Technology executives, national security analysts, and U.S. officials all seem to agree that the United States must win the AI competition with China. In October 2024, Biden administration National Security Adviser Jake Sullivan warned that the United States risked “squandering [its] hard-earned lead” if it did not “deploy AI more quickly and more comprehensively to strengthen [the country’s] national security.” And in one of its first executive orders, the second Trump administration declared its goal “to sustain and enhance America’s global AI dominance.” Washington has pursued a two-pronged strategy in its bid for supremacy: constraining China by restricting the export of key technological components and accelerating domestic innovation on foundational AI models. To achieve the latter goal, both administrations have pursued a program of relatively light regulatory oversight of industry leaders, targeted investment in semiconductors and energy infrastructure, and encouraging the adoption of AI across the federal government, especially defense and intelligence agencies, for uses as varied as investigating outbreaks of food-borne diseases and detecting financial fraud.

So far, these policies have enabled American firms to maintain their lead in market share and model performance over their Chinese counterparts. But Washington cannot and should not expect its lead to last forever. Recent breakthroughs by Chinese AI companies such as DeepSeek, Alibaba Cloud, Baidu, and Tencent suggest that the gap between U.S. and Chinese cutting-edge AI capabilities is narrowing, and that American supremacy in AI is far from assured. Even as it continues to vie for dominance, Washington needs to plan for a possible future in which the United States loses the AI competition to China—or, at the very least, one in which Chinese AI models are as popular globally. 

Preparing for the possibility that the United States will finish second, however, does not mean that Washington is doomed to repeat the failures of the 5G competition, in which the United States has struggled to provide affordable, innovative products as China charged ahead. Rather, by promoting richer frameworks that properly account for the features that make AI most attractive to emerging markets, by making it easier for developers to migrate between models, by building systems that allow users to compare the outputs of competing AI models, and by securely sharing U.S. data with model builders, partners, and allies, Washington can ensure that, even if it is unsuccessful in its bid for global AI supremacy, it will safely benefit from the AI revolution. 

BLOWN LEAD

Until the summer of 2024, the United States appeared to have a winning formula for AI dominance. Rapid model innovation emerged from a U.S. ecosystem that consisted of academic research, publications, and talent, private sector capital and development, and extremely light regulatory oversight. U.S. foundational models such as OpenAI’s GPT and Google’s Gemini had exhibited significant advancements since 2022, when ChatGPT first took the world by storm. In the past two years, AI tools have reduced instances of “hallucination” (the term used to describe the phenomenon of AI generating false or contradictory information), gained the ability to consume and produce images and sounds, mastered more complex tasks, and demonstrated enhanced reasoning capabilities. More U.S. tech companies, including Anthropic, xAi, and Meta moved quickly to develop larger models capable of mastering benchmarking tasks such as understanding speech and images, programming, and solving complex scientific problems as fast as they were concocted. As American models demonstrated their speed and mastery, their global market share and international users grew rapidly. 

By the end of 2023, the leading U.S. models were overperforming their Chinese counterparts by double-digit percentage points in response accuracy. But China has quickly closed the gap with a savvy combination of government initiatives such as the Next Generation AI Development Plan, an emphasis on AI education and workforce training, mighty research investment, close coordination between Beijing and the tech industry, and massive public investment in data centers, energy transmission, and semiconductor manufacturing. These efforts helped narrow the gulf between the performance of U.S. and Chinese AI tools on most popular benchmarks to single-digit percentages by the end of 2024. In the last six months, DeepSeek and Qwen have matched the performance of U.S. models, prompting fear that the United States’ once commanding lead has evaporated. 

Meanwhile, China has taken the lead in integrating AI into high-tech manufacturing. Xiaomi, which started as a smartphone manufacturer, for example, uses over 700 AI-guided robots in its Beijing factory to produce a new electric vehicle every 76 seconds, on average. AI is widely used in Chinese cities for traffic management, surveillance, and law enforcement, and provincial and municipal governments are experimenting with AI-innovation zones to foster new applications in governance, health, and education. 

Moreover, U.S. export controls have proved less effective as barriers against the acquisition of advanced chips by China than many policymakers and industry analysts expected. Beijing has used shell companies and stockpiled chips to get around the controls and has accelerated its own domestic chip development programs. Chinese firms have also pioneered software development techniques to maximize already available hardware in order to optimize training and inference time and lower overall costs. Whether or not these gains have definitively put China in the lead, it is clear that the days of absolute American AI dominance are over. 

MAKING THE BEST OF SECOND

American AI firms may yet retain their global lead on many aspects of building foundational AI models. OpenAI, in partnership with SoftBank and Oracle, announced a $500 billion AI infrastructure investment project known as Stargate in January, and Amazon, Meta, Microsoft, and Google all continue to invest billions of dollars in startups, AI labs, and talent. Amazon AWS, Microsoft Azure, and Google Cloud make up over 60 percent of the global cloud market, a critical resource for building and deploying models, compared with top Chinese provider Alibaba’s four percent market share. But the pace of technological innovation that has propelled U.S. companies over the past three years could fall short of China’s or prove unsustainable as advancements at the frontiers of AI discovery become increasingly difficult, the United States slips in the global competition for talent, and cuts to federal research funding undermine innovation. Washington should try to prevent this scenario. The White House Office of Science and Technology, for its part, has begun to develop a new national AI Action Plan, which is expected to be released in July. 

Still, policymakers should also plan for a world in which competing AI ecosystems coexist. Fortunately, Washington can find alternative strategies to ensure that the United States benefits from AI progress even if it does not win the innovation competition outright. 

First, the United States should look for new ways to demonstrate the merits of its models to global markets. The U.S. National Institute of Standards and Technology, through the AI Safety Institute and industry partners, could promote new evaluation frameworks for foundational AI models. Standard benchmarks mostly focus on raw capabilities such as language understanding, reasoning, and conversation abilities at the exclusion of other metrics—such as the models’ transparency, accountability, and accessibility, cost of operation, and the ease with which the models’ weights (the “knobs” to fine-tune a model to make accurate predictions) can be modified. New evaluative frameworks that incorporate these measures could be used to appeal to new markets and users, keeping U.S. companies competitive even if their models can no longer consistently outcompete Chinese models on the standard benchmarks.

As more models emerge, users will want to avoid having to decide between (and thus getting locked into) either the United States’ or China’s current offerings. In this coming world of AI consumer choice, minimizing migration costs between models will become an attractive selling point. The American AI industry can reduce the cost and complexity of transitioning to its models by lowering purchase prices and reducing the amount of software modification, hardware upgrading, and personnel training it takes to migrate between models. And Washington could lead efforts in the International Organization for Standards to standardize the application programming interfaces—the protocols that allow different software programs to communicate and exchange data—of foundational models, which could lower the costs of moving between models and reduce dependence on any one country’s AI offerings. If Chinese models do take the lead, giving global users a degree of confidence that they can tap into the benefits of multiple models, as well as switch back to American models if the United States were to become number one once again, would be a wise hedge.

Policymakers should also plan for a world in which competing AI ecosystems coexist.

As Chinese (and eventually other) models grow more powerful and penetrate global markets, Washington cannot simply highlight the risks of censorship and espionage and expect U.S. and foreign companies to refuse to adopt them. Instead, U.S. companies must build systems and applications that run on foundational models but mitigate the risks of relying on any one specific model. Incorporating another software layer between applications and foundation models, known as an intermediate abstraction layer, can isolate the downstream systems from the foundational model, making them more independent and resilient. If the foundational model changes in ways that negatively affect the performance of applications or other models running on the foundational model or a new, better foundational model emerges, companies building downstream applications can quickly shift to a different foundational model.

The adoption of Chinese foundational models by U.S. users and companies does create real risks, including vulnerability to incorrect, skewed, exploitative, and even damaging outputs; exposure of sensitive data to a potential adversary; and the threat of service disruptions that could paralyze entire economic sectors. When they do build on foreign foundational models, U.S. companies will need to make sure their downstream applications are resilient against such threats. U.S. firms should build adjudication systems capable of comparing the responses of less powerful but trusted models and untrusted but more powerful models, determining whether a response from the untrusted model is satisfactory, warning users about potential risks, and preventing the use of incorrect responses. Such systems would raise development and maintenance costs and slow down response times as they weigh the competing responses, so U.S. policymakers should prioritize integrating an adjudicating system into the downstream applications that would suffer most from the introduction of unreliable foreign technology, such as medical diagnosis, fraud identification, and traffic control.

Finally, the United States should regulate the types of data U.S. developers and companies share with foreign model builders without defaulting to blanket bans on data transfers. Washington may understandably be wary of sharing U.S. data with China for privacy and national security reasons, but there may be instances in which the economic or societal benefits of fine-tuning a Chinese model with U.S. data outweigh the risks. For example, if a Chinese AI tool emerges that makes medical diagnoses and suggests treatments with far more accuracy than its U.S. counterpart, American hospitals should use the Chinese model even at the risk of sharing individual patient information, or even data from a larger population of patients that could be used to fine-tune the model to the U.S. population. In those instances, U.S. firms can ameliorate the privacy and security risks by data anonymization; data masking, which replaces sensitive data with fake or scrambled data; and differential privacy—mathematical frameworks that allow for the sharing of the data on groups while limiting the sharing of information on individuals.

Washington will want to standardize the new evaluation metrics and to develop guidelines for sharing data with allies and friends. It will also need to provide technical and financial assistance to partners who lack the expertise to migrate between models, as well as build and apply the systems that will adjudicate competing models. 

LOSING THE RIGHT WAY

Washington faces an evolving global AI landscape in which absolute dominance is no longer assured. American policymakers cannot simply rely on jingoistic calls to win the AI race while ignoring the possibility that the country’s early lead will not hold forever. Washington should still, of course, attempt to keep its lead. A more responsible—and realistic—strategy would promote policies that help the United States thrive while preparing the country in case it fails to achieve outright dominance. 

Otherwise, Washington will face the worst possible outcome: a superior competitor with increasing economic and military power enabled by AI, and a domestic AI industry unable to keep up, handicapped by its inability to build on Chinese models if necessary. Finishing second is not a death knell for American AI, but refusing to adapt to compete would be. 

Loading... Please enable JavaScript for this site to function properly.