America Can Win the AI Race

Throughout history, technology has been critical to determining which countries dominate global politics. By rapidly industrializing in the 1800s, Germany and the United Kingdom overtook Russia in economic strength. Europe’s broader industrialization had an even more profound effect. In 1790, Europe, China, and India held roughly the same shares of global manufacturing output, but by 1900, Europe—then home to a quarter of the world’s people—controlled 62 percent of the world’s manufacturing. By contrast, China had six percent and India had less than two.

European powers translated their economic might into military power, launching a wave of colonial expansion. By 1914, Europeans occupied or controlled over 80 percent of the planet’s land surfaces. The states were able to make this translation because the Industrial Revolution had altered the key metrics of power, transforming coal, steel, and oil production into critical components of military success. In World War II, the United States turned its mighty manufacturing capacity to the business of war, retooling factories to build tanks and airplanes and making its military into the world’s most powerful. At the height of the war, Allied factories were producing over 3.5 times as many aircraft and tanks as the Axis powers, burying Germany, Japan, and Italy beneath an onslaught of iron.

Washington has maintained its leading position in the intervening century in large part because of technology. After the Soviet Union’s launch of Sputnik, the United States surged its investments in science and technology, building the world’s leading universities and technology companies. But technology is never static, and today the world is reckoning with an innovation that could prove just as transformative as nineteenth-century industrialization: artificial intelligence. Its powers, once largely confined to science fiction, are becoming common and ubiquitous. Just as the Industrial Revolution created machines that were physically stronger than humans, the AI revolution is creating machines that are cognitively smarter than humans. GPT-4, the successor to ChatGPT, recently achieved human-level performance on the SAT, the GRE, and the bar exam. And AI is rapidly improving. It is already transforming jobs from computer programming to fighter jet piloting, and it will continue to alter professions in the future.

Evaluating which state is leading in AI is tricky, especially because knowledge of the algorithms behind the technology can easily spread across borders. But researchers do know what drives AI advancements: massive amounts of data and computing hardware, talented AI scientists and engineers, and resources for AI initiatives. Countries can accrue clear advantages in each of these domains, and scholars can assess these as metrics of national AI power.

Right now, it is clear that the United States leads in AI, with advantages in computing hardware and human talent that other countries cannot match. But China is rapidly catching up. The AI ecosystem is highly open and breakthroughs rapidly proliferate, and Beijing has a better government strategy for advancing AI than Washington. China is ahead of the United States in AI adoption, and it has a large and growing community of high-quality AI experts. Beijing could also inadvertently benefit from U.S. immigration restrictions, which might help China keep more talent. Washington’s effort to cut off China from U.S. technology, meanwhile, could hasten the day when Beijing no longer needs U.S. computing hardware.

If the United States wants to win the AI competition, it must approach Beijing carefully and construct its own initiatives thoughtfully. It needs a strategy that will keep China dependent on foreign-made chips, and it needs to continue attracting and retaining the world’s top AI talent. It must make sure that its institutions, especially the military, fully adopt new innovations. And it must harness its existing advantages, working hard to mobilize academic, corporate, and government resources to improve its mastery of the technology that will govern the future.

ADVANTAGE, AMERICA

In the race for AI dominance, Washington has the upper hand. U.S. companies dominate the key chokepoints in the equipment needed to produce advanced semiconductors, giving the United States unparalleled leverage over the AI supply chain. This advantage is compounded by trends in AI, where the amount of computing hardware used to train cutting-edge machine-learning models is doubling every six months. The most sophisticated AI models use thousands of highly specialized advanced chips, and these chips can only be built using U.S. technology.

To capitalize on this advantage, in October 2022 the Biden administration banned companies from selling to China advanced semiconductor manufacturing equipment and AI chips made with U.S. technology. For Beijing, these restrictions could prove devastating. China is highly dependent on foreign chips, importing over $400 billion worth of them per year. The bans have the dual effect of denying China both the ability to buy high-end AI hardware and the tools it needs to build its own. In early 2023, Japan and the Netherlands—the other two primary makers of semiconductor manufacturing gear—reportedly joined U.S. controls, although the public details of the deal are murky. But provided the two states did properly join with Washington, the bans will prove highly successful. Collectively, the three nations control 90 percent of the global market for semiconductor manufacturing equipment, and so the restrictions could ensure that China’s domestic chip production falls behind as the global industry advances.

Control over chips gives the United States an advantage over China. But as AI research becomes more computing intensive, the United States will also need to invest in more computing resources at home to capitalize on its upper hand. To train the largest machine-learning models, researchers need thousands of advanced, specialized chips, at the cost of millions of dollars per project—expenses that most organizations struggle to afford. Many leading AI labs are backed by major tech corporations with deep pockets. DeepMind and Google Brain, for example, are owned by Google’s parent company. OpenAI secured a $10 billion investment from Microsoft after ChatGPT was released. Academic AI researchers, meanwhile, find themselves priced out of training large models. The U.S. government has proposed a National Artificial Intelligence Research Resource to provide academics with additional data and computing hardware. But Congress must fund these federal resources to help U.S. academics remain competitive in cutting-edge AI research.

China is bleeding AI talent.

The United States also has an advantage over China in the competition for talent. Of the top 15 institutions publishing deep learning research, 13 are American universities or corporate labs. Only one, Tsinghua University, is Chinese. U.S. universities and firms recruit the best researchers from around the world, so much so that two-thirds of the top AI scientists in the United States did their undergraduate studies overseas before coming to the country, including many who came from China.

That’s not to say that China lacks AI talent. The country is home to the world’s fastest-growing AI research community—and one that is quickly improving in quality. China puts out more AI research papers than does the United States, and the number of Chinese authors contributing to top AI journals increased by a factor of 12 from 2009 to 2019 and is now roughly 2.5 times higher than the number of U.S. contributors.

But the United States still leads in quality, with papers that are cited 70 percent more often than Chinese ones. And China is bleeding talent. A 2020 study by MacroPolo, a U.S. think tank, tracked the flow of international AI talent based on a sample of papers accepted to one of the world’s top artificial intelligence conferences. They found that although more leading AI researchers did their undergraduate studies in China than in any other country, the vast majority of these experts left to pursue their graduate work abroad. More than half of them came to the United States, and over 90 percent of those who came to the United States ultimately stayed after graduation. China may be the biggest source of AI talent, but the United States is the biggest beneficiary of the talents of Chinese researchers.

The exodus is not the only way that China’s larger population has failed to give the country a leg up in the AI race. The country has 1.4 billion people, and so Chinese researchers should have access to more data than their U.S. counterparts. But technology is not always constrained by borders, and major U.S. tech firms—the ones that are funding some of the biggest AI breakthroughs—have a global reach that exceeds that of Chinese companies. Facebook, for example, has 2.7 billion users, and YouTube has over 2 billion. WeChat, China’s biggest app, has 1.2 billion users. U.S. tech firms also have a more diverse database. With the exception of ByteDance’s TikTok, Chinese social media platforms have struggled to gain a foothold outside China, putting them at a disadvantage in gaining access to diverse data. These apps may be well suited to predict the behavior of Chinese netizens, but their AI models may not hold for outside markets.

GAINING GROUND

Washington, however, does not dominate every part of the AI race. When it comes to some sectors of data collection, for example, China may be ahead. U.S. businesses have a wider reach, but the Chinese Communist Party has built a massive domestic surveillance apparatus that will give it larger data sets and faster AI development for some applications than the United States can muster. China is home to half the world’s roughly one billion surveillance cameras, which are spread across airports, hotels, banks, train stations, subways, factories, apartment complexes, and even public toilets. This system will give Chinese companies an edge over their U.S. competitors in facial recognition, where a grassroots backlash has slowed U.S. efforts to deploy public cameras on a mass scale. Two of the United States’ largest tech companies—Amazon and Microsoft—have a moratorium on selling facial recognition to law enforcement. IBM canceled its work on facial recognition altogether. Several cities and states have even banned law enforcement from using facial recognition software.

Beijing’s advantages in facial recognition may not confer a wide advantage. Better data on Chinese faces won’t necessarily translate to non-Chinese faces, and it certainly won’t help train better AI fighter pilots. And U.S. companies have ample data not only because of their international reach but also because Americans have willingly given their personal data over to them. Beijing, meanwhile, has cracked down on the power of tech firms, including by increasing consumer data privacy protections after a series of scandals. Yet Washington’s failure to regulate tech firms is unlikely to give it much of an advantage in data. In emerging AI fields, rule-making uncertainty may hold back innovation as companies struggle to figure out what they can or cannot do. By clarifying copyright laws on training AI models and AI model outputs, the United States could help its businesses unlock new opportunities.

It is not just in the data realm where Washington could struggle. In other parts of the AI contest, China is working hard to close the gap with the United States. China is improving its domestic talent base, growing its number of scientists and engineers. It also has more than 200 talent recruitment programs focused on getting the estimated 400,000 Chinese scientists abroad to bring scientific knowledge back to China. The U.S. government has long been concerned about Chinese efforts to acquire technology adopted in the United States, and so it has responded to Beijing’s initiatives with a crackdown on espionage. But if increased investigations of Chinese researchers prompt some of them to feel unwelcome in the United States, Washington could cut off the flow of Chinese talent. Such moves would be a gift to Beijing.

Washington is also at risk of losing high-skilled immigrants from all over the world thanks to its immigration policies. A numerical cap on H-1B work visas—the visas typically awarded to college-educated immigrants—arbitrarily constrains U.S. companies from hiring foreign talent. Per-country caps on green cards mean that Indian immigrant scientists, especially, have to wait absurdly long periods before they can become permanent U.S. residents, and Indian scientists make up a quarter of Silicon Valley’s workforce. (One study estimated that Indian nationals who already have visas and are applying for employment-based green cards have to wait 89 years before they can receive one.) It is not surprising, then, that nearly 70 percent of top machine learning researchers residing in the United States said that visa and other immigration problems were an obstacle to recruiting foreign scientists. It is one policy area where government regulation is clearly harming American competitiveness.

The Biden administration has taken some steps to make it easier for science, technology, engineering, and mathematics students to come to the United States. The White House has, for example, increased the number of foreign exchange programs, expanded the number of fields qualifying for the STEM Optional Practical Training visa for recent college graduates, and made it easier for STEM Ph.D.s to apply for the O-1A “extraordinary ability” visa. These are valuable steps, but they will have a minor impact relative to the problem. The United States needs comprehensive reform of its immigration policies for high-skilled workers to make it easier for U.S. universities and companies to recruit international scientists and engineers. Washington especially has to exempt STEM Ph.D. graduates from the H-1B cap and ease their pathway to permanent residency. And it must do so soon: as a country of 330 million people, the United States will always be at a disadvantage competing against a country of 1.4 billion if it restricts itself to homegrown talent. The United States’ unique advantage is its ability to draw on the best and brightest from around the world, and it cannot afford to lose that edge.

KEEP YOUR ENEMIES CLOSER

Immigration is not the only domain where Washington’s policies could harm its AI drive. The Biden administration’s export restrictions may be hurting Beijing right now, but the bans risk hastening Beijing’s drive toward chip independence. U.S. restrictions, for instance, could turn China’s $400 billion of chip buying power inward, fueling its domestic semiconductor industry. This purchasing would add to the government’s already extensive efforts to grow domestic chip production. And once Beijing no longer needs foreign chips, the United States will have lost its leverage over its adversary.

Washington’s restrictions could also prompt companies to excise U.S. technology from their chip supply chains. Although U.S. chip restrictions cover only approximately one percent of China’s chip market today, the size of the banned market will grow over as technology advances, provided that the restrictions stay in place (as U.S. officials have promised). Even American companies have previously looked for ways to work around Washington’s prohibitions. After the U.S. government banned Huawei from receiving U.S.-made chips in 2019, Intel and Micron continued shipping chips to Huawei by cutting out components made in the United States.

Instead of implementing broad bans, U.S. policymakers should work closely with allies to maintain China’s dependence on foreign chips. The United States should prohibit sales of chips to China for military applications or that will facilitate China’s human rights abuses. But it should still permit sales to commercial data centers. Additionally, the United States should work with allies to create a broader, more multilateral system of semiconductor export controls to close off opportunities for China to circumvent U.S. controls. These steps will help keep China reliant on foreign chips built using U.S. technology, ensuring that Washington maintains the upper hand as the AI revolution progresses.

To stay ahead, the United States must also invest in new forms of microelectronics research that can ensure that U.S. companies lead in future semiconductor technology. Smaller transistor sizes have driven chip advances for decades, but as the most sophisticated chips become as tiny as is physically possible, future innovations will likely come from new areas, such as advanced packaging techniques that pack more functionality onto chips. The United States’ recent $52 billion CHIPS and Science Act is an opportunity to not only bring cutting-edge chip manufacturing back home but also to invest in new kinds of semiconductor innovations.

STRUCTURAL ADJUSTMENT

To be the world’s leading AI power, China and the United States will certainly need top-notch resources, researchers, and manufacturing. But to transform expertise and innovations into hard power, the countries need to find ways to integrate AI inventions into their militaries. It is not a simple task. Unlike stealth technology or hypersonic missiles, most AI advancements come from the civilian sector. They must then be reworked and scaled to have a battlefield impact.

Both states are aware of this challenge, and they are working hard to meet it. The U.S. Defense Department, for example, has gone on an organization building spree, creating dozens of new offices designed to bring private-sector technology into the armed forces. The department has had some positive results. The Defense Innovation Unit, for example, established in 2015 as the Defense Innovation Unit Experimental (DIUx), has contracted with 120 companies that do not traditionally do defense work and 60 companies that are new to working with the Department of Defense entirely. It has helped foster new military innovations, including by helping launch the United States’ Project Maven—which uses AI technology to process drone video feeds.

The Chinese military has followed Washington’s lead. In 2018, the country’s Central Military Commission Science and Technology Commission created a self-described “rapid response small group” for adopting market AI inventions, apparently modeled on the Defense Innovation Unit. (Some Chinese reporting even referred to the group as “China’s DIUx.”) China has used competitions modeled on the U.S. Defense Advanced Research Projects Agency to draw in private sector innovators to take on tough problems. After the United States made advancements in swarming, in which robots autonomously cooperate to perform a task, and in AI dogfighting, China carried out public demonstrations of its own advances in both technologies.

Once Beijing no longer needs foreign chips, the United States will have lost its leverage.

Even though Beijing is copying Washington’s playbook, U.S. leaders have worried that their country might ultimately fall behind China—in large part because some American tech workers are reluctant to collaborate with the U.S. military. When protests by Google employees led the company to stop working on Project Maven in 2018, U.S. military leaders panicked that they would be shut out of future game-changing commercial technology. But these fears have not come to fruition. Tech companies have, in fact, proved eager to work with the Department of Defense on artificial intelligence. When the Pentagon asked for bids on a $10 billion AI-enabled cloud computing contract, Amazon, IBM, Microsoft, and Oracle all fought to win it. There has been an explosion in defense-oriented AI startups, from now large firms such as Anduril to small ones such as Heron Systems, the latter of which beat out Lockheed Martin in a 2020 AI dogfighting competition. Even Google has returned to defense contracting.

If anything, the U.S. government appears to have the opposite problem: its commercial players are so determined to profit off the military’s needs that they are making it hard for the bureaucracy to properly adopt new tech. Tech companies launched lawsuits, filed protests, and called for investigations into the selection process for the DOD’s cloud computing contract in an attempt to take the contract for themselves. Their efforts paralyzed the program for three and a half years, before the DOD finally surrendered and canceled the contract. To proceed, it had to craft a multi-vendor solution, with Amazon, Google, Microsoft, and Oracle all on contract. To avoid being mired in more pointless delays, the United States must reform acquisition rules to cut back on objections.

The Department of Defense faces other, bigger structural obstacles. The DOD has succeeded in adopting commercial AI technology, but mostly for one-off solutions in small projects. It has not yet succeeded in integrating artificial intelligence into its core functions, and history shows that what matters most in periods of technological disruption is finding the best ways to use new inventions. The United Kingdom, for example, was the first country to develop aircraft carriers, but by the start of World War II it had fallen behind Japan and the United States, which embraced carrier aviation as central to the future of naval warfare. The two states’ advantages came thanks to bureaucratic foresight; war games at the Naval War College, for instance, showed U.S. Navy leaders the carriers’ potential. London, by contrast, made the mistake of consolidating airpower under the Royal Air Force instead of prioritizing naval aviation. Government decisions, not technology, were decisive in determining which states led in the carrier race.

The pieces of a better AI strategy are falling into place.

To properly integrate AI into the armed forces, the Department of Defense must build an iterative process of experimentation, prototyping, testing, and concept development—something it has struggled to do thus far. After Project Maven’s early success, the DOD established the Joint AI Center in 2018 to accelerate the use of AI technology, but the body had difficulty scaling applications across the rest of the department. In response, in 2022, Pentagon officials created a new chief digital and artificial intelligence office, consolidating AI and data-related functions across the DOD, including the Joint AI Center’s duties. This is a welcome shift, but it remains to be seen whether the new office will succeed in making sure AI is fully deployed across the Department of Defense. Building the right institutions will take continued attention and investment from senior defense leaders.

And ultimately, to truly integrate and capitalize on AI, U.S. defense leaders will need to shift how they measure military capability. The Pentagon will never give artificial intelligence its due when it does not consider AI to be a key component of military strength. When military leaders appear before Congress to advocate for their budgets, they make their case in terms of industrial-age metrics. The navy, for example, lays out how many ships it requires, while the air force spells out how many aircraft it has to purchase. These measurements still matter, but what matters more today is the digital capabilities of these systems, such as whether the ships and planes have sensors to detect enemy forces, algorithms that can process information and enable better decision-making, and intelligent munitions to precisely strike targets. All these capabilities can be improved with artificial intelligence, and U.S. leaders must begin taking them into account.

It will not be easy for the armed forces to make these changes: the American military is a vast and unwieldy bureaucracy. It will also be hard for the U.S. government as a whole to adjust to the rise of AI, given how polarized Washington is. Reforms to high-skilled immigration, in particular, have run into repeat resistance from conservatives on Capitol Hill. And the fact that today’s AI systems have major limitations—and therefore require great caution and care during implementation—further complicates the process. Military service members will not use systems they do not trust, and so military officials must make sure that when AI is deployed it works as intended.

But the pieces of a better AI strategy are falling into place. The Pentagon may not yet properly measure the power of artificial intelligence, but it is paying much more attention to the technology. The federal government has increased spending and is exploring data and computing resources for academics. The White House is trying to make it easier for foreign STEM workers to come to the country. The United States, in other words, is working to ensure that China cannot fully catch up. If Washington ultimately maintains control over the semiconductor supply chain, maximizes the inflow of talent, and fields trustworthy systems, it will succeed in staying ahead. As the AI revolution reshapes global power, the United States can come out on top.