By Sean Kucer
AI is "rapidly transforming our world
." It is being used to diagnose patients more accurately than doctors
, manipulate voters
and advise traders
on Wall Street. It may also reshape the international military balance. In the words of Secretary of Defense Mark Esper
, we must “lead in [AI’s] national security applications to maintain our strategic position…” One way to ensure the U.S. continues to lead in this space is by controlling the resources necessary for our competitors to accelerate their own AI development. Our best opportunity for this may be in the case of semiconductor manufacturing equipment (SME).
Advances in AI are partly driven by increases in computational power (“compute”), as has been noted by the Department of Defense
. One way of seeing this is by looking at the amount of compute used for advanced AI algorithms. AlphaZero, a famous game-playing AI developed by DeepMind, used orders of magnitude more compute
than AI algorithms of the last decade, like DeepSpeech2. The amount of compute used in the largest AI training runs has doubled every 3.4 months
. More compute generally means better performance
Compute comes from improved hardware. State-of-the-art AI chips, with 5nm transistors, are some 25x more economical
than the 65nm chips of 2006. Whatever China’s economic prowess may be, without cutting-edge manufacturing technology China would need to outspend the US and its allies 25 to 1 just to keep pace.
AI chips are similar to normal computer chips but are specialized for AI. AI chips share the trend towards more and smaller transistors, which provide speed and energy efficiency. However, AI chips are unique in running many calculations in parallel rather than sequentially, imprecisely calculating numbers to boost efficiency for AI algorithms, and sometimes storing entire AI algorithms on a single chip. There are also different types of AI chips—notably graphics processing units, field-programmable gate arrays, and application-specific integrated circuits—which have their own unique functions within the broader AI system. But together, AI chips allow for training more powerful programs at a small fraction of the cost that CPUs can, making AI chips an essential resource for AI development.
This is where export controls come in. China is dependent on the United States and its allies for cutting-edge chips. They can make some
AI chips, but they are not state-of-the-art. This limitation is due not only to a lack of technical know-how, but also a lack of access to manufacturing equipment with the precision necessary to create the most advanced chips
. The production of the SME capable of producing state-of-the-art AI chips is dominated by the US, the Netherlands and Japan, which together account for 94% of SME production
(47%, 17%, and 30%, respectively). This is why China is rushing to buy the SME necessary for them to create their own state-of-the-art fabrication plants. But if we enact strict export controls on SME, we will remain in a much better position to secure our national interests. Many expert organizations have noticed this and released reports suggesting such export controls on SME, including The National Security Commission on Artificial Intelligence
, the Center for Security and Emerging Technology
, and the Center for a New American Security
Export controls remain essential to national security and are relatively straightforward to update. In fact, the US government has already had success in reducing SME exports to China, evident in the recent reversal of the Netherlands government’s decision to grant a license for the company ASML to sell its most advanced manufacturing machine to China. The powerful $150 million machines did not get shipped to China. Export controls have also historically received support from both parties. Recent export controls
hold some promise but are still somewhat lax on SME while imposing unnecessary restrictions on other exports (such as consumer or commodity chips). The essential thing to control is SME to produce cutting-edge chips.
It is time to enact stricter export controls on semiconductor manufacturing equipment. This is an excellent opportunity for us to collaborate with our allies, bolster national security, and keep the United States at the forefront of technological development.
Sean Kucer is a former research intern with the Technology Policy Program at the Center for Strategic and International Studies in Washington, DC.
The Technology Policy Blog is produced by the Technology Policy Program at the Center for Strategic and International Studies (CSIS), a private, tax-exempt institution focusing on international public policy issues. Its research is nonpartisan and nonproprietary. CSIS does not take specific policy positions. Accordingly, all views, positions, and conclusions expressed in this publication should be understood to be solely those of the author(s).