Paul Scharre
Recent U.S. export controls are a major escalation in the U.S.-China tech competition. In September, National Security Advisor Jake Sullivan announced a change from the U.S. goal of being “only a couple of generations ahead” of China in key technologies to maintaining “as large of a lead as possible.” The Biden administration’s aggressive new semiconductor export controls put this principle into practice, aiming to stop Chinese chip development in its tracks. Yet key elements of the policy are likely to backfire.
Semiconductors, or computer chips, are vital to the global economy today, and trends in machine learning suggest they are poised to become an even more vital strategic resource in the future. Recent supply-chain shortages have highlighted semiconductors’ importance, and in August Congress passed $52 billion in subsidies to re-shore semiconductor manufacturing to the United States. Yet chips are more critical than most policymakers or CEOs realize. Computing hardware is one of four key battlegrounds, alongside data, talent, and institutions, that will determine which countries lead an AI-driven future.
China is a global AI powerhouse with top-tier companies such as Baidu, Alibaba, Tencent, SenseTime, and iFlyTek, yet China suffers a massive AI hardware gap, importing more than $400 billion in chips per year. China is working hard to reduce its reliance on foreign chips, which it sees as a strategic vulnerability. Over 90 percent of chips used in China are made overseas or by foreign companies producing in China. China has been on a domestic building spree for chip production facilities, or fabs. Before 2020, China accounted for the fastest-growing share of global semiconductor production.
The Biden administration’s new export controls restrict sales of U.S. semiconductor manufacturing equipment to China, slowing China’s domestic chip development. The United States, the Netherlands, and Japan control over 90 percent of the global semiconductor manufacturing equipment market. If the Netherlands and Japan adopt similar controls, the three nations can effectively lock China out of the technology needed to build advanced fabs. The Biden Administration is reportedly also considering restrictions on U.S. investments in Chinese AI and semiconductor firms. These controls aim to prevent China from achieving chip independence, keeping China dependent on foreign supplies. Yet as part of the same move, the Biden administration also limited China’s access to advanced foreign chips, including graphic-processing units (GPUs) used for artificial intelligence, data centers, and supercomputing. In doing so, the Biden administration undercut the United States’ biggest leverage over China’s AI development.
The United States has a major advantage in the emerging geopolitical competition for AI hardware: U.S. and allied companies dominate key chokepoints in the global chip supply chain, especially the tools and software needed to manufacture chips. Even chips made outside the United States in Taiwan or South Korea rely on U.S. equipment for manufacturing.
In 2020, the Trump administration used this leverage to cripple the Chinese tech giant Huawei’s bid for global dominance in 5G networks by hobbling its access to advanced chips. Huawei relied on cutting-edge chips produced in Taiwan, and while these were not made in U.S. factories, they were made using American equipment. The latest U.S. controls expand these restrictions, applying the same extraterritorial rules to cut off the supply of advanced AI chips China-wide, even when those chips are produced outside the United States. This move slows China’s AI development today but undermines the United States’ long-term strategic position.
The ability to deny China access to advanced chips is a powerful advantage whose value is growing exponentially. The past decade has seen an explosion in artificial intelligence. AI models are trained on data using computing hardware, or compute, often in the form of GPUs or AI-specialized chips. Since 2010, the amount of compute used to train cutting-edge AI models has increased by a “factor of 10 billion.” Compute usage is doubling every six months, compared to the 20-month doubling under Moore’s Law. (Compute for the largest AI models is doubling every 10 months.)
Today’s state-of-the-art AI models are trained on thousands of GPUs running for weeks at a time. Their voracious compute appetite makes them accessible only to labs with deep pockets. Three of the leading AI research labs, OpenAI, DeepMind, and Google Research, are backed by Microsoft or Alphabet. Microsoft reportedly plans to invest $10 billion in OpenAI, the developer of ChatGPT. Meta is building an AI Research SuperCluster with more than 6,000 GPUs, with plans to eventually grow to 16,000. Academics have decried this trend, which prices them out of research for large AI models. The U.S. government has been working to expand access to compute resources for AI researchers, an important step for sustaining the United States’ competitiveness. Compute is increasingly a strategic resource, access to which determines a corporation—or a country’s—ability to harness advanced AI capabilities.
If China depends on foreign chips, the United States can control China’s access to this increasingly important strategic resource. Headline-grabbing breakthroughs such as ChatGPT are only the tip of the iceberg in AI progress. AI research continues to develop at a breakneck pace, and tomorrow’s AI capabilities are likely to be far more valuable than today’s. Using U.S. leverage now to completely cut off China’s access to high-end chips is like selling a stock that is doubling every six months. Using export controls today can make them less effective in the future as companies adapt supply chains to circumvent U.S. restrictions.
The U.S. government has effectively just created a massive market for a U.S.-independent semiconductor supply chain. The Chinese government has long pursued wasteful and largely ineffective industrial policies to boost domestic semiconductor manufacturing. In the wake of U.S. export controls, the Chinese government pledged another $140 billion in spending. U.S. controls supercharge the Chinese government’s efforts by creating incentives for the private sector as well, mobilizing the global marketplace to accelerate Chinese chip independence. Foreign companies have strong incentives to design U.S. components out of their supply chains over time to circumvent U.S. restrictions and sell to the Chinese market. Replacing U.S. suppliers will take time, but all of the incentives are now in place to do so.
Even if Japan and the Netherlands join the United States on tooling restrictions in China, foreign companies could build new fabs entirely free of U.S. equipment outside of China to service the Chinese market. Restricting access to foreign chips also turns China’s $400 billion of buying power inward, boosting domestic chip manufacturing. Chinese commercial data and cloud companies could choose to turn to second-tier domestic chip suppliers, rather than continue to rely on foreign chips whose supplies are uncertain. These market incentives could mobilize the private sector, inside and outside China, to finally achieve the chip independence the Chinese government has long sought yet been unable to achieve to date.
More targeted controls on military applications and human rights abuses while permitting commercial use would be a better approach and could keep China dependent on foreign chips produced with U.S. equipment. Targeted controls are more cumbersome, as the United States is forced into a game of whack-a-mole while the Chinese military tries to circumvent U.S. restrictions through other buyers.
However, the United States just created a powerful tool to crack down on China’s military-civil fusion. As part of the October 2022 controls, the Commerce Department established that firms that do not cooperate with end-use checks would be placed on the Unverified List and, after 60 days, would transfer to the Entity List if they still have not complied. The U.S. government already successfully used this tool in December to verify and clear 27 Chinese companies that demonstrated compliance with end-use restrictions while escalating others, including Chinese chipmaker YMTC, to the Entity List. The United States has the tools it needs to enforce compliance with targeted restrictions. A more targeted approach that allows sales of high-end chips to China’s commercial sector could suppress demand for chips made without U.S. technology, delaying China’s chip independence.
The United States should seek to foster strategic dependencies by China on U.S. technology. In the U.S.-China tech competition, the leverage the United States used to kneecap Huawei is a priceless strategic advantage. Yet the Biden administration’s new chip restrictions undermine the United States’ long-term position. Because U.S. export controls will only have a temporary effect as global markets adapt over time, the United States is better off holding this leverage in reserve for now. Artificial-intelligence capabilities are rapidly advancing, and computing hardware is becoming an even more valuable input for AI power.
The deep learning revolution is only a little over 10 years old, and the second decade looks to be even more dramatic than the first. The United States will be in a stronger position for the changes ahead if it retains the ability to deny China access to powerful AI capabilities, if necessary. Using narrower export controls today could still deny China’s military access to advanced chips while keeping Chinese commercial companies reliant on foreign supplies. Keeping China dependent on U.S. technology is a stronger strategy than decoupling and will give the United States more control over China’s access to advanced technology in the long run.
No comments:
Post a Comment