The Biden administration’s use of export controls to maintain an edge over China in the development of artificial intelligence has already been effective and could be increasingly crucial as high-performance chips play a greater role in training the next generation of frontier models, say analysts from the Center for Strategic and International Studies.
“The alternative to American AI chips is not Chinese AI chips. The alternative to American AI chips is no AI chips,” said Gregory Allen, the director of CSIS’ Wadwhani AI Center who formerly led strategy and policy at the Defense Department’s Joint Artificial Intelligence Center.
In conversation with Alan Estevez -- head of the Commerce Department’s Bureau of Industry and Security -- during a Jan. 14 event, Allen provided a counterpoint to critics of a new BIS rule published Jan. 15 which is ultimately aimed at limiting China’s access to components such as Graphics Processing Unit chips and model weights used in AI development.
It comes as observers note China’s circumvention of pre-existing restrictions on such components by accessing computing clusters in other countries.
Major tech industry companies and their associations say the rule could have the unintended consequence of strengthening China’s position in the marketplace for AI as it now requires a license to export or transfer the AI components to all but a short list of trusted countries.
“Even if Chinese-made GPUs aren’t as capable as U.S.-made GPUs in the short term, Chinese firms (or those from other competitors) will happily supply the computing power needed to meet customers’ challenges, vitiating the administration’s goal of limiting overall AI computing while simultaneously harming U.S. firms’ leadership in global AI compute technology and market share,” reads a Jan. 7 blog post from Information Technology and Innovation Foundation vice president of global innovation policy Stephen Ezell.
But CSIS’ Allen argued previous restrictions have not only hampered China’s development of large language models, but also the scale at which they’ve been able to produce relevant chips at all.
“The Chinese AI firm DeepSeek … they released an open source model that is by some performance parameters, really competitive, with what's coming out of Meta, what's coming out from everyone else,” he said. But then he noted, “The CEO of DeepSeek in a recent interview said the number one challenge facing the company is not financing, it's not human resources, it is those export controls that is the biggest challenge facing the future of his company.”
In addition, Allen said, “there's the question about not just buying chips, but making chips locally in China. I had the opportunity to speak to somebody who was talking to folks in Huawei's supply chain in the very recent past, and what they said is that SMIC, Huawei’s, preferred logic chip manufacturer for AI, is still stuck making fewer than 20 wafers per month, and that bottleneck in their process is exactly as you predicted ... [because it’s dependent on] equipment that is on export controls, which I think is a phenomenal outcome.”
“Their yields are dreadful, they're stuck as of November 2024, at 20 percent of the chips that come off that line [being] actually useful,” Allen said. “I think those are really strong data points as an endorsement of the actions that you've taken.”
Separately, another CSIS analyst told Inside AI Policy use of newer frontier models, such as OpenAI’s o1 -- which the company claims can reason -- will further increase demand for chips.
Generally “frontier models get better when you use more data, bigger compute, and smarter post-training to create their model weights,” said CSIS missile defense fellow Masao Dahlgren. “But the magic of o1 and other ‘reasoning’ models stem from test-time scaling -- improved techniques for performing inference, when the model is run.”
“This could mean … hardware controls could become more impactful, since reasoning models use more compute every time they are run,” he said. “This trend toward test-time scaling is one of the reasons analysts are bullish on AI chip company revenues, since users might need more compute to run reasoning models.”
At the same time though, Dahlgren said he’s concerned about how efficiently China’s DeepSeek has been able to use the chips they are able to access.
“Chip controls choke Chinese AI labs in the short run,” he said. “But we have to worry about the anaerobic exercise they're getting. High-performance Chinese models like DeepSeek V3 show how much performance they can squeeze from less compute; I worry that Chinese labs could gain a market foothold in compute-efficient, low-cost models that undercut U.S. competition at the low end.”
For that reason, he said, it’s important that the rule exempts open-weight models from the license requirements.
“The United States needs to maintain its lead in open model research to lock out Chinese attempts at gaining influence and market share with compute-efficient models,” he said, adding, “I’m … relieved the rule does not onerously restrict open-weights models. America wins when the world builds on American technology.”
Taking a long-term view, Dahlgren made a point that Estevez also acknowledged on the CSIS stage: export controls alone will not be enough to ensure U.S. national security.
“For U.S. national security, what matters more in the long run is how widely the United States adopts AI, rather than how completely it can delay its competitors,” Dahlgren said. “We should make the most of this breathing room.”