameters are published, allowing anyone to download, modify, and run them instead of paying for access—fits a long-standing strategic pattern. "China has always framed itself as a provider of more freely accessible, low-cost technology to the rest of the world while the United States provides these proprietary systems," McElwee said. This pitch, aimed at developing countries and the Global South, frames American firms as primarily profit-motivated.
When asked whether the open-weight push resembles China's dumping of subsidized steel to drive out competitors, McElwee acknowledged the parallel. "I'm not saying that Beijing is acting in good faith. I'm saying it's a marketing ploy." She pointed to Huawei and 5G as the closest precedent. Huawei became the world's largest supplier of telecom network equipment, in part by undercutting Western rivals on price with state subsidies and cheap state financing for developing-world customers. The U.S. blacklisted the company in 2019 and pressed allies to exclude it from their 5G networks. Yet much of Europe, Africa, and Latin America built 5G infrastructure with Huawei equipment anyway.
Security vulnerabilities loom. A Booz Allen Hamilton study published in June found that three of four Chinese models tested produced significantly more vulnerable code when users were identified as working for the U.S. government. McElwee framed the concern clearly: "Washington is rightly concerned about data vulnerabilities when U.S. customers use open-weight Chinese models, because the output of those models is generating data on U.S. citizens." The U.S. bans Chinese electric vehicles and other Chinese technology products, she noted, yet "if we're not banning AI systems—which are increasingly used to train those products and access consumer data in the United States—then we have a security vulnerability."
Kimi K3 itself faces specific accusations. Michael Kratsios, director of the White House Office of Science and Technology Policy, claimed Moonshot trained K3 on outputs from Anthropic's Claude and used restricted Nvidia chips routed through Thailand. Moonshot has not responded publicly to these allegations.
The Biden administration imposed the first significant restrictions on advanced AI chip exports to China, which Beijing views as a core grievance since chips represent its choke point in the AI race. The Trump administration eased some controls in January, moving Nvidia's H200 from near-automatic denial to case-by-case review, though the most advanced Blackwell processors remain prohibited. "There's not a blanket restriction anymore. It's case-by-case," McElwee said. "But to my knowledge, there have been no cases that are approved."
Yet China is obtaining chips anyway—through partners in Southeast Asia, stockpiling, and by recruiting the scientists who design them. McElwee argues that enforcement should focus there. Washington has also tightened imports, with the FCC banning new imports of foreign-made humanoid robots, robot dogs, and power inverters last week on national security grounds. The administration is weighing restrictions on Chinese open-weight models as well, though several outlets reported the White House may exempt such models from its review regime.
McElwee is skeptical that restrictions on open-weight models would be effective, given that they are freely downloadable by design. She argues that controls should target the tangible advantage the U.S. actually holds: advanced chips. "Let's take advantage of our strong points in this AI race, which are advanced Nvidia chips, which are produced right here in Arizona," she said.