China’s Overlooked AI Strategy | Foreign Affairs

China’s Overlooked AI Strategy | Foreign Affairs

In early 2025, the Chinese company DeepSeek released its R1 artificial intelligence model, sending shock waves throughout policy circles in the United States. Despite U.S. export controls on advanced semiconductors, the company had managed to develop a customizable open technology that could compete with some of the most advanced proprietary American AI models, and many feared that U.S. leadership in AI might soon be eclipsed. Now, another Chinese company, Moonshot AI, has released a state-of-the-art open model, Kimi K2, that is capable of autonomously achieving complex tasks, prompting some commentators to call it another DeepSeek moment.

But the threat posed by Chinese open models is not simply about China catching up to the United States in the AI race. It is also about the broader global adoption of AI. For the month of January 2025, the DeepSeek R1 app had 33 million active users across the world; by April, that number had nearly tripled to 97 million. Moreover, the CEO of the open-model repository Hugging Face noted that over 500 derivative versions of the original R1 model had been downloaded a combined total of 2.5 million times in January. In other words, derivative versions of R1, which are customized and tailored specifically from the original model to meet users’ needs, were downloaded five times as often as R1 itself, underscoring the value users saw in R1’s adaptability. Given this extraordinary interest, it has become clear that the low-cost, open-model approach favored by DeepSeek, Moonshot AI, and other Chinese companies could offer China an overwhelming advantage in meeting researcher demand for cutting-edge models, particularly in developing countries that are eager to access AI’s benefits.

The question of which country’s AI models achieve global preeminence has policy implications that extend beyond market competition or military applications. Open models such as R1 and Kimi K2 offer users around the world a chance to develop AI systems that can be customized for local needs, including in areas such as health care, education, and the workforce, at a lower cost than their American counterparts. In this sense, the greatest advantage open models could offer China may be in the realm of soft power. By enabling AI’s benefits to be broadly shared, Chinese open models could win international goodwill and position China as an AI benefactor to countries across the developing world, including in Africa, Asia, Latin America, and the Middle East. Moreover, Chinese government policy has embraced open-innovation ecosystems; in the United States, where companies have primarily focused on closed AI models, government policies have emphasized maintaining security and minimizing the risk that other countries, particularly China, could adapt U.S. technologies for military or economic advantage.

If Washington’s new AI strategy does not adequately account for open models, American AI companies, despite their world-leading models, will risk ceding international AI influence to China. The larger danger is that the United States will lose strategic leverage in emerging technology diplomacy in key regions of the world. Whether the country’s leading producers of AI are perceived as sharing or guarding their technologies thus has crucial ramifications that policymakers need to recognize. In striving for global AI leadership, the United States must carefully balance the need to mitigate national security risks with the imperative to bring innovative U.S. technology to other parts of the world. At present, China’s soft-power push in AI is limited by its access to computing power. With a recalibrated strategy, the United States and American AI companies have a chance to promote their own attractive, open models and prevent China’s rise as the leading purveyor of AI to the world.

MOST FAVORED NATION

It is crucial to understand that part of DeepSeek’s global success stems from its particular approach to AI. As with several other Chinese labs, DeepSeek has focused on developing smaller, more efficient models than their American counterparts that are far cheaper to train and deploy. Generally speaking, the Chinese approach has been shaped by U.S.-led export controls on advanced semiconductors, cost efficiency goals, and a drive for viable, real-world AI applications as opposed to building the most powerful general-purpose models. Moreover, China’s open models lag only slightly behind leading closed models produced by U.S. companies such as OpenAI, and their performance could persuade many countries to begin using Chinese models as the foundations of their AI infrastructure. At the core of this approach is Beijing’s broader push to gain more soft-power influence, especially in the global South, by sharing and promoting China’s advanced technologies and infrastructure—a state policy goal that has been articulated by senior government officials and reinforced in large-scale Chinese initiatives. Speaking about DeepSeek in February, for example, Foreign Ministry spokesperson Guo Jiakun made a point of stressing that China wants to share the benefits of open AI with other countries.

As it has been formulated in international relations theory, soft power is the ability to influence the preference of other actors through perceived shared values, cultural attractiveness, and persuasive ideas. Whereas hard power depends on military strength and the ability to credibly threaten the use of force to achieve strategic goals, soft power draws on such disparate sources as technology, education, and commerce to extend a country’s influence. Throughout and after the Cold War, for example, the global appeal of American culture and the U.S. economy prompted many countries to collaborate with Washington, creating trade and political interdependence among states and attracting highly skilled immigrants to the United States, who in turn helped fuel further U.S. technological breakthroughs.

Today, AI’s dual-use potential makes it a complicated asset in the soft-power portfolio. It is well documented that China and the United States are competing to develop game-changing, hard-power applications for AI, such as autonomous drones, advanced data integration, and command and control systems. In recent years, concerns about these security issues have had an outsize influence on Washington’s AI strategy. As U.S. companies introduced powerful generative AI models in late 2022, the Biden administration began to fear that China could find military applications for the new technology. As a result, its policies focused mostly on restricting access to the hardware needed to develop powerful AI models.

But that security focus has overshadowed AI’s extraordinary potential for transforming the global economy, workforce, health care, and society at large in beneficial ways—including in many developing countries. For example, if American AI models can be used to develop new cancer treatments or novel agricultural practices in countries with few resources, it would enhance the United States’ image and influence in a similar way to programs previously run by development and assistance agencies like USAID. It is precisely this potential for beneficial AI that Beijing is focused on developing and sharing—positioning itself to gain a powerful edge in global AI influence as a result.

The United States still has a chance to win the contest for global AI diffusion.

At the heart of AI’s role in the United States’ soft-power competition with China is the difference between open- and closed-model ecosystems. Open models such as DeepSeek’s R1 or Meta’s Llama family, one of the few existing big U.S. models to adopt this approach, feature open weights—the numerical parameters that determine the model’s predictions—that any user can adjust to tailor the model for a specific task. In contrast, the closed models used by most leading U.S. companies, including the models underlying OpenAI’s ChatGPT releases, require more computing resources to train, are closed off to users behind tightly controlled interfaces, and are generally more expensive to develop. Although closed models are usually more powerful, these attributes have made them in some ways less attractive to global AI developers than their open-model counterparts. Moreover, as DeepSeek’s R1 and Moonshot AI’s Kimi K2 have demonstrated, the performance gap between closed and open models appears to be shrinking.

Meanwhile, China’s government and some of its major firms are betting that open models—available for developers to access and adapt freely—can win over users and thus gain influence around the world. The release of R1 and Kimi K2 has already underscored the huge potential of China’s AI models. If successful, this strategy could entrench Chinese technology in global AI stacks, the digital infrastructure underlying AI, including the models, chips, and data centers.

Such a strategy may also cater to popular and consumer sentiment. Survey research suggests that the populations of developing countries are more optimistic about the economic and social benefits of AI than their counterparts in advanced countries. For instance, the 2025 Edelman Trust Barometer found that popular trust in AI was highest in China, India, Indonesia, Nigeria, Thailand, and the developing world generally; the proportion of Indians who indicated that they trusted AI—77 percent—was more than twice as large as the proportion of Americans who did. Similarly, in 2024, Google and Ipsos found that more than 70 percent of respondents in emerging markets such as Brazil, Mexico, South Africa, and the United Arab Emirates believed AI would positively affect work, learning, the treatment of disease, and information access, compared with around 50 percent of respondents who believed so in the United States.

Seeking to capitalize on this AI optimism, Beijing has embraced openness as a selling point. In May 2025, at an AI capacity-building conference in China, Deputy Foreign Minister Ma Zhaoxu emphasized the key advantages of Chinese AI models—open source, low cost, and high performance. Leveraging their models’ adaptability, China’s open-model developers could thus position China as a primary technological partner for countries that are looking to make a major AI push but that do not have the financial or computational resources to develop their own powerful models from scratch. For example, a local health department in a developing country might fine-tune an open model by retraining it on data sets of local disease statistics in remote regions to provide better medical services. This kind of deep customization is generally not possible with closed models.

DIGITAL EMPIRE

China’s soft-power AI push is not happening in a vacuum. It should be understood as part of Beijing’s broader effort to promote domestic self-sufficiency and expand its digital influence abroad. Since the 2010s, Beijing policy documents, including the AI Development Plan, announced in 2017, and the 14th Five-Year Plan, released in 2021, have underscored the Chinese government’s desire to leverage open-source technologies to push domestic innovation and reduce China’s reliance on the West. China’s promotion of its open AI models fits in lockstep with its aim of serving as the developing world’s leading provider of digital infrastructure through, for instance, the Digital Silk Road initiative. Over the last decade, Chinese companies such as Huawei and ZTE have invested hundreds of millions of dollars in digital infrastructure projects globally, including telecom networks, undersea cables, and surveillance equipment. Now, Alibaba Cloud and Huawei Cloud are ramping up overseas data center development for cloud computing in countries such as Malaysia, Mexico, the Philippines, and Thailand. These investments provide a strong foundation to help China deploy its AI applications around the world, particularly if it can access more high-end chips in the future than it has been able to under recent U.S. export controls.

Already, China is taking steps to solidify its global influence over the AI sector. Since 2023, Beijing has led a multilateral effort called the Global AI Governance Initiative to create responsible AI development and regulation. Although the initiative has rightfully trumpeted the importance of including developing countries in AI policy formulation, it is also light on specifics and emphasizes China’s willingness to develop others’ infrastructure. Some global AI industry leaders are concerned that Chinese open generative models could spread Chinese Communist Party propaganda to users by censoring or distorting information on Chinese history, global politics, and human rights issues.

The rapid advances of Chinese open models suggest that existing U.S. policy assumptions and private-sector AI strategy need to evolve. By prioritizing closed models, many U.S. firms have made a bet on AI as an exclusive product. At the same time, some U.S. firms appear to be giving more emphasis to cutting-edge innovation, including the goal of creating artificial general intelligence, than to developing reliable applications for existing generative models. Meanwhile, Chinese application developers are already exploring practical uses for DeepSeek’s and Alibaba’s models in automobiles, home appliances, health care, and other areas. Success in these applications is not guaranteed, but China’s push for finding practical uses for powerful AI models is notable.

To a degree, U.S. firms have begun to respond. In the aftermath of the DeepSeek R1 release, Sam Altman, the CEO of OpenAI, admitted that he felt the company was “on the wrong side of history” and would be rethinking its open-source strategy. Shortly thereafter, OpenAI announced it would be releasing an open-source model, which is expected in the late summer of 2025. Google, too, has since released a family of open-source models called Gemma, although its performance lags significantly behind the company’s closed models. What’s more, in June 2025, OpenAI dropped the price of its o3 model by 80 percent to make it more accessible to developers, researchers, and startups—and likely to be more competitive with some of the Chinese models.

OPEN OPPORTUNITY

The United States still has a chance to win the contest for global AI diffusion. China’s bid to become AI’s leading distributor already faces several hurdles. By preventing China from obtaining the tools needed to fabricate advanced chips for AI and by blocking Chinese firms’ access to U.S.-designed AI chips, American-led export controls likely will hinder China’s ability to spread its models widely. Moreover, demand for the chips needed for model inference—the deployment and use of trained models—will likely further outpace supply as inference-hungry models become widely integrated into applications.

These limits present a narrow window for the United States to offer better alternatives, but it must act quickly. Promoting the growth of an open-model AI ecosystem should be a first order priority for U.S. policymakers. The Trump administration’s newly released AI Action Plan appears to recognize the geostrategic significance of open models, recommending that “We need to ensure America has leading open models founded on American values.” Meta is the leading U.S. open-model player, but others exist, and OpenAI’s plans for its own powerful open model reflect an emerging trend. U.S. policymakers should further encourage this shift by encouraging collaboration between model developers and research institutions and helping provide additional resources to researchers and educators. Moreover, the U.S. government should offer more support to American AI researchers through programs such as the National Artificial Intelligence Research Resource, which allows small and underresourced open-model developers to better compete with larger tech developers. The AI Action plan nods toward these solutions, but they will require further fleshing out in the months ahead. A more diverse AI ecosystem would broaden the base for potential innovations that could benefit society and the economy.

At the same time, the United States should recalibrate its AI export control strategy. The recent successes of DeepSeek and Moonshot AI show that export controls cannot prevent Chinese companies from producing advanced open models. Nevertheless, the increasing importance of inference-based computing underscores the need for the United States to realign its export controls with the current direction of AI development. The controversy surrounding the Trump administration’s decision in mid-July to resume exports of Nvidia’s H20—a chip suitable for inference—underscores the need for a recalibration of the strategy. Washington’s reversal appears to have been more a concession aimed at trade negotiations with Beijing than a strategic pivot toward loosening advanced chip restrictions over the long-term. The administration will ultimately need to establish clear performance-based guidelines for chip exports as Nvidia and competitors release new chips. Controlling the semiconductors most suitable for inference could become a more central point of a future U.S. export control strategy, particularly as Chinese companies incorporate models like DeepSeek’s into new real-world applications.

By promoting open-model development and more carefully tailored export restrictions, Washington can also advance American soft power. A stronger open model ecosystem in the United States, for example, could provide U.S. allies and partners with attractive and affordable alternatives to Chinese models, fostering innovation abroad and reinforcing U.S. global influence. By taking steps to share AI’s benefits, the United States can enhance its global image, even if this must be carefully balanced with security concerns. If it fails to address China’s growing soft-power lead quickly, however, the cost may be high: Chinese diffusion of cheap, powerful AI capabilities—and the global clout that will accompany it—could ultimately prove too hard to displace.

Loading…

Source link

Visited 1 times, 1 visit(s) today

Leave a Reply

Your email address will not be published. Required fields are marked *