Table of Contents
ToggleThis article is compiled from an analysis by anthropic.com .
In its latest policy research report, Anthropic warns that the AI competition is escalating from a battle over model performance to a systemic one. The report points out that while the US and its allies still hold a significant advantage in computing power, China is rapidly closing the gap by exploiting export control loopholes and model distillation. If Washington does not now curb chip smuggling, access to overseas data centers, and distillation attacks, a situation could emerge in 2028 where the US and China are neck-and-neck in AI development, or even with China surpassing the US—this will not only be about technological leadership but will also determine who sets the rules and governance framework for future AI.
Anthropic has published a new paper outlining its views on the AI competition between China and the United States.
The United States and its allies need to maintain their lead over major competitors such as China in the field of AI. As AI's capabilities rapidly improve, this technology will soon profoundly impact social governance, national defense, and the international power structure. At the same time, the pace of AI development is accelerating, leaving limited time for all parties to set the rules of competition, manage technological risks, and shape the global governance framework. It is against this backdrop that Anthropic outlines the measures needed to ensure the United States maintains its leading position.
One of the most crucial elements in AI development is access to computing chips, or "computing power," for training models. Since the most advanced chips are primarily developed by companies within the US and its allies, the US government currently restricts China's access to these chips through export controls. Recent experience shows that these controls have been effective. In fact, Chinese AI labs have been able to develop models approaching US standards mainly due to their talent pool, exploitation of export control loopholes, and large-scale model distillation—that is, rapidly replicating some technological achievements by extracting the output and performance of US models.
In this article, Anthropic outlines two possible scenarios for the world's future in 2028. Anthropic predicts that by then, transformative AI systems will have already emerged.
In the first scenario, the United States successfully maintains its computing power advantage. Policymakers further tighten export controls, reducing China's ability to acquire cutting-edge U.S. capabilities through methods such as model distillation, and accelerating the adoption of AI by the U.S. and its allies. In this world, the U.S.-led technological ecosystem can exert greater influence on AI rules, standards, and governance frameworks. It is in this scenario that the U.S. is more likely to engage in effective communication with China regarding AI protection; Anthropic supports this to the extent feasible.
In the second scenario, the United States fails to take sufficient action. Policymakers fail to block China's access to advanced computing power, allowing Chinese AI companies to rapidly leverage these resources, approaching the forefront of AI and even surpassing it in some areas. In this world, AI rules and standards will be contested by more countries, and the most advanced models may be used for larger-scale social governance, cyber operations, and defense capabilities. Even if this situation is based on US computing power and the spillover of US technology, it is not in the long-term interests of the United States and its allies.
The United States and its allies entered the AI competition with a significant advantage. The core tools needed for AI dominance were built upon a highly innovative corporate ecosystem within the US and its allies' systems. Past success means that the most important task now is largely to avoid squandering existing advantages: to prevent China from closing the gap more easily.
The development and deployment of AI will determine the future direction of global technology rules, industry standards, and governance frameworks. Whoever can maintain a leading position in the field of AI is more likely to shape how these systems are implemented.
Currently, the United States and its allies hold a significant lead in computing power, one of the most crucial elements for developing cutting-edge AI models. This lead stems from technological innovation by the US and its allies, as well as from bipartisan export control policies in the US. However, in terms of model intelligence, Chinese AI labs are not far behind. Anthropic's focus on China's AI development is not to deny the effectiveness and contributions of the Chinese people and the Chinese AI community, but because China is, besides the United States, the only country with sufficient resources and top talent, and is systematically approaching the forefront of AI.
China has already applied AI technology in areas such as information censorship, social governance, cybersecurity, and military effectiveness. Chinese AI labs possess world-class talent. What truly limits their ability to continue approaching the forefront is the constraint of computing power. Part of the reason Chinese labs have been able to maintain a near-cutting-edge position is that they have exploited loopholes in US export control policies and obtained some of the performance of US models through large-scale model distillation, thereby accelerating their own model training and performance approximation.
With the rapid expansion of computing power and the increasing use of AI to enhance the training of new models, Anthropic is entering a period of rapid acceleration in AI efficiency. The so-called "genius kingdom in the data center"—which Anthropic understands as a transformative level of AI intelligence—may be just around the corner. This acceleration makes policy action more urgent.
To date, due to persistent issues such as export control circumvention and model distillation, China's AI system has been able to continue advancing close to the forefront curve. However, if the US and its allies take action now to address both the issues of access to computing power and the spillover of model performance, they could still potentially secure a 12- to 24-month lead in frontier performance. By 2028, such a lead would be of significant strategic importance. This advantage would also enhance the effectiveness of communication between US and Chinese AI experts regarding AI protection and governance, and Anthropic supports such engagement. But the window of opportunity to secure this lead will not last forever.
The window for the computing power advantage has only 12-24 months left.
Here, Anthropic proposes two possible scenarios for the state of the US-China AI competition in 2028. The first scenario is that the US and its allies establish a significant lead in model intelligence, application adoption, and global distribution. This scenario is likely to occur if policymakers act now to tighten controls on the advanced computing power of Chinese laboratories, reducing their space to catch up by distilling the best US AI models, and to accelerate the adoption of AI by the US and its allies.
The second scenario is that China is competitive near the forefront. This scenario would occur if policymakers do not continue to build on their current lead or relax restrictions on Chinese companies' access to advanced computing power.
Many in the U.S. Congress and the Trump administration have voiced support for export controls, curbing model distillation attacks, and promoting the export of U.S. AI technology stacks. As these policies progress, Anthropic hopes the U.S. and its allies can secure a significant lead by 2028, avoiding a close race with China two years later.
Anthropic anticipates that cutting-edge AI will have a profound impact on the economy and society in the coming years, as described in *Machines of Loving Grace* and *The Adolescence of Technology*. Anthropic's mission is to ensure that humanity can navigate the transition to transformative AI in a protective and beneficial way. Anthropic believes that a successful transition will lead to significant breakthroughs in medicine, invention, and economic growth.
Whether this transition can proceed smoothly depends in part on which technological ecosystems first build the most powerful systems. The industrial systems, regulatory environments, and governance frameworks in which the most advanced AI resides will shape the rules governing the development and deployment of this technology. In turn, these rules will also influence whether this technology is protected, whose protection it protects, and which interests it ultimately serves.
If the frontiers of AI are primarily defined by systems that use it for military superiority, cyber operations, social governance, and information control, then this technological transformation will face greater uncertainty and security risks.
Historically, the effectiveness of large-scale governance and monitoring has often been limited by the cost of human intervention. Powerful AI systems can reduce these costs, enabling automated governance, identification, and decision-making to be implemented on a larger scale. Therefore, China's leading position in AI could have a significant impact on the global landscape of AI governance and protection.
China possesses vast economic, military, and national governance resources. It is also the only country outside the United States with well-resourced and highly concentrated AI laboratories, and is rapidly approaching the forefront of AI development. Furthermore, China places great importance on becoming a leading AI power. Beijing has already invested tens of billions of dollars in China's AI and semiconductor industries.
China has already applied AI systems in areas such as information censorship, social governance, cyber operations, and the construction of security capabilities. The deployment of related technologies in some regions, including facial recognition, biometric data collection, and communications surveillance, also demonstrates AI's potential for large-scale governance. Cutting-edge AI systems will make these capabilities more cost-effective to maintain, have wider coverage, and be more automated. As these technologies spread overseas, AI may be used by more countries to strengthen governance and surveillance capabilities. The AI frontier led by China may significantly change the way technology is used and governance models are implemented globally.
Cutting-edge AI will shape the future balance of military power. China has identified AI as a key variable in the future battlefield and is advancing the intelligentization of its military system. Chinese military strategists view the "intelligentization" of military forces as a crucial path to approaching and ultimately enhancing their military effectiveness. The Chinese military has begun procuring commercially developed AI systems for military purposes, including deploying DeepSeek models to coordinate unmanned vehicle swarms and enhance cyber operations effectiveness.
These capabilities do not spread slowly. When a new model reaches a new level of performance in areas such as autonomous targeting, vulnerability discovery, or cluster collaboration, the party that controls it can put it into practical use within weeks, rather than years.
The risks will compound further as cutting-edge AI becomes an accelerator for other core technologies. Advanced AI models will be able to compress R&D cycles in fields such as semiconductors, biotechnology, and advanced materials. Leadership in cutting-edge AI will enable a nation to continuously expand its advantage across its entire national defense technology stack.
If a Chinese AI lab develops a model at the level of Claude Mythos Preview before its American counterpart, China will be the first to acquire a system capable of independently discovering and identifying software vulnerabilities, potentially enabling it to further enhance cybersecurity capabilities. The future performance of these models will improve exponentially, thus having a greater impact on the defense interests of the United States and other countries.
China rapidly replicates US AI through model distillation.
The competition between US and Chinese AI labs could make industry- and government-led safeguards and governance efforts more difficult. If Chinese labs closely follow or reach the same level as US models, private AI companies in both the US and China may feel greater pressure to release new models and products faster, before completing adequate safeguard assessments for deployment. Governments may also be reluctant to enact policies that encourage responsible AI development and deployment for fear of falling behind.
While a growing number of researchers in Chinese AI labs and policy circles are paying attention to AI protection risks, this trend has not yet translated into protection practices comparable to those of US labs. As of last year, only three of China's 13 top AI labs had published protection assessments, and none disclosed assessments of chemical, biological, radioactive, and nuclear (CBRN) risks. The AI Standards and Innovation Center (CAISI) found that, under a common jailbreak technique, DeepSeek's R1-0528 model responded to 94% of clearly malicious requests, compared to 8% for the US reference model. This pattern continues in recently released models. For example, an independent assessment of Moonshot's Kimi K2.5, released in April of this year, found that it failed to reject a higher percentage of CBRN-related requests compared to leading US models.
More seriously, Chinese labs frequently release models with dual-use (military and civilian) capabilities in the form of open weights. Once the model's weights are made open, existing protective measures can be removed, allowing any state or non-state actor to use the model for malicious purposes, including cyberattacks and CBRN abuse, which were originally designed to prevent such abuse.
Anthropic supports the United States and other countries in adopting policies to establish and maintain a defensive, near-term lead over China in terms of AI sophistication, domestic adoption, and global distribution. This lead is crucial for protecting the national defense interests of the United States and its allies and preventing the misuse of AI technology. It is also a fundamental prerequisite for ensuring that the United States and its allies are well-positioned in future global AI governance.
Anthropic deeply respects the Chinese people and the achievements of the Chinese AI community. Anthropic hopes that China will maintain peaceful relations with the world. Anthropic's concerns specifically point to the risks that any powerful nation's acquisition of cutting-edge AI systems may pose to global security and governance.
Where feasible, Anthropic supports international dialogue with Chinese AI experts on AI protection. Regardless of where AI is developed and deployed, the world shares a common interest in protecting it. Cutting-edge AI systems may pose a range of risks that require communication between the United States and China. Identifying shared challenges and promoting relevant concepts to prepare for and mitigate these risks is in the mutual interest of both sides.
Constructive engagement is most promising when the United States maintains a significant performance advantage. Establishing a leading margin in the responsible development and deployment of cutting-edge AI will enhance the effectiveness of the United States in influencing AI defense practices in China and elsewhere.
Mythos Preview, a model released by Anthropic to select partners in April as part of Project Glasswing, indicates an acceleration in performance, making policy action more urgent. After gaining access to this model, Firefox patched more vulnerabilities last month than it would in its entire 2025 plan, nearly 20 times its average monthly vulnerability patching rate for 2025. Regarding this model, a Chinese cybersecurity analyst wrote that China was "still sharpening its knives, while the other side suddenly set up a fully automatic Gatling gun."
Cutting-edge AI performance will rapidly approach the transformative AI vision of a "genius haven in the data center." This acceleration will be driven by the logic of the augmentation suite law: as computing power and data input increase, model performance will predictably improve; at the same time, AI itself will be increasingly used to accelerate the development of new models.
Anthropic will likely view 2026 as a window of opportunity for the US to achieve a breakthrough lead in AI. US labs possess state-of-the-art AI models, a significant lead in both the quantity and quality of advanced AI chips needed to drive the frontiers, and a substantial capital advantage through revenue and funding to support related investments. Chinese labs do possess real advantages: world-class innovative talent, abundant and inexpensive energy, and vast amounts of data. These are all necessary conditions for developing cutting-edge intelligence. However, they lack sufficient domestic computing power to compete, as well as sufficient revenue and capital to fund this competition.
The United States and China are engaged in a competition for strategic advantage over cutting-edge technologies such as AI. Public statements from both Beijing and Washington reflect this assessment. Calling this competition a "race" might be misleading, suggesting a finish line that guarantees victory for one side. In reality, it will be an ongoing struggle for dominance. Whether democratic or non-democratic nations will ultimately shape the values, rules, and norms of the AI era depends on the trajectory of this long-term competition.
This competition is unfolding on four fronts:
Smart Efficiency: Which countries are able to develop the most efficient AI models?
Domestic adoption: Which countries are most effective at integrating AI into the business and public sectors?
Global distribution: Which countries are capable of deploying AI technology stacks to support the global economy?
Resilience: Which countries are able to maintain political stability during economic transition?
Two possible scenarios for 2028: US dominance or a competitive race?
Of these four fronts, intelligence effectiveness is paramount. Anthropic anticipates that cutting-edge model effectiveness will have the most profound impact on geopolitical competition. Model effectiveness is also a core factor driving market adoption and global distribution.
However, intelligent efficiency alone is not enough. If China can integrate near-cutting-edge AI systems into its economic and defense systems more quickly and effectively, and promote the global adoption of low-cost, subsidized AI, then even if there are gaps in model intelligence, China may gain an advantage sufficient to offset these gaps. Beijing's "AI+" initiative and its emphasis on "embodied intelligence" demonstrate its policy direction of prioritizing the integration of cutting-edge intelligence into its economic and national systems. The Trump administration's AI action plan, and its focus on "promoting the export of US AI technology stacks," also illustrates the strategic advantages brought about by promoting global adoption.
While this article will not focus on the "resilience" front, Anthropic believes it will be a crucial aspect of the AI competition. Maintaining stability, cohesion, and effective policymaking during this period will be a core advantage; conversely, for countries unable to achieve this, it will become a vulnerability.
Computing power—the advanced semiconductors needed to train and deploy cutting-edge AI—is a core investment on each of these competitive fronts. The global battle for AI leadership is, to a large extent, a battle for computing power. Over the past decade, model performance has increased with the scale of computing power, and historically, most improvements in AI performance have come from the large-scale use of computing power.
Furthermore, computing power is used not only to train new models but also to support users in using AI, also known as "inference" performance. Whether training the most intelligent models or deploying them in commercial and national defense sectors, computing power is crucial. Top talent, massive amounts of data, and breakthroughs in core algorithms are certainly important in the race for intelligence; however, without sufficient computing power, these investments will hardly have a real impact.
Currently, democracies are vying for leadership in computing power. Some worry that export controls could accelerate China's efforts to develop its own advanced chip supply chain, but there is little evidence to suggest that China's self-sufficiency efforts can challenge the US and its allies' dominance in advanced computing technologies. Even before the export controls were implemented, Beijing had already invested heavily in China's chip industry, launching major industrial policies such as "Made in China 2025" and the National Integrated Circuit Industry Investment Fund. Despite this state-supported investment, Chinese AI labs and chip manufacturers remain constrained by US and allies' 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's product roadmaps found that, in terms of total processing performance, Huawei will only be able to produce products equivalent to 4% of NVIDIA's total computing power in 2026, and this figure will drop to 2% by 2027. More importantly, NVIDIA is only one part of the computing power ecosystem of the United States and its allies. Google and Amazon are also accelerating the production of their respective chips, namely TPU and Trainium, to meet the needs of cutting-edge AI labs and their customers in the United States.
Further exacerbating China's computing power shortage is its limited progress in several of the most technologically complex segments of the semiconductor supply chain. Without access to extreme ultraviolet (EUV) lithography technology, especially as policymakers further close loopholes in deep ultraviolet (DUV) lithography and its service maintenance, Chinese chipmakers will struggle to produce a sufficient quantity and quality of chips to challenge the United States' computing power leadership. China's inability to mass-produce high-bandwidth memory further widens this gap. One study estimates that if the U.S. tightens restrictions on China's access to U.S. computing power, the U.S. could gain approximately 11 times the computing power of China's AI industry.
The leading computing power comes from two main reasons.
The first reason is the continuous innovation of companies like NVIDIA, AMD, Micron, TSMC, Samsung, and ASML in democratic economies such as the United States, Japan, South Korea, Taiwan, and the Netherlands. It is these companies that have collectively built the unique technologies needed 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-thinking and decisive policy actions taken by the past three US administrations. Bipartisan policy actions, by restricting access to the US AI technology stack by companies under Chinese jurisdiction, have protected the innovation engine of the US and its allies. Anthropic's CEO has also publicly commented on the importance of export controls. In recent years, these controls have limited the sale of high-end AI chips and semiconductor manufacturing equipment to China, hindering China's development of cutting-edge AI despite Beijing's significant investment of national resources in this field. Without action to restrict China's access to US computing power, China could potentially possess all the necessary conditions to develop AI comparable to or even superior to that of the US.
Some observers worry that restricting access to computing power will force Chinese AI labs to innovate in other areas, thus weakening the US lead. While Chinese labs are indeed innovating, so far, their innovations have not been sufficient to compensate for their computing power gap. Algorithmic improvements are both a function of computing power and a multiplier of it, not a substitute for it. Discovering these algorithmic advancements is itself a highly computationally dependent process: more computing power means labs can perform more experiments, thus discovering more algorithmic improvements. This cycle will tighten further as more cutting-edge models become involved in AI development, helping to build their own next generation. In short, computing power advantage will further translate into algorithmic advantage, and ultimately into a sustained lead in AI itself.
Currently, US-based cutting-edge systems are estimated to be at least several months ahead of top Chinese models in terms of intelligence, although such estimates inevitably contain uncertainties. While China's open-source weighted models have gained considerable attention, they lag behind closed-source cutting-edge models in enterprise adoption, and public market investors are beginning to focus on their commercialization. Furthermore, Chinese AI labs appear to be moving away from the open-source route, opting instead to keep their best models proprietary.
Blocking vulnerabilities: Chip smuggling and overseas data centers
Leaders in China's AI sector have also acknowledged the impact of export controls and the core demand for US chips. Executives at China's top AI labs have expressed concern that China will fall further behind due to computing power limitations. Leading Chinese labs cite computing power scarcity as a major constraint on accelerating model performance improvements, and export controls as the cause of this constraint. An executive at a major Chinese cloud provider stated that supplying China with export-controlled US chips would have a "huge, really huge" impact, adding that any supply gap would severely affect China's AI development; he also refuted concerns that importing US chips would slow down China's self-sufficiency efforts. The main voices within China advocating the "ineffectiveness of export controls" seem to originate more from official statements and state-owned media, likely aimed at influencing US policymakers.
While export controls have been effective in establishing the current advantage, they are still insufficient. Although China cannot manufacture enough advanced chips domestically, nor can it legally purchase them overseas, Chinese AI labs still manage to maintain a near-cutting-edge position in model intelligence through two workarounds.
The first method is circumventing the acquisition of computing power, including smuggling AI chips directly into China or accessing overseas data centers. The second method is illegal model access, which involves distilling cutting-edge US models and using these models as tools to accelerate one's own AI research and development.
It's an open secret that China circumvents US export controls. For example, US federal prosecutors have charged a co-founder of Supermicro 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 models using advanced US chips that are prohibited from being sold to China. The Financial Times reported that Alibaba and ByteDance are now using export-controlled US chips to train their flagship models in data centers 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 gaining access to advanced US computing power.
Distillation attacks are another tactic used to approximate U.S. counterparts and weaken the impact of export controls. In this approach, Chinese labs create numerous fake accounts to bypass access controls on U.S. AI models and systematically collect their outputs to replicate cutting-edge performance. This allows these labs to free-ride on decades of basic research, billions of dollars in investment, and the collaborative development of cutting-edge models by top engineers worldwide. As a result, China can obtain near-cutting-edge performance at extremely low cost, effectively subsidized by the U.S. From a long-term national defense perspective, this equates to systematic acquisition of industrial intelligence on core technologies. 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 to China's AI development. A recent article in a state-run media outlet described distillation attacks targeting US models as "backdoors" relied upon by Chinese AI labs, calling them a core component of their business models. A former ByteDance researcher stated that Chinese AI labs use distillation as a shortcut to train models, thereby avoiding investment in building their own data pipelines.
U.S. policymakers have acted swiftly to address this threat. The White House Office of Science and Technology Policy issued a memo on distillation attacks. Senior officials from the White House, the U.S. War Department, and members of Congress have also expressed concern about the issue. Recently, legislation proposed by the U.S. House Foreign Affairs Committee to address distillation attacks has been unanimously passed by the committee.
If policymakers in the United States and its allies can block the two channels supporting the development of Chinese AI models—circumventing access to computing power and unauthorized access to models—Anthropic could have a rare opportunity to lock in a lead.
Below, Anthropic describes two hypothetical future scenarios to illustrate how policy actions taken today will shape the competitive landscape in 2028.
The United States' computing power advantage remains solid. Despite increased state support for China's semiconductor industry, Chinese chipmakers still lag behind the US and its allies by several years, partly due to their lack of access to advanced semiconductor manufacturing equipment, related services, and maintenance. The computing power gap between the US and China is widening as US and allied chip manufacturing capabilities come online and advanced chipmakers continue to develop more efficient and powerful chips.
At the same time, U.S. policymakers are taking action to close loopholes in the U.S. economic protection apparatus. With more abundant law enforcement resources, efforts to smuggle chips into China and access export-controlled chips in overseas data centers are becoming increasingly difficult.
Therefore, US AI models are 12 to 24 months ahead in terms of intelligence and efficiency, and this lead is widening. A few AI labs are at the forefront with the most intelligent, powerful, and efficient models, and all of these labs are located in the United States. The "nation of geniuses in data centers" has already become a reality in core industries such as cybersecurity, finance, healthcare, and life sciences.
When cutting-edge U.S. labs release new models in 2028 that achieve a leap in performance—similar to the relative impact of the Mythos Preview in April 2026—China may not achieve similar AI performance until 2029 or 2030. This will buy democratic countries crucial time to develop rules and standards for advanced AI systems.
US AI Exports: Locking in Trusted Infrastructure
American AI has become the infrastructure of the global economy, driving new economic and scientific vitality. The Trump administration's efforts to promote domestic AI adoption and boost US AI exports have been effective; powerful AI is being widely adopted both domestically and internationally, and the resulting benefits are driving unprecedented economic growth and technological progress. The global adoption rate of US AI has risen sharply. The efficiency and computing power leadership of democratic countries means that Chinese AI companies will find it difficult to compete for global market share outside of a few national markets. The fact that the world's top-tier, cutting-edge AI systems are shaped by democratic values also makes it more difficult for certain countries to use AI systems to infringe on rights and civil liberties.
Cybersecurity and other national security advantages have further expanded. Public and private sector cybersecurity professionals are using advanced AI systems to narrow the attack surface of the United States and other democracies, and to weaken China's ability to gain and maintain cyber foothold in relevant systems, thereby making national security assets, intellectual property rights, and communications networks more secure. The overwhelming AI advantage of the United States has also become an important force in containing external risks.
A self-reinforcing cycle will further solidify the leadership of democratic nations. Overwhelming AI superiority makes the United States and its allies more attractive partners. This alliance expands the US AI market and the coalition for setting global AI standards. In turn, this promotes the development and deployment of secure, reliable AI systems that protect civil liberties. The world's top technology and scientific talent continues to flow to the heart of cutting-edge technology development. The US also gains significant leverage to push for cooperation with Beijing on core issues such as AI governance, strategic competition, and trade.
This cycle will continuously reinforce itself: the lead strengthens the alliance, and the alliance strengthens the lead; the international order dominated by democratic countries will also be anchored in the transition to transformative AI.
China's AI development and deployment are approaching the forefront in model intelligence. Despite weaker semiconductor manufacturing efficiency, models trained in Chinese AI labs are only a few months behind their American counterparts. Continued distillation attacks, access to overseas computing power, weak enforcement of semiconductor manufacturing equipment exports, and the relaxation of US semiconductor export controls have all contributed to China's progress. Continued access to cutting-edge US AI for research and development has also enabled Chinese AI labs to close the gap and approach their American counterparts.
Commercial and national adoption is progressing rapidly. Beijing is driving domestic adoption nationwide through its "AI+" policy. Even though Chinese AI models are slightly less effective than their American counterparts, China's efforts to promote adoption have been fruitful. As a result, China is better positioned to deploy near-cutting-edge AI capabilities in the economic, military, and technological fields, thus shifting the balance of power in its favor.
China's AI-enabled cyber capabilities pose a serious threat. By integrating AI-enabled cyber capabilities into its already highly sophisticated cyber power structure, China's military remains a threatening cyber competitor. Related cyber actors have gained greater access to core infrastructure and dual-use infrastructure in the United States and most other countries, enabling them to interfere with core national defense and societal functions. As AI becomes more deeply integrated into core systems, even if democratic countries were to develop this technology first, they will be unable to gain an advantage over China in AI defense.
Beijing is winning global adoption through cost and the flexibility of localized deployment. Huawei and Alibaba's data centers are widespread globally, particularly in low-cost markets in the Global South, but not limited to these regions. These data centers rely on older chip upgrade kits, and China is able to export these chips because its domestic market can meet demand by obtaining export licenses to purchase US chips, smuggling chips into China, or remotely accessing overseas data centers. These data centers host second-tier models produced in Chinese labs; while not top-of-the-line, they are cheaper and still effective.
Similar to Huawei's past strategy of "cheap and sufficient," China's near-cutting-edge models and hardware support a significant and rapidly growing segment of the global economy. This infrastructure advantage will give China considerable influence in relevant markets.
To ensure that the first scenario is ultimately achieved, Anthropic supports the following policy action directions.
Plug the loopholes: smuggling chips, access to overseas data centers, and semiconductor manufacturing equipment.
Currently, Chinese labs obtain export-controlled US chips through smuggling and overseas data centers, while loopholes in semiconductor manufacturing equipment controls are accelerating their self-sufficiency efforts. Tightening regulations and increasing enforcement budgets would help close these vulnerabilities supporting China's AI ecosystem. This would lower China's computing power ceiling and correspondingly slow its AI progress, thereby maintaining and expanding the AI lead of democratic countries. It's important to note that a lower computing power ceiling could also substantially weaken distillation attacks, as Chinese AI labs would still need to reach a certain computing power threshold to effectively conduct illicit distillation.
Protecting Anthropic's innovations: restricting model access and curbing distillation attacks.
Policymakers in the U.S. Congress and the executive branch can continue to support policy actions to punish and deter distillation attacks originating from Chinese laboratories, while also taking steps to help U.S. laboratories themselves detect and prevent such attacks. These measures could include legislating to explicitly define distillation attacks as illegal and promoting threat intelligence and technology sharing among U.S. peer laboratories and between laboratories and the U.S. government. Curbing such activities could substantially extend the lead of democracies in the coming months and years.
Promote US AI exports.
As AI is increasingly adopted by the global public and commercial sectors, the Trump administration should continue to promote the global adoption of credible AI hardware and models developed and shaped by democratic principles. Locking in credible U.S. infrastructure now can prevent the Chinese AI ecosystem from gaining the global foothold it needs to compete on cost and adoption rates in the future.
The United States and its allies have developed some of the world's most powerful cutting-edge AI models and possess the most advanced core AI technologies globally. This gives them a significant advantage. If Anthropic can maintain its priority access to these technologies, this advantage can continue to grow. However, if these technologies are handed over directly to competitors, this advantage will be lost. The decisions made by policymakers this year will determine the future of transformative AI. Anthropic supports those committed to ensuring that the United States and its democratic allies continue to lead the world in 2028.
[ Original Link ]
Related reports
