The Paradox of US AI Export Controls: Effectiveness of Strategies Against China's AI Militarization and Challenges for Multilateral Governance

Executive Summary

Executive Summary (Executive Summary)

The strengthening of US AI export controls has been an ad hoc accumulation of security imperatives mixed with industrial-political motives. The case of Chinese startup Z.ai launching a model with comparable performance without any restrictions, immediately after the ban on non-US persons accessing Anthropic's top-tier models, symbolically demonstrates that the current control system is failing to substantially curb China's AI capabilities while weakening the global competitiveness of US companies. The most likely future scenario is a fragmented competitive landscape where US export controls maintain partial effectiveness while China's indigenous development accelerates in parallel, solidifying a complex structure where no single strategy can simultaneously resolve all risks. Consequently, a strategic shift is urgently needed for the US government to move beyond ad hoc access restrictions and establish a risk-based, tiered regulatory framework, linked with a multilateral AI governance framework with allies. Allied governments, in turn, must pursue a dual strategy of actively building their own AI sovereign capabilities, rather than passively integrating into the US-led system. For global corporations, the key survival strategy will be to accept geopolitical fragmentation as an irreversible structural change and proactively build resilient business models adaptable to any scenario through supply chain diversification and technology portfolio restructuring.

Step 1: Issue Situation Analysis

Analysis of the Issue Situation: US AI Model Export Controls and China's Response to AI Militarization

1. Background and Progress of the Issue

Artificial intelligence (AI) has rapidly emerged as a strategic asset that determines national security, military power, and economic hegemony, extending beyond the realm of mere technological innovation. Historically, each technological revolution has fundamentally altered the nature of geopolitical competition, and the AI era is no different, with the ability to learn, innovate, and adapt more rapidly becoming a key variable of national competitiveness, in addition to traditional power indicators such as territory, population, and military strength [5]. In this context, the United States and China have engaged in all-around competition over AI technology leadership, encompassing semiconductor supply chains, model development capabilities, military applications, and international standard-setting.

The US has progressively strengthened export controls based on concerns that AI technology could be directly utilized to enhance China's military and security capabilities. Initially, the focus was on restricting exports of advanced semiconductors, particularly high-performance GPUs from Nvidia. Subsequently, the scope expanded to include access controls on AI models themselves. The US government, through executive orders, has required AI developers to provide early access to the government before releasing frontier models, and based on this, has begun earnest negotiations to establish voluntary standards for model release criteria and the scope of foreign access [2]. In this process, analyses have suggested that some key players in the US AI industry intentionally highlighted the conflict with China to create a regulatory environment favorable to themselves. It is also noteworthy that while AI competition in the US is unfolding in a 'winner-takes-all' dynamic, this characteristic is relatively less pronounced in Chinese policy documents [1].

2. Current Situation (Latest Trends)

Currently, the US government is implementing measures to block foreign access to the most powerful AI models on grounds of national security. Specifically, the Trump administration directed a ban on non-US persons accessing Anthropic's top-tier models, Fable 5 and Mythos 5 [7], stemming from an intention to block high-risk capabilities that could be weaponized, such as the detection of cybersecurity vulnerabilities. However, mere days after the implementation of this measure, the emergence of a Chinese startup, Z.ai, releasing an AI model with performance close to Anthropic's models, and at a competitive price, with no access restrictions, has amplified doubts about the effectiveness of export controls [9]. Chinese cybersecurity firm 360 has also unveiled an AI tool, 'Tulongfeng,' which it claims rivals Mythos [12], suggesting that US control measures are paradoxically accelerating China's indigenous development.

In terms of semiconductor export controls, smuggling issues are also becoming prominent. With the explosive growth in demand for Nvidia AI servers from China, Chinese companies such as Alibaba and Tencent are showing willingness to pay almost any price, and smuggling activities are surfacing across the entire AI server supply chain [13]. Consequently, Nvidia is responding by further strengthening inspections of AI servers, but complete blockage is practically difficult due to the complexity of the supply chain. Meanwhile, the landscape of the large language model market itself is rapidly changing. According to the 2026 AI Industry Report, OpenAI's ChatGPT's market share in the global AI assistant market has fallen below 50% for the first time [4], indicating that numerous competitors, including those in China, are rapidly closing the gap.

China is also taking proactive steps in the competition for technical standards. China has announced seven national standards for the interoperability of AI agents, establishing an integrated framework for how AI agents identify, discover, collaborate with, and utilize external tools [15]. This is interpreted as part of China's strategy to strengthen its technological sovereignty by preemptively establishing its own AI ecosystem standards to counter export controls.

3. Key Actors and Their Positions/Interests

The US Governmentperceives the military application of AI technology as the greatest threat and utilizes export controls and model access restrictions as key instruments. However, it faces criticism that ad hoc and piecemeal measures may instead stimulate China's indigenous development and weaken the trust of allies [3]. The core challenge for the US strategy is to ensure coherence between the two goals of maintaining technological superiority and strengthening allied cooperation.

US AI Companies (Anthropic, OpenAI, etc.)hold complex positions caught between the government's tightening regulations and their own commercial interests. Anthropic has blocked foreign access to its top-tier models in compliance with the government's export restriction order [7], but this is leading to a loss of global market share. Criticisms that some AI companies exaggerated the threat from China to create a regulatory environment favorable to themselves suggest that their interests may not align solely with national security logic [1].

The Chinese Government and AI Companiesare leveraging US export controls as a justification for strengthening their technological sovereignty and accelerating the development of an independent AI ecosystem. China plans to invest approximately 2 trillion yuan ($295 billion) in AI data center infrastructure over the next five years [17], which is part of a systematic strategy toward long-term technological independence. Furthermore, China has implemented a reverse regulation, State Council Order No. 837, to strengthen state control over technology transfer during overseas investments [6]. Chinese AI companies are expanding their global presence by successively releasing open-source models that match the performance of US models at a lower cost [9][12].

Allies and Third Countriesare seeking to secure their own AI sovereignty amidst the US-China AI cold war, while simultaneously pursuing strategic autonomy to avoid subordination to either side. Canada and France have declared their commitment to strengthening their indigenous AI capabilities [4], and India is using the direct impact of US access restrictions as an opportunity to redefine its own AI sovereignty strategy [7]. The launch of Fugu, an independent model by Tokyo-based startup Sakana AI, can be understood in the same context [12]. Global South countries are at risk of becoming dependent on either the US or Chinese AI ecosystem due to the AI accessibility gap and lack of regulatory capacity [16], and the need for non-aligned AI cooperation frameworks for them is also being raised [8].

Global Corporate Usersare accelerating their shift to Chinese open-source models as an alternative to the US government's tightening export controls. According to a UBS report, there is a clear trend of companies switching from advanced closed-source models to cost-effective Chinese open-source models, such as Alibaba's Qwen series, for tasks that do not require them [10]. This demonstrates that US export controls are inadvertently helping Chinese AI companies expand their global market presence.

4. Summary of Key Issues

First, the effectiveness of export controlsis the most critical issue. While the US is blocking access to top-tier AI models from Anthropic, Chinese companies are releasing models with similar performance, raising fundamental questions about whether control measures genuinely contribute to maintaining the technological gap [9][12]. As warned by PIIE, ad hoc measures may instead stimulate China's indigenous development, leading to unfavorable outcomes for the US in the long run.

Second, the uncontrollability of open-source modelsreveals the structural limitations of the export control system. If Chinese companies release high-performance models in open-source form, any US access restriction measures will have little practical effect. The trend of global companies switching to Chinese open-source models for cost savings clearly illustrates the severity of this problem [10].

Third, the issue of alignment with alliesis significant. The US's unilateral and abrupt access restriction measures infringe upon the AI access rights of its allies, stimulating them to pursue independent AI sovereignty paths [4][7]. The proposed establishment of a free world AI computing alliance by the Carnegie Endowment is an alternative approach to address these issues, but coordinating the interests of allied nations itself presents another challenge.

Fourth, China's AI militarization and the competition for technical standardsare critical. The AI innovation activities of China's People's Liberation Army Strategic Support Force suggest that the military application of AI is already considerably advanced. China's preemptive announcement of national standards for AI agents reflects a strategic intention to expand its influence in the global AI ecosystem by securing technical standards first. The attempt to secure superiority in wireless infrastructure through preempting 6G spectrum is also an important issue intertwined with AI militarization [14].

Fifth, the absence of global AI governanceis evident. As the US and China build their respective technological ecosystems and strengthen export controls and technology transfer restrictions, Global South countries are at risk of becoming dependent on one side due to the AI accessibility gap and lack of regulatory capacity [16][8]. As the preliminary report of the UN Independent International Panel of Scientists points out, the unequal distribution of AI benefits and the lack of regulatory capacity in developing countries remain fundamental challenges for AI governance.

Step 2: In-depth Issue Analysis

In-depth Analysis of the Issue: US AI Model Export Controls and China's Response to AI Militarization

1. Analysis of the Root Causes of the Issue

The US's strengthening of AI export controls ostensibly aims to prevent technology leakage, but at its core lies a more complex structural anxiety. The most critical root cause is the shift in perception that AI has transcended the category of mere commercial software to become an element of national security infrastructure. AI encompasses a wide range of military applications, including automated cyberattacks, intelligence gathering and analysis, autonomous weapon systems, and support for military decision-making. The US government particularly believes that if frontier AI models capable of automatically detecting software vulnerabilities and exploiting them fall into the hands of adversarial forces, a non-symmetric security threat could materialize [20]. This perception has led to measures such as the ban on non-US persons accessing top-tier models like Anthropic's Fable 5 and Mythos 5 [7].

However, delving deeper into this root cause reveals that the US export control policy cannot be explained solely by security logic. An analysis scheduled for publication in a Cambridge University journal suggests that some key players in the US AI industry have intentionally highlighted the technological conflict with China to elicit a regulatory environment and government support favorable to themselves [1]. In other words, the strengthening of export controls stems not only from pure security imperatives but also from the industrial-political motives of US AI companies seeking to institutionally solidify their competitive advantage within a 'winner-takes-all' dynamic [1]. In this regard, export controls function as both security policy and industrial policy, operating in a mutually reinforcing manner.

Another root cause stems from the inherent limitations of control due to the intrinsic nature of AI technology. Unlike hardware with a physical form, such as semiconductors, AI models are digital information that can be easily replicated and disseminated, and their boundaries can be circumvented through the open-source ecosystem. Indeed, immediately after the US blocked access to Anthropic's models, Chinese startup Z.ai released a model with comparable performance without any access restrictions [9], and Chinese cybersecurity firm 360 unveiled 'Tulongfeng,' an AI tool claimed to rival Mythos [12]. This demonstrates that US control measures are not substantially curbing China's indigenous development capabilities and are instead paradoxically creating an opening for Chinese companies to freely penetrate the global market.

2. Structural Context

Political Structure

At the political level, the US's AI export controls reveal a structural vulnerability where ad hoc measures led by the executive branch accumulate without a medium- to long-term strategic framework. The Trump administration, through executive orders, required AI developers to provide early government access before releasing frontier models and pursued the establishment of voluntary standards for model release criteria and the scope of foreign access [2]. However, this process, relying on executive orders without a legislative basis, makes policy consistency and predictability fragile. Furthermore, the dual relationship where the US government treats private companies like Anthropic and OpenAI as both regulated entities and partners in national security further complicates policy design [3].

In contrast, China is pursuing a long-term and systematic state-led AI strategy. Through State Council Order No. 837, effective July 1, 2025, China has established a system where the state strictly controls the scope of technologies, dispatched personnel, and shareable know-how that Chinese companies can transfer during overseas investments [6]. This indicates that China is strengthening its institutional safeguards for preventing technology leakage and protecting strategic technologies, moving in a direction that counters US export controls while simultaneously enhancing its own technological sovereignty. Moreover, China is accelerating the standardization and interoperability of its AI ecosystem by announcing seven national standards for AI agent interoperability [15]. Preempting technical standards is not merely a technical issue but also a political act that determines the future governance structure of the global AI infrastructure.

Economic Structure

At the economic level, US export controls entail a structural dilemma that self-limits the global competitiveness of its own AI industry. While the US blocks foreign access to top-tier AI models, Chinese open-source models are penetrating the global market without restrictions. According to an analysis by Swiss UBS Securities, there is a clear trend of global companies switching from high-performance closed-source models to cost-effective Chinese open-source models, such as Alibaba's Qwen series, for tasks that do not require them [10]. This implies that US control measures are paradoxically facilitating the expansion of Chinese AI models' global market share.

China's scale of AI investment is also a crucial variable for understanding the economic structure. China is pursuing an ambitious plan to invest approximately 2 trillion yuan (about $295 billion) in AI data center infrastructure over the next five years [17], which is not merely corporate investment but strategic infrastructure development at the national level. Such large-scale investment aligns with China's long-term strategy to achieve self-sufficiency in AI computing infrastructure and gradually reduce its dependence on US semiconductor export controls. Indeed, the fact that smuggling activities are spreading throughout the supply chain due to the explosive increase in demand for Nvidia AI servers from China [13] demonstrates that export controls are not suppressing China's desire for technological access but are instead creating side effects that invigorate the underground economy.

Security Structure

At the security level, the core of this issue lies in the military application potential of AI and the asymmetric response structure to it. The AI innovation activities of China's People's Liberation Army Strategic Support Force are a focus of intense analysis within the US security community, and the integration of AI into core elements of military operations, such as cyber warfare, electronic warfare, information warfare, and autonomous weapon systems, is rapidly increasing. China's AI-based cybersecurity capabilities are also rapidly improving, leading to assessments that they are eroding US cyber superiority [11]. In this security environment, US export controls serve as a means to block short-term threats, but they possess a structural limitation where their effectiveness is bound to rapidly diminish over time, considering the speed at which China is developing its indigenous capabilities.

The competition for 6G wireless infrastructure also forms an important layer of the AI security structure. Paul Rosenzweig, former Deputy Assistant Secretary for Policy at the US Department of Homeland Security, warned that if the US falls behind China in 6G spectrum allocation, its mobile AI competitiveness could be severely weakened [14]. This indicates that the competition for AI hegemony is expanding beyond model development and semiconductors to include wireless infrastructure. China is establishing early superiority in the 6G field through long-term planning and systematic investment, suggesting that the convergence of AI and communication infrastructure will be a key foundation of future military power [14].

3. Comparative Analysis of Historical Precedents and Similar Cases

Lessons from COCOM during the Cold War

The US AI export control regime shares structural similarities with the COCOM (Coordinating Committee for Multilateral Export Controls) system, which controlled technology exports to the Soviet Union during the Cold War. Established in 1949, COCOM was a multilateral system where Western allies jointly controlled the export of strategic technologies to the Soviet bloc. Two key lessons emerge from COCOM. First, the effectiveness of export controls is fundamentally limited without the joint participation of allies. Even if the US prohibits the export of a specific technology, its impact is halved if allies export the same technology. In the current context of AI export controls, the fact that allied companies are developing similar capabilities outside the scope of US control, as exemplified by Japan's Sakana AI releasing a model comparable to Mythos, Fugu [12], reaffirms this lesson. Second, COCOM did not completely suppress the Soviet Union's indigenous technological development; instead, it spurred the Soviet Union to strengthen its self-reliance in certain technological fields. This finding aligns precisely with analyses suggesting that current US AI control measures are paradoxically accelerating China's indigenous development.

Precedent of Semiconductor Export Controls: The Case of the Japanese Semiconductor Industry

The US pressure on the Japanese semiconductor industry in the 1980s and the Toshiba incident in 1987 offer another relevant precedent. The Toshiba incident, involving the illegal export of precision machine tools capable of military diversion to the Soviet Union, exposed the vulnerability of technology leakage channels through allied companies, leading to the strengthening of export control systems. However, this case also demonstrated that export controls can create tensions in alliances and incentivize the target country to accelerate its indigenous development. The current situation where India, facing access restrictions to Anthropic's models due to US export control measures, is re-recognizing the need to strengthen its own AI sovereignty [7] suggests that this historical pattern is repeating.

Comparison with the Nuclear Non-Proliferation Regime

The Nuclear Non-Proliferation Treaty (NPT) and the Nuclear Suppliers Group (NSG) represent another historical precedent for controlling the proliferation of dual-use technologies. The nuclear non-proliferation regime shares a similar objective with AI export controls in that it established international agreements and verification systems to prevent the military diversion of technology. However, fundamental differences exist between nuclear technology and AI technology. Nuclear technology requires highly specialized physical infrastructure and rare materials, whereas AI technology, being digital information, is inherently easy to replicate and disseminate. This difference implies that applying nuclear non-proliferation control methods to AI will inevitably have structural limitations. The Carnegie Endowment's proposal for a coalition of free-world AI computing can be interpreted as an attempt to address these limitations and enhance control effectiveness through a multilateral approach.

Precedent of Internet Governance Fragmentation

The US-China conflict surrounding internet governance since the 2010s holds significant implications as a precedent for AI standards competition. Just as China has advocated for 'cyber sovereignty' and moved towards separating its domestic internet ecosystem from global standards, it is now moving towards establishing its own national standards and an interoperability framework for AI agents [15]. This increases the likelihood of the AI ecosystem fragmenting into a US-led global standards system and a China-led independent standards system, similar to the internet. In such a scenario, the effectiveness of export controls would be further weakened.

4. Key Variables in Issue Development

Variable 1: Speed of Open-Source AI Model Proliferation

One of the most decisive variables in the future development of this issue will be the speed at which open-source AI models improve in performance. ChatGPT's global market share as an AI assistant has fallen below 50% for the first time [4], and the large model market is rapidly shifting from a monopoly by a single giant corporation to a multi-player competitive landscape. The faster Chinese open-source models achieve performance levels close to US closed-source frontier models, the more meaningless US export controls will become. In particular, if the trend of global companies switching to Chinese open-source models for cost savings accelerates [10], US control measures could end up merely weakening the global market position of US AI companies.

Variable 2: Policy Coherence among Allied Nations

Policy coherence among key allies is essential for the effectiveness of US export controls. However, Western countries such as Canada and France have declared their intention to build independent AI capabilities [4], and India is seeking to strengthen its AI sovereignty due to access restrictions imposed by US control measures [7]. The emergence of calls for a non-aligned movement in an AI Cold War [8] suggests that global AI governance may not converge into a single US-led system. Whether the coalition of free-world AI computing proposed by the Carnegie Endowment can be concretely realized, and how allied nations' interests will be coordinated in that process, will be key variables.

Variable 3: Speed of China's Semiconductor Self-Sufficiency

The ultimate effectiveness of US export controls hinges on when China achieves the capability to domestically produce advanced semiconductors. Currently, China is resorting to smuggling to acquire Nvidia AI servers [13], demonstrating its continued high dependence. However, if China continues its efforts toward semiconductor self-sufficiency alongside a 2 trillion yuan investment in AI data centers [17], US hardware export controls will lose their efficacy over time. This variable is the key factor determining the 'shelf life' of export controls.

Variable 4: Speed and Transparency of AI Militarization

The speed of the Chinese People's Liberation Army's AI militarization and its level of transparency are also important variables in the issue's development. As the military application of AI accelerates and its reality becomes clearer, pressure for stronger export controls from the US and its allies will increase. Conversely, if the reality of AI militarization proves to be unclear or exaggerated, the justification for export controls will weaken, and industry backlash may grow. This variable determines the political sustainability of export control policies.

Variable 5: Outcome of Global AI Standards Competition

Finally, the outcome of the competition surrounding AI technology standards is a crucial variable. If China successfully establishes its own national standards for AI agents [15] and promotes them internationally, the effectiveness of US export controls will be further weakened by the fragmentation of the standards system itself. Conversely, if the US, along with its allies, preempts the establishment of AI technology standards and governance systems, more effective technological hegemony can be maintained through the synergy of export controls and standard setting. The outcome of this standards competition will be the most critical structural variable determining the landscape of the global AI ecosystem over the next decade.