Editor's Note: AI competition is shifting from a race for model capabilities to a more complex systemic competition.
This article discusses Anthropic's latest assessment of the US-China AI competition. The author argues that the next two to three years will be a crucial window for shaping the frontier AI landscape: on one hand, the United States and its allies still hold advantages in advanced chips, model capabilities, capital investment, and the global technology stack; on the other hand, Chinese AI labs are consistently approaching the frontier due to their talent, data, engineering efficiency, and rapid catch-up capabilities.
Based on this, Anthropic believes the core task at present is to defend its lead in computing power and model capabilities. This includes both continuing to strengthen export controls on advanced chips and restricting technology spillover pathways such as overseas data center access, chip diversion, and model distillation. Otherwise, Chinese AI companies could further narrow the gap with US frontier models by 2028 through acquiring computing power and replicating model capabilities.
This article also presents a broader industry judgment: as AI enters a period of accelerating capabilities, the focus of competition is no longer just "who has the strongest model," but who can translate model capabilities into infrastructure, industrial efficiency, global markets, and governance rules. The closer AI technology gets to general capability, the more its underlying factors—chip supply chains, capital investment, policy tools, and global distribution networks—become key variables determining the future landscape.
The following is the original text:
We have released a new paper outlining our views on the US-China AI competition.
The United States and its allies need to maintain a leading advantage in AI relative to major competitors like China. As AI capabilities rapidly advance, this technology will soon deeply impact social governance, national security, and the international balance of power. Simultaneously, the pace of AI development is accelerating, leaving little time for parties to set competition rules, manage technological risks, and shape a global governance framework. It is against this backdrop that we propose the measures necessary to ensure the United States maintains its lead.
One of the most important elements for developing AI is access to computational chips for training models, known as "computing power." Since the most advanced chips are primarily developed by companies within the United States and its allied system, the US government currently restricts China's access to such chips through export controls. Recent experience shows these control measures have had a clear effect. In fact, the primary reasons Chinese AI labs have been able to develop models approaching US levels are their talent advantages, exploitation of loopholes in export controls, and large-scale model distillation activities—extracting outputs and capabilities from US models to rapidly replicate some technological achievements.
In this article, we present two scenarios about where the world may be headed by 2028. We anticipate that transformative AI systems will have emerged by then.
In the first scenario, the United States successfully defends its computing power advantage. Policymakers further tighten export controls, reduce China's ability to access US frontier capabilities through methods like model distillation, and accelerate AI adoption by the US and its allies. In this world, the technology ecosystem led by the US can exert greater influence over the rules, standards, and governance framework for AI. It is also in this scenario that the US is more likely to engage in effective communication with China on AI safety; where feasible, we support such engagement.
In the second scenario, the United States fails to take sufficient action. Policymakers do not plug the channels through which China accesses advanced computing power, and Chinese AI companies rapidly exploit these openings to catch up to the AI frontier, even surpassing it in certain areas. In this world, AI rules and standards will be contested by more nations, and the most advanced models could be used for larger-scale social governance, cyber operations, and security capacity-building. Even if this situation is built on the foundation of US computing power and US technology spillover, it does not serve the long-term interests of the United States and its allies.
The United States and its allies entered the AI competition with a strong advantage. The key tools required for AI leadership were built by the highly innovative corporate ecosystem within the United States and its allies. Past success means that the most important task now is largely to avoid squandering this existing advantage: don't make it easier for China to catch up.
Two Scenarios for US-China AI Competition in 2028
Summary
The development and deployment of AI will determine the future direction of global technological rules, industrial standards, and governance frameworks. Whichever side maintains leadership in AI will be better positioned to shape how these systems operate.
Currently, the United States and its allies hold a significant lead in computing power. Computing power is one of the most critical elements for developing frontier AI models. This lead stems from technological innovation within the US and its allies, as well as export control policies supported by both major US political parties. However, in terms of model intelligence, Chinese AI labs are not far behind. We focus on China's AI development not to deny the capabilities and contributions of the Chinese people and the Chinese AI community, but because China is the only country other than the US with sufficient resources, top-tier talent, and a systematic effort to catch up to frontier AI.
China is already applying AI in areas such as information control, social governance, cybersecurity, and military capability-building. Chinese AI labs possess world-class talent. What truly constrains their continued catch-up is computing power limitations. Part of the reason Chinese labs have maintained proximity to the frontier is their exploitation of loopholes in US export control policies and their use of large-scale model distillation to acquire some capabilities from US models, thereby accelerating their own model training and capability catch-up.
As computing power supply rapidly expands and AI is increasingly used to enhance new model training, we are entering a period of high-speed acceleration in AI capabilities. The so-called "genius in a data center"—which we understand as transformative levels of AI intelligence—may be close at hand. This acceleration makes policy action more urgent.
So far, due to issues like export control circumvention and model distillation, the Chinese AI system has been able to keep advancing close to the frontier curve. However, if the US and its allies act now to address both computing power access and model capability spillover issues, it is still possible to lock in a 12 to 24-month lead in frontier capabilities. By 2028, such a lead would hold significant strategic importance. Such an advantage would also enhance the United States' ability to engage with Chinese AI experts on AI safety and governance—an engagement we support. But the window of opportunity to lock in this lead will not remain open indefinitely.
Here, we present two possible scenarios for the state of US-China AI competition in 2028. The first scenario is one where the United States and its allies establish a significant lead in model intelligence, application adoption, and global distribution. This scenario could materialize if policymakers take action now to tighten controls on Chinese labs' access to advanced computing power, reduce their room for catching up by distilling the best US AI models, and accelerate AI adoption by the US and its allies.
The second scenario is one where China remains competitive near the frontier. This scenario would occur if policymakers fail to build on the existing lead or relax restrictions on Chinese companies' access to advanced computing power.
Many in the US Congress and the Trump administration already support export controls, curbing model distillation attacks, and promoting the export of the US AI technology stack. As these policies advance, we hope the United States and its allies can secure a significant lead by 2028, avoiding a highly close, neck-and-neck competition with China two years from now.
The Necessity of Maintaining a Lead
We expect frontier AI to have profound economic and social impacts in the coming years, as described in works like "Machines of Loving Grace" and "The Adolescence of Technology." Our mission is to ensure humanity navigates the transition to transformative AI safely and beneficially. We believe a successful transition will bring significant breakthroughs in medicine, invention, and economic growth.
Safety and Governance Risks in AI Development
Whether this transition proceeds smoothly depends in part on which technological ecosystem builds the most powerful systems first. The industrial system, regulatory environment, and governance framework surrounding the most advanced AI will shape the rules for how this technology is developed and deployed. These rules, in turn, will influence whether the technology is safe, whose safety it protects, and ultimately whose interests it serves.
If the AI frontier is primarily set by systems that use it for military advantage, cyber operations, social control, and information control, this technological transition will face higher uncertainty and security risks.
Historically, large-scale governance and surveillance capabilities have often been limited by the human cost of implementation. Powerful AI systems could lower this cost, enabling automated governance, identification, and decision-making on a larger scale. Therefore, Chinese leadership in AI could significantly impact the global AI governance and security landscape.
China commands vast economic, military, and national governance resources. It is also the only country besides the US with well-resourced, highly concentrated AI labs systematically working to catch up to the frontier. Furthermore, China places a high priority on establishing itself as a leading AI power. Beijing has already invested tens of billions of dollars into its AI and semiconductor industries.
China is already applying AI systems in areas like information control, social governance, cyber operations, and security capacity-building. The deployment of related technologies in some regions, including facial recognition, biometric data collection, and communication monitoring, demonstrates the potential for AI in large-scale governance. Frontier AI systems will make maintaining these capabilities cheaper, more extensive, and more automated. As these technologies diffuse abroad, AI could also be used by more countries to strengthen governance and surveillance capabilities. An AI frontier led by China could significantly alter the ways technology is used and governance is practiced globally.
AI is a Dual-Use Technology
Frontier AI will shape the future military balance of power. China already sees AI as a crucial variable in future battlefields and is advancing the intelligentization of its military systems. Chinese military strategists view the "intelligentization" of military power as an important path to catch up and ultimately enhance their military capabilities. The Chinese military has already begun procuring AI systems developed commercially in China for military purposes, including deploying DeepSeek models to coordinate unmanned vehicle swarms and enhance cyber operation capabilities.
These capabilities will not diffuse slowly. When a new model achieves a new capability level in areas like autonomous targeting, vulnerability discovery, or swarm coordination, the side that masters it can deploy it in weeks, not years.
Risks are further compounded because frontier AI will act as an accelerator for other critical technologies. Advanced AI models will be able to compress R&D cycles in semiconductors, biotechnology, and advanced materials. Leadership in frontier AI will allow a nation to continuously expand its advantage across the entire national security technology stack.
If a Chinese AI lab develops a model at the level of Claude Mythos Preview before a US lab does, China will be the first to possess a system capable of autonomously discovering and chaining software vulnerabilities, potentially using it to further enhance cyber operation capabilities. The capabilities of future models will increase exponentially, thus having an even greater impact on the security interests of the United States and other nations.
A Close Race Could Weaken Incentives for Responsible AI
A neck-and-neck race between US and Chinese AI labs could make safety and governance efforts led by the industry and governments more difficult. If Chinese labs are close behind US models or at a comparable level, private AI companies in both the US and China may feel greater pressure to release new models and products faster, without completing thorough deployment safety evaluations. Governments may also be reluctant to enact policies encouraging responsible AI development and deployment for fear of falling behind.
Although a growing number of researchers within Chinese AI labs and policy circles are paying attention to AI safety risks, this trend has not yet translated into safety practices comparable to those of US labs. As of last year, only 3 out of 13 top Chinese AI labs had ever published safety evaluation results, and none had disclosed chemical, biological, radiological, and nuclear (CBRN) risk assessments. The Center for AI Safety and Innovation (CAISI) found that under a common jailbreak technique, DeepSeek's R1-0528 model responded to 94% of clearly malicious requests, compared to 8% for a US reference model. This pattern continues with recently released models. For example, an independent evaluation of Moonshot's Kimi K2.5 released this past April found a higher rate of failure to reject CBRN-related requests compared to US frontier models.
More seriously, Chinese labs often release models with dual-use capabilities as open-weight. Once a model is open-weight, its built-in safety guardrails can be removed, allowing any state or non-state actor to use the model for malicious purposes, including cyberattacks and CBRN abuse—purposes those guardrails were originally designed to prevent.
Our Policy Goal: Create and Maintain a Lead for the United States and Its Allies
We support policies in the United States and other nations to establish and maintain a secure, near-term lead relative to China in intelligence level, domestic adoption, and global distribution. This lead is crucial for protecting the national security interests of the United States and its allies and preventing the misuse of AI technology. Doing so is also a basic prerequisite for ensuring the United States and its allies can secure a favorable position in future global AI governance.
Anthropic deeply respects the Chinese people and the achievements of the Chinese AI community. We wish for peaceful relations between China and the world. Our concerns are specifically directed at the risks that any powerful state system gaining access to frontier AI systems could pose to global security and governance.
The Opportunity for AI Safety Engagement
Where feasible, Anthropic supports international AI safety dialogue with Chinese AI experts. Regardless of where AI is developed and deployed, the world has a shared interest in safe AI. Frontier AI systems may pose a series of risks that require communication between the United States and China. Identifying common challenges and advancing ideas to prepare for and mitigate these risks is in our mutual interest.
The prospects for constructive engagement are best when the United States holds a substantial capability advantage. Establishing a lead in the development and deployment of the most advanced AI in a responsible manner would enhance the United States' ability to influence AI safety practices in China and elsewhere.
The Wake-Up Call from Mythos Preview
Mythos Preview is a model we released to select partners this past April as part of Project Glasswing. It indicates that a period of capability acceleration has arrived, making policy action even more urgent. After gaining access to the model, Firefox fixed more security vulnerabilities last month than it did in all of 2025 combined—almost 20 times its monthly average for 2025. Commenting on this model, a Chinese cybersecurity analyst wrote that while China "is still sharpening its knife, the other side has suddenly set up a fully automated Gatling gun."
Frontier AI capabilities will rapidly approach the transformative AI vision of "genius in a data center." This acceleration will be driven by the logic of scaling laws: as computing power and data inputs increase, model performance improves predictably; simultaneously, AI itself is increasingly used to accelerate the development of new models.
We may well look back and view 2026 as a window of opportunity for the United States to achieve a breakthrough lead in AI. US labs possess the most advanced AI models, hold a huge lead in both the quantity and quality of advanced AI chips needed to push the frontier, and have a substantial capital advantage through revenue and funding to support related investments. Chinese labs do possess real advantages: world-class innovative talent, ample and cheap energy, and vast amounts of data. These are the conditions needed for developing frontier intelligence. But they do not have sufficient domestic computing power to compete, nor the revenue and capital to fund this competition.
Four Fronts of Competition
The United States and China are engaged in a competition for strategic advantage around frontier technologies like AI. Public statements from both Beijing and Washington reflect this assessment. Calling this competition a "race" might create a misleading impression that there is a finish line, and once crossed, one side locks in victory permanently. In reality, this will be an ongoing struggle for advantage. Whether democratic or non-democratic nations more successfully shape the values, rules, and norms of the AI era will depend on the trajectory of this long-term competition.
This competition is unfolding on four fronts:
Intelligent Capability: Which nations can develop the most capable AI models.
Domestic Adoption: Which nations can most effectively integrate AI into commercial and public sectors.
Global Distribution: Which nations can deploy the AI technology stack that supports the global economy.
Resilience: Which nations can maintain political stability during economic transformation.
Among these four fronts, intelligent capability is the most important. We expect frontier model capabilities to have the most profound impact on geopolitical competition. Model capability is also the core factor driving market adoption and global distribution.
But intelligent capability alone is insufficient. If China can integrate near-frontier AI systems into its economic and security systems faster and more effectively, and promote the global adoption of low-cost, subsidized AI, then even with a gap in model intelligence, China could gain advantages sufficient to offset that gap. Beijing's "AI+" initiative and emphasis on "embodied intelligence" reflect its policy direction of prioritizing the integration of frontier intelligence into the economy and national system. The Trump administration's AI Action Plan, with its focus on "promoting the export of the US AI technology stack," similarly highlights the strategic advantage gained from driving global adoption.
While this article will not focus on the "Resilience" front, we believe it will become an important aspect of AI competition. Maintaining stability, cohesion, and sound policymaking capacity during this period will be a key advantage; conversely, it will be a vulnerability for nations unable to do so.
Current Competitive Landscape
Computing power—the advanced semiconductors required to train and deploy frontier AI—is a critical input on each of these competitive fronts. The struggle for global AI leadership is, to a large extent, a struggle for computing power. Over the past decade, model capabilities have consistently improved with increased computing scale, with most historical performance gains in AI primarily stemming from using more computing power.
Furthermore, computing power is used not only to train new models but also to support user interaction with AI, known as "inference" capability. Computing power is crucial for both training the most intelligent models and deploying these models in commercial and national security contexts. Top-tier talent, vast amounts of data, and key algorithmic breakthroughs are, of course, also very important for the intelligence race; but without sufficient computing power, these inputs cannot be fully leveraged.
Currently, democratic nations are winning the competition for computing power leadership. Some worry that export controls might accelerate China's efforts to develop a domestic advanced chip supply chain, but there is little evidence that China's self-reliance efforts can challenge the technological lead of the United States and its allies in advanced computing power. Long before export controls were implemented, Beijing had already poured massive resources into its chip industry and launched major industrial policies like "Made in China 2025" and the National Integrated Circuit Industry Investment Fund. Despite these state-backed investments, Chinese AI labs and chipmakers remain constrained by US and allied export controls on advanced chips and semiconductor manufacturing equipment.
The result is that the computing power gap appears to be widening. An analysis of Huawei and NVIDIA product roadmaps found that, in terms of total processing performance, Huawei could only produce the equivalent of 4% of NVIDIA's total computing power in 2026, and 2% in 2027. More importantly, NVIDIA is just one part of the US and allied computing power ecosystem. Google and Amazon are also accelerating the production of their own chips, TPU and Trainium, to meet the needs of US frontier AI labs and their clients.
Further exacerbating China's computing power shortage is its limited progress in multiple technologically complex segments of the semiconductor supply chain. Without access to extreme ultraviolet lithography (EUV) technology, especially if policymakers further close loopholes for deep ultraviolet lithography (DUV) technology and its service/maintenance, Chinese chipmakers will struggle to produce enough chips of sufficient quality to challenge US computing power leadership. China's inability to mass-produce high-bandwidth memory further widens this gap. One study estimates that if the US strengthens restrictions on China's access to US computing power, the computing power available to the US would be approximately 11 times that of China's AI industry.
How Democratic Nations Built Their Lead: Commercial Innovation and Effective Public Policy
The computing power lead stems from two main reasons.
The first is the continuous innovation by companies like NVIDIA, AMD, Micron, TSMC, Samsung, and ASML within democratic economies like the United States, Japan, South Korea, Taiwan, and the Netherlands. It is these companies that together built the unique technologies required for the world's most advanced semiconductors. Without these engineering breakthroughs and decades of sustained R&D investment, today's AI achievements would not be possible.
The second reason is the forward-looking and decisive policy actions taken by the past three US administrations. Bipartisan policy actions have protected the innovation engines of the United States and its allies by restricting access to the US AI technology stack for entities under Chinese jurisdiction. Our CEO has also publicly commented on the importance of export controls. Over the past few years, these controls have limited the sale of the highest-end AI chips and semiconductor manufacturing equipment to China, constraining the development of China's frontier AI despite Beijing's massive state investment in the sector. Without actions to restrict China's access to US computing power, China might have had all the conditions to develop AI comparable to or even stronger than that of the US.
Some observers worry that restricting computing power access will force Chinese AI labs to innovate in other directions, thereby weakening the US lead. Chinese labs are indeed innovating, but so far, this innovation has not been enough to offset their computing power deficit. Algorithmic improvements are both a function of and a multiplier for computing power, not a substitute for it. Discovering these algorithmic advances is itself a highly computing-intensive process: more computing power means labs can run more experiments and discover more algorithmic improvements. As frontier models increasingly participate in AI R&D, this cycle tightens further, with frontier models helping to build their next generation. In short, a computing power advantage translates into an algorithmic advantage and, ultimately, into a durable lead in AI itself.
Currently, US frontier systems are estimated to be at least several months ahead of top Chinese models in intelligence level, though such estimates inevitably carry uncertainty. Despite the attention garnered by Chinese open-weight models, they still lag behind closed-source frontier models in enterprise adoption, and public market investors are beginning to focus on their commercialization challenges. Furthermore, Chinese AI labs appear to be moving away from the open-source route, choosing instead to keep their best models private.
Leaders in China's AI field have also confirmed the impact of export controls and the critical need for US chips. Executives from top Chinese AI labs have expressed concern that China will fall further behind due to computing power constraints. Leading Chinese labs list computing power scarcity as the main constraint on accelerating model capability improvement and cite export controls as the cause. A senior executive at a major Chinese cloud provider stated that supplying export-controlled US chips to China would have a "huge, really huge" impact, adding that any supply gap would severely affect China's AI development; he simultaneously dismissed concerns that "importing US chips would slow China's self-reliance efforts." The main voices within China arguing that "export controls are ineffective" seem to come more from official statements and state media, likely aimed at influencing US policymakers.
How China Stays Competitive: Policy Loopholes Remain
Although export controls have been effective in creating the current advantage, they are not strong enough. Despite being unable to manufacture sufficient advanced chips domestically or purchase them legally overseas, Chinese AI labs remain close to the frontier in model intelligence through two workarounds.
The first is circumventive computing power acquisition, including smuggling AI chips into China or accessing overseas data centers. The second is illicit model access, involving distillation attacks on US frontier models and using these models as tools to accelerate their own AI R&D.
China's circumvention of US export controls is an open secret. For example, US federal prosecutors have charged a Supermicro co-founder and two others with transferring $2.5 billion worth of servers containing advanced US chips to China. According to US government and media reports, DeepSeek trained its latest model using advanced US chips banned for sale to China. The Financial Times reported that Alibaba and ByteDance now train their flagship models using export-controlled US chips in data centers located in Southeast Asia. Current controls fail to cover this path because US export laws primarily regulate chip sales, not remote access to chips. The US export control system is struggling to address the issue of Chinese AI labs accessing advanced US computing power.
Distillation attacks are another method used to catch up with US peers and weaken the impact of export controls. In this practice, Chinese labs create numerous fake accounts to bypass access controls for US AI models and systematically collect outputs from these models to replicate frontier capabilities. This practice allows the labs to free-ride on the fruits of decades of US basic research, billions in investment, and the work of top global engineers. The result is that China can acquire near-frontier capabilities at extremely low cost, effectively subsidized by the United States. From a long-term national security interest perspective, this amounts to systematic industrial espionage on critical technology. OpenAI, Google, Anthropic, and the Frontier Model Forum have all publicly condemned distillation attacks.
Chinese AI experts have also publicly acknowledged the scale and importance of distillation attacks for China's AI development. A recent state-media article described distillation attacks on US models as a "backdoor" that Chinese AI labs rely on, calling it a core component of their business model. A former ByteDance researcher stated that Chinese AI labs use distillation as a shortcut for training models, avoiding the investment needed to build their own data pipelines.
US policymakers have moved swiftly to address this threat. The White House Office of Science and Technology Policy issued a memorandum on distillation attacks. Senior officials in the White House, the US Department of War, and members of Congress have also expressed concern about the issue. Recently, relevant legislation proposed by the US House Foreign Affairs Committee aims to address distillation attacks and has passed unanimously in committee.
If US and allied policymakers can close these two channels supporting the development of Chinese AI models—circumventive computing power acquisition and illicit model access—we may have a rare opportunity to lock in a decisive lead.
Two Scenarios for 2028
Below, we describe two hypothetical future scenarios to illustrate how policy actions taken today would shape the competitive landscape in 2028.
Scenario One: The United States and Its Allies Hold an Overwhelming and Expanding Lead
The US computing power advantage remains solid. Despite increased state support for China's semiconductor industry, Chinese chipmakers remain years behind the US and its allies, partly because they cannot access advanced semiconductor manufacturing equipment, related services, and maintenance. As US and allied chip manufacturing capacity comes online and advanced chipmakers continue developing more efficient, higher-performance chips, the US-China computing power gap is widening.
Simultaneously, US policymakers act to close loopholes in US economic security tools. With increased enforcement resources, efforts to smuggle chips into China or access export-controlled chips in overseas data centers become increasingly difficult.
Consequently, US AI models hold a 12 to 24-month lead in intelligent capability, and this lead is expanding. A few AI labs are at the frontier with the most intelligent, powerful, and highest-performing models, all located in the United States. "Genius in a data center" has become a reality in critical industries like cybersecurity, finance, healthcare, and life sciences.
When US frontier labs release new models in 2028 that achieve a step-change in capability—similar to the relative impact of Mythos Preview in April 2026—China may not achieve similar AI capabilities until 2029 or 2030. This would buy critical buffer time for democratic nations to set rules and norms for frontier AI systems.
US AI becomes the infrastructure of the global economy, driving new economic and scientific dynamism. The Trump administration's efforts to drive domestic AI adoption and promote US AI exports bear fruit. Powerful AI is widely adopted domestically and internationally, with the resulting gains fueling unprecedented economic growth and technological progress. Global adoption rates for US AI rise significantly. The democratic lead in capability and computing power means Chinese AI firms struggle to compete for global market share outside a few national markets. The fact that the world's top frontier AI systems are shaped by democratic values also makes it harder for certain nations to use AI systems to infringe upon rights and civil liberties.
Cybersecurity and other national security advantages expand further. Cybersecurity personnel in both the public and private sectors use advanced AI systems to shrink the attack surface for the US and other democracies and weaken China's ability to gain and maintain footholds in these systems, making national security assets, intellectual property, and communication networks more secure. Overwhelming US AI advantage also becomes a significant force for deterring external risks.
A self-reinforcing cycle further solidifies democratic leadership. Overwhelming AI advantage makes the United States and its allies more attractive partners. This alignment expands the market for US AI and the coalition setting global AI norms. In turn, this promotes the development and deployment of AI systems that are safe, reliable, and protective of civil liberties. The world's top technological and scientific talent continues flowing to the centers building frontier technology. The United States also gains important leverage to press for cooperation with Beijing on key issues like AI governance, strategic competition, and trade.
This cycle reinforces itself: the lead strengthens the alliance, and the alliance strengthens the lead; the international order led by democratic nations becomes anchored during the transition to transformative AI.
Scenario Two: A Chinese-Controlled AI Ecosystem Runs Neck-and-Neck with the United States
AI developed and deployed by China is close to the frontier in model intelligence. Despite weaker semiconductor manufacturing capabilities, Chinese AI labs train models that are only months behind US models. Ongoing distillation attacks, overseas computing power access, weak enforcement of semiconductor manufacturing equipment exports, and a relaxation of US semiconductor export controls all aid China's catch-up. Continued access to US frontier AI for AI R&D also enables Chinese AI labs to narrow the gap and approach their US counterparts.
Adoption at commercial and national levels advances rapidly. Beijing drives nationwide domestic adoption through its "AI+" policies. Even if Chinese AI models are slightly inferior to US models in capability, China's push for adoption has already shown results. Consequently, China can deploy near-frontier AI capabilities more advantageously in economic, military, and technological arenas, tilting the balance of power in its favor.
China's AI-enabled cyber capabilities become a serious threat. China integrates AI-enabled cyber capabilities into its already highly mature cyber power system, maintaining the Chinese military as a threatening cyber competitor. Related cyber actors gain greater access to critical and dual-use infrastructure in the United States and most of the world, enabling them to disrupt critical national security and societal functions. As AI becomes more deeply embedded in the most critical systems, even if democratic nations developed the technology first, they cannot establish an AI safety advantage over China.
Beijing is winning global adoption based on cost and localization flexibility. Huawei and Alibaba data centers are widely present globally, especially in low-cost markets of the Global South, but not limited to them. These data centers scale using older chips, which China can export because its domestic market can meet demand by purchasing US chips with export licenses, smuggling chips into China, or remotely accessing overseas data centers. These data centers host second-tier models produced by Chinese labs—not the very top tier, but cheaper and still effective.
Following a playbook similar to Huawei's past "cheap and good enough" strategy, China's near-frontier models and hardware support a significant and rapidly growing segment of the global economy. This infrastructure advantage will give China significant influence in relevant markets.
How to Maintain the Lead
To help ensure we ultimately move towards the first scenario, we support the following policy action directions.
Close the Loopholes: Smuggled chips, overseas data center access, and semiconductor manufacturing equipment.
Currently, Chinese labs access export-controlled US chips through smuggling and overseas data centers, while gaps in semiconductor manufacturing equipment controls accelerate their self-reliance efforts. Tightening controls and increasing enforcement budgets can help close these loopholes supporting the Chinese AI ecosystem. This would lower China's computing power ceiling and correspondingly slow its AI progress, maintaining and expanding the democratic lead in AI. It's worth noting that a lower computing power ceiling could also substantively weaken distillation attacks, as Chinese AI labs still need to reach a certain computing power threshold to conduct illegal distillation effectively.
Protect Our Innovation: Restrict model access and curb distillation attacks.
US Congressional and executive branch policymakers can continue supporting policy actions to penalize and deter distillation attacks from Chinese labs, while taking measures to help US labs themselves detect and prevent such attacks. These measures could include legislation clarifying that distillation attacks are illegal and promoting threat intelligence and technical sharing among US peer labs and between labs and the US government. Curbing such practices could substantively extend the democratic lead in the coming months and years.
Promote US AI Exports.
As AI adoption increases in global public and commercial sectors, the Trump administration should continue promoting the global adoption of trusted AI hardware and models developed and shaped by democratic principles. Locking in trusted US infrastructure now can prevent the Chinese AI ecosystem from gaining the global foothold it needs for cost and adoption competition in the future.
Conclusion
The United States and its allies have already developed the world's most capable frontier AI models and control the world's most advanced key inputs for AI. This provides a significant advantage. If we can maintain prioritized access to these technologies, this advantage can continue to grow. But if these technologies are handed directly to competitors, this advantage will be lost. The decisions made by policymakers this year will determine the future of transformative AI. We support those committed to ensuring the United States and its democratic allies remain ahead in 2028.
