In July, the Trump administration released an artificial intelligence action plan titled “Winning the AI Race,” which framed global competition over AI in stark terms: whichever country achieves dominance in the technology will reap overwhelming economic, military, and geopolitical advantages. As it did during the Cold War with the space race or the nuclear buildup, the U.S. government is now treating AI as a contest with a single finish line and a single victor.
But that premise is misleading. The United States and China, the world’s two AI superpowers, are not converging on the same path to AI leadership, nor are they competing across a single dimension. Instead, the AI competition is fragmenting across many domains, including the development of the most advanced large language and multimodal models; control over computing infrastructure such as data centers and top-of-the-line chips used to train and run models; influence over which technologies and standards are used throughout the world; and integration of AI into physical systems such as robots, factories, vehicles, and military platforms. Having an edge in one area does not automatically translate into an advantage in the others. As a result, it is plausible that Washington and Beijing could each emerge as leaders in different parts of the AI ecosystem rather than one side decisively outpacing the other across the board.
This outcome is even more likely in the wake of the Trump administration’s decision to lift some export controls on advanced AI chips to China. In December, President Donald Trump announced that the U.S. government would permit the sale of Nvidia’s H200—the company’s second most powerful AI chip—to approved customers in China. The decision reflects a belief that allowing China access to “good enough” computing power can generate revenue for U.S. companies and reinforce American technological standards without risking the United States’ edge in AI innovation. But the danger of selling high-end U.S. chips to China is that it could lead to a more divided AI landscape—one in which U.S. firms maintain a lead in providing advanced AI-based services, but Chinese companies gain ground in disseminating their slightly less advanced but cheaper technology around the world and applying AI to machines, factories, and infrastructure.
The most plausible outcome of the AI race, then, may not be decisive American or Chinese victory, but something more complex and more consequential: an asymmetric form of AI bipolarity. In a world without a clear winner, the United States will need to adapt to a longer-term competition while engaging China to manage the shared risks that advanced AI is likely to produce.
PLAYING CATCH-UP
The United States still enjoys a clear advantage at the cutting edge of AI. The world’s most capable large language models and multimodal systems are produced by U.S. firms such as OpenAI, Google, and Anthropic. These models demonstrate superior reasoning and tool-use capabilities—such as autonomously writing and debugging code, querying live databases, and analyzing spreadsheets—and anchor the most commercially valuable AI services, including AI assistants that help manage cloud platforms, productivity software, and customer service.
But the United States’ lead at the frontier is narrower than it once appeared. Chinese firms including DeepSeek, Alibaba (through its Qwen models), and Moonshot AI (with its Kimi series) are catching up. For many practical applications, such as drafting text, summarizing and translating documents, writing routine code, or powering customer service chatbots, the difference between the best U.S. models and the best Chinese ones is already marginal.
For now, the United States’ most significant advantage lies not in models but in compute—the quality and quantity of computing resources to train and run AI models. U.S. companies design the world’s most advanced AI chips, primarily through Nvidia, and the United States is far ahead of China in the scale of AI data centers. U.S. firms control roughly 70 percent of global AI compute, whereas Chinese companies control around ten percent. This capacity allows U.S. companies to train larger and more capable models and absorb the enormous computational costs of customers making requests of models in ways that Chinese competitors cannot easily match. U.S. companies, such as Amazon, Google, Meta, and Microsoft, plan to spend trillions of dollars on specialized chips, AI-focused data centers, and the energy infrastructure to power them over the next few years, likely widening the computing power gap between the United States and China, at least in the near term.
Export controls that were enacted during Trump’s first term and dramatically strengthened under the Biden administration reinforced this advantage. Restrictions on advanced AI chips and on semiconductor manufacturing equipment have made it difficult for Chinese firms to acquire or produce sufficient quantities of leading-edge chips for AI, which has slowed China’s ability to create the computing power required to train and deploy the most advanced models.
China has still managed to make some decent chips. Huawei’s Ascend 910 series—the best Chinese semiconductors—perform about 60 to 70 percent as well as Nvidia’s H100 or H200 on some AI workloads. But Huawei can make only hundreds of thousands of them, whereas Nvidia currently produces and exports millions of far more capable AI chips each year.
HANDS OFF
China has access to vast quantities of data and deep pools of AI talent. It can also easily and quickly build AI-related infrastructure and generate the energy to power it. Access to computing power, then, remains the single most binding constraint on China’s global AI ambitions—a constraint that the Trump administration just eased with its decision to allow some Chinese firms to buy Nvidia’s H200 chips. Although Chinese companies still won’t have access to Nvidia’s newest Blackwell generation or its forthcoming Rubin line, the H200 remains highly capable. It was released in 2024, is still used in major AI data centers run by U.S. companies, and is about ten times more powerful than the chips that could be sold to China under U.S. President Joe Biden’s export regulations. The Trump administration has hinted that other U.S. chipmakers, including AMD and Intel, might also be permitted to sell advanced chips.
The White House seems to believe that allowing the sale of powerful but not leading-edge chips will generate revenue for U.S. firms that can be put toward research and development while preserving U.S. leadership at the frontier of AI research. The Trump administration also reasons that continued Chinese reliance on U.S.-designed hardware and software—particularly Nvidia’s CUDA platform—will enable the United States to influence programming frameworks, development tools, and data-center architectures used by Chinese AI firms. Another motivation seems to be the conviction that selling chips that outperform China’s domestic alternatives could reduce Beijing’s incentives to speed up indigenous development of advanced AI chips.
The risks of selling H200 chips to China, however, outweigh the benefits. Depending on the number of H200 chips that ultimately reach China and how efficiently they are used, the United States could lose its massive advantage in compute capacity. According to analysis by the Institute for Progress, if the United States exported no advanced chips to China, its compute capacity in 2026 would be more than ten times that of China’s. With aggressive H200 exports, however, the U.S. advantage could dwindle to the single digits—or, under some scenarios, disappear. In other words, with unrestricted H200 exports, Chinese AI labs could build supercomputers approaching the performance of top U.S. systems, albeit at a higher cost.
Just as important, exporting H200s is unlikely to slow China’s efforts at making its own advanced chips in the long run. China’s domestic chip production is constrained by manufacturing bottlenecks, not by lack of demand. Since Trump’s December announcement, Chinese firms have already placed orders for more than two million H200s—far exceeding what Huawei or other Chinese companies can currently produce. As a result, U.S. chip sales are likely to add to, rather than substitute for, China’s total available compute. Moreover, there are some signs that Beijing may require potential buyers of the H200 chips to justify why domestic alternatives will not suffice, suggesting that Chinese authorities are prepared to maintain artificial demand for homegrown chips through procurement mandates and restrictions on foreign hardware in sensitive sectors.
The United States still enjoys a clear advantage at the cutting edge of AI.
The decision to export H200 chips to China also risks eroding the broader export controls that the United States has negotiated with its allies. In 2019, the Netherlands—home to ASML, the world’s leading manufacturer of advanced lithography equipment—agreed to restrict exports of its most sophisticated tools to China, recognizing that these machines are essential for producing leading-edge semiconductors. Dutch officials are now asking why they should continue to limit exports of critical manufacturing equipment when U.S. firms are allowed to sell the finished chips produced using the same equipment. If the Netherlands or other key allies, such as Japan and South Korea, were to loosen their export controls, China’s ability to domestically produce high-end chips could improve sharply—eventually undercutting not only Nvidia but also U.S. data center companies that rely on sustained hardware advantages.
The implications of the Trump administration’s export reversal, however, extend beyond China’s domestic market. Chinese firms such as Alibaba, ByteDance, and Tencent are increasingly building and operating—or partnering to expand—data center infrastructure in Africa, Latin America, the Middle East, and Southeast Asia. Even if Beijing restricts H200 imports for domestic use, these companies could deploy U.S.-designed chips overseas, offering subsidized, vertically integrated AI infrastructure bundled with power, connectivity, and talent programs.
China is already skilled at disseminating its technology to other countries. U.S. labs typically rely on proprietary, closed-weight models that are accessed through cloud services. They are powerful and easy to use but tightly controlled by their developers and difficult for customers to modify. Chinese firms, by contrast, have embraced open-weight models, which are appealing because they are cheaper, can be more easily tailored to specific industries or languages, and can be run through local rather than U.S.-based cloud providers—which, in turn, reduces concerns about data localization and foreign dependence. Although these open-weight models are generally less reliable than leading U.S. systems, China’s approach embeds its AI in global AI ecosystems.
The Trump administration is keen to promote the global diffusion of an American AI technology stack in which U.S. data centers, chips, and models are bundled together and the world remains dependent on U.S. hardware, software, and services. But in the wake of the H200 decision, Chinese firms are likely to build data centers in foreign countries using advanced U.S. chips running attractive Chinese open-weight models. This is not an American AI stack; it is a U.S.-enabled Chinese one.
AI, ROBOT
Even if the United States continues to lead at the AI frontier—and even if U.S. cloud providers remain the backbone of global AI services—it may not be sufficient to beat China in the AI race. This is because beyond models, compute, and diffusion lies another dimension of the race that may prove decisive: embodied AI. Unlike models that generate text or images, embodied AI systems integrate sensing, perception, control, and decision-making to operate in physical environments. They underpin industrial robots, autonomous vehicles, and intelligent machines that learn by interacting with the world.
Here, China may be particularly well positioned. Beijing has explicitly elevated embodied AI as a national priority. Central government plans have identified intelligent manufacturing and humanoid robotics as critical emerging industries, while local governments have offered grants, tax incentives, subsidized land, and preferential procurement to firms deploying AI-enabled automation. Beijing, Guangdong, Hubei, Shanghai, and Zhejiang are piloting large-scale programs focused on humanoid robotics and industrial automation, often pairing research institutes with manufacturing partners to accelerate real-world testing and deployment.
These efforts are already translating into productivity gains. AI-enabled automation has helped Chinese factories reduce defect rates, shorten production cycles, and operate continuously with fewer workers. According to the International Federation of Robotics, China’s stock of industrial robots exceeded two million in 2024. That year, Chinese factories installed roughly 300,000 new robots—more than the rest of the world combined—whereas U.S. factories put in place just 34,000. Some Chinese factories for electronics and electric cars are already operating with minimal human supervision.
In the years to come, the benefits of AI will not only depend on making smarter models but also on turning bits into atoms—that is, translating the gains from greater intelligence into economic productivity, industrial competitiveness, and novel military capabilities. All of this hinges on the ability to embed intelligence into machines that act in the real world and shape the real economy—areas in which China is well positioned to dominate.
A GRUELING DECATHLON
Taken together, these trends point toward an emerging end state that defies simple narratives of victory or defeat. The AI race is no longer a sprint toward a single finish line nor is it even a marathon. Instead, the United States and China are competing in an AI decathlon and the United States must shift its strategy accordingly. As AI bipolarity comes into view, the United States must lean into its strengths where they matter most, disseminate its technology, and accept the necessity of sustained AI dialogue with Beijing even amid intensifying rivalry.
For starters, Washington should avoid further eroding its advantages in computing power—the bedrock of global AI leadership. If the administration is unwilling to reverse its H200 decision, and Congress does not intervene, the Commerce Department should approve licenses for H200 exports slowly and apply heightened scrutiny to Chinese firms with close ties to China’s national security agencies.
At the same time, the Trump administration must follow through on the pledge it made in its AI action plan to rigorously enforce remaining export controls—most notably on high-bandwidth memory, advanced lithography equipment, and other critical semiconductor manufacturing and packaging tools needed to produce top-tier chips. These controls remain among the few ways Washington can influence the tempo and scale of China’s progress in AI.
To compete with China’s “good enough” AI infrastructure and open-weight model strategy, the administration should direct the International Development Finance Corporation and the Export-Import Bank to fund AI projects in countries across the so-called global South to compete with subsidized, state-backed Chinese alternatives—and work with Congress to secure expanded legal authorities and financial resources to do so. Absent such efforts, Chinese AI ecosystems are likely to become the default option in many developing countries, providing Beijing new levers of influence and entrenching AI models and standards that normalize surveillance and censorship.
Simultaneously, the administration must work with Congress to prepare for the domestic economic shocks that AI bipolarity is likely to intensify. As the United States comes to dominate services on the frontier of AI, a growing number of white-collar jobs—particularly entry-level positions—will probably be displaced by machine intelligence. Meanwhile, China’s dominance of the industrial applications of AI risks further hollowing out U.S. manufacturing and creating new dependencies on Chinese goods and supply chains.
These trends cannot be reversed, but their risks can be mitigated. The administration should work with Congress to invest more in STEM education, vocational training, and midcareer retraining and should encourage the adoption of AI systems that complement rather than replace human labor. Washington must also update labor laws, regulations, and safety guidelines to account for the ways AI is likely to reshape the workforce and deploy AI aggressively in health care, education, and government to make public services more efficient, accessible, and affordable. At the same time, reducing dependence on China will require supporting AI-enabled manufacturing in the United States, strengthening supply chains with trusted partners, and ensuring that productivity gains from automation translate into durable economic capacity at home.
Although AI bipolarity will likely intensify tensions between the United States and China, it also strengthens the case for sustained superpower dialogue. In a bipolar AI ecosystem, neither side can fully insulate itself from the risks generated by the other. Both countries therefore have strong incentives to manage those risks, including by coordinating efforts to prevent nonstate actors from using AI for catastrophic cyber or biological attacks—something Trump’s AI action plan warns is a growing danger. They also share an interest in ensuring that increasingly advanced AI remains under human control, even if Washington and Beijing disagree profoundly on the values those systems should reflect. In a world where the AI race is multifaceted and where neither side is likely to emerge as a clear winner, avoiding a destabilizing race to the bottom remains more important than ever.
Loading…
