Artificial intelligence has become an important dimension of relations between China and the United States, affecting industrial policy, international commerce, scientific exchange, military planning, and diplomatic negotiation. Its significance extends beyond competition to develop better software. Advanced AI depends on interconnected systems of computing hardware, research institutions, skilled workers, data, investment, and electricity. Government decisions affecting any of these resources can influence both technological development and the broader bilateral relationship.
The two countries occupy central but different positions in this technological system. American firms and universities have played leading roles in advanced chip design, cloud computing, and frontier AI development. Chinese companies and research institutions have developed substantial capabilities in AI research, industrial applications, digital services, and increasingly sophisticated models. These activities have grown through international exchange even as both governments have introduced restrictions justified by national security and economic considerations.
The relationship is therefore neither a simple technological race nor a complete separation into independent systems. Competition coexists with commercial dependence, shared scientific foundations, and limited cooperation over risk. The developments examined here, principally through the end of 2024, show how AI has become embedded in existing disputes over trade, security, and political authority while creating new questions about access to computing resources, the international distribution of technological benefits, and control over increasingly capable systems.
From scientific exchange to strategic competition
AI entered an already complex bilateral relationship. Following the normalization of diplomatic relations in 1979, scientific and educational exchanges expanded substantially. Chinese students and researchers entered American graduate programs, multinational firms established research operations in China, and cross-border publication became a routine feature of computer science. Advances in machine learning during the 2000s and 2010s developed within this international research environment. Shared datasets, conference publications, and software libraries allowed techniques to circulate more freely than many physical technologies.
The rise of deep learning increased the strategic visibility of AI. Improvements in image recognition, speech processing, and machine translation demonstrated that artificial neural networks could support commercially valuable applications. In 2016, Google DeepMind’s AlphaGo defeated Lee Sedol in Go, and in 2017 it defeated Ke Jie. These events attracted considerable attention in China, although Chinese AI research and investment long predated them. Their importance lay partly in making rapid progress legible to policymakers, investors, and the wider public.
In July 2017, China’s State Council issued the New Generation Artificial Intelligence Development Plan. It set staged objectives for 2020, 2025, and 2030, including the ambition to become a leading global AI innovation center by 2030. The plan connected basic research, industrial development, education, and security applications. In the United States, the February 2019 executive order launching the American AI Initiative emphasized federal research, access to resources, technical standards, and workforce development. The National Artificial Intelligence Initiative Act of 2020, enacted in January 2021, established a statutory framework for federal coordination.
These initiatives overlapped with a broader deterioration in bilateral relations. Trade disputes, concerns about technology transfer, and restrictions involving telecommunications companies helped recast AI as a strategic capability rather than merely an expanding commercial sector. American officials increasingly linked access to advanced technology with the modernization of China’s military. Chinese policy documents emphasized indigenous innovation and reduced exposure to foreign technological restrictions.
The public release of ChatGPT in November 2022 added another dimension. General-purpose systems based on large language models suggested that AI could affect a much wider range of professional and creative activities than earlier specialized applications. American and Chinese firms accelerated model development, while governments focused more closely on computing capacity and model governance. This chronology does not establish that technological rivalry was inevitable: it shows how scientific advances interacted with existing political disputes and policy choices to make AI a prominent component of strategic competition.
Capabilities, resources, and the problem of measuring leadership
AI capability is distributed across several layers rather than captured by a single national ranking. Relevant resources include researchers, algorithms, computing infrastructure, specialized chips, high-quality training data, financing, and the ability to turn experimental systems into dependable products. A country can perform strongly in academic publishing while depending on imported manufacturing equipment. A company can produce a competitive model without controlling the hardware supply chain on which it runs. These distinctions complicate claims that either country has achieved comprehensive AI leadership.
The United States benefits from a concentration of major technology firms, venture investment, internationally connected universities, and advanced chip-design companies. Firms including Google, Microsoft, Amazon, Meta, and OpenAI occupied important positions in AI development and deployment by 2024. American universities such as Stanford University, the Massachusetts Institute of Technology, and Carnegie Mellon University have longstanding research strengths. These capabilities nevertheless depend heavily on internationally mobile researchers and globally distributed production networks; they are not exclusively domestic resources.
China combines a large research community with extensive digital markets and manufacturing capacity. Baidu, Alibaba, Tencent, Huawei, and specialized AI firms have supported research and commercial deployment across language technology, computer vision, logistics, and industrial systems. Universities and public research institutions contribute both fundamental work and trained personnel. Stanford’s 2024 AI Index reported that, under its classification, U.S.-based institutions produced 61 notable machine-learning models in 2023, compared with 15 from China. The same report identified China as a major leader in AI publication volume. These measures describe different activities and should not be treated as interchangeable indicators of technological superiority.
Computing capacity is particularly consequential for large-scale models. A simplified approximation for the training computation of some dense transformer models is , where is the number of floating-point operations, the number of model parameters, and the number of training tokens. This is an engineering approximation, not a universal law or a direct measure of intelligence. Architecture, data quality, numerical precision, and training procedures can substantially change the resources required to reach a given level of performance.
Commercial value depends on additional factors. Running a model for users—usually called inference—creates continuing costs that differ from the one-time cost of a training run. Reliable deployment requires integration with databases, security controls, monitoring, and human workflows. Benchmark scores can also overstate practical usefulness if test material overlaps with training data or if evaluations omit important languages and operating conditions. Consequently, bilateral comparisons require separate assessments of research quality, frontier performance, deployment scale, cost efficiency, and dependence on external suppliers.
Semiconductors, export controls, and technological interdependence
Advanced semiconductors connect AI competition to a highly international production system. American companies design many leading AI accelerators and supply important chip-design tools. Taiwan is central to leading-edge contract manufacturing, South Korea to advanced memory, the Netherlands to essential lithography equipment, and Japan to several categories of manufacturing equipment and materials. China is a major electronics manufacturing center, semiconductor market, and growing producer of chips and equipment. None of these roles can be understood adequately through a purely bilateral map.
The importance of graphics processing units arises from their ability to perform large numbers of mathematical operations in parallel. Training modern neural networks involves repeated operations on arrays of numbers, making specialized accelerators particularly valuable. Effective AI computing also requires high-bandwidth memory, fast connections among processors, advanced packaging, networking equipment, and software. Restricting access to a particular accelerator therefore affects one component of a larger system, while restrictions on equipment can influence the ability to manufacture future generations.
On October 7, 2022, the U.S. Department of Commerce’s Bureau of Industry and Security announced controls covering certain advanced computing chips, semiconductor manufacturing items, and specified activities by U.S. persons supporting advanced semiconductor production in China. The measures included licensing requirements and end-use restrictions. In October 2023, the department revised the rules, including technical criteria intended to address products and transactions that could fall outside the initial restrictions. The stated American rationale centered on limiting capabilities relevant to military modernization and other national-security concerns.
These export controls differ from tariffs. A tariff changes the cost of an otherwise permitted transaction; an export-control regime may require authorization or prohibit a transfer regardless of the price offered. Its operation depends on product specifications, destination, end user, end use, and sometimes the involvement of American technology in foreign-produced goods. Commercial cloud services create a further complication: organizations may obtain computing services without purchasing the underlying processors, making the location and identity of users relevant to enforcement debates.
China has pursued greater domestic semiconductor capability through public support, corporate investment, and research programs. Restrictions can encourage substitution and more efficient use of available hardware, but they can also constrain access to equipment and components that are difficult to replace. In 2023, China introduced export licensing requirements for certain gallium- and germanium-related items and for specified graphite products, illustrating that supply-chain dependencies extend beyond finished chips. Such measures have distinct legal scopes and cannot automatically be treated as exact equivalents.
The resulting relationship is one of selective restriction within continuing interdependence. Export controls can delay access, raise costs, and redirect investment, but their effects depend on enforcement, supplier coordination, inventories, technical adaptation, and the availability of substitutes. They do not provide a simple mechanism for halting all AI development in a large technological economy.
Commercial competition, investment, and industrial adoption
Commercial AI competition includes both companies building general-purpose models and businesses applying AI to particular tasks. Model developers require substantial computing resources and specialized personnel, while application developers may purchase access through an application programming interface rather than train their own systems. This distinction allows firms with relatively limited resources to enter AI markets, but it can also make them dependent on a small number of infrastructure providers. Cloud access, software compatibility, and pricing therefore influence the distribution of commercial power.
American firms have held prominent positions in cloud services, model platforms, accelerator design, and enterprise software. Chinese firms have developed models and applications within large domestic markets, including online commerce, digital payments, manufacturing, and consumer services. Their opportunities differ by sector. A factory using machine vision to identify defective components faces different technical and regulatory requirements from a chatbot serving the general public. The commercial significance of AI cannot consequently be measured only by the popularity of conversational systems.
Openly distributed models complicate the idea of separate national technology blocs. An open-weight model makes trained numerical parameters available under specified conditions, allowing others to run or modify it without relying exclusively on the original developer’s hosted service. American and Chinese developers have both used this approach. However, released weights do not necessarily come with training data, complete development code, or unrestricted legal permissions. “Open” describes a range of arrangements rather than one uniform commercial or technical category.
Investment policy has become another connection between AI and bilateral relations. In August 2023, President Joe Biden issued Executive Order 14105, establishing the basis for restrictions on certain U.S. investments involving specified technologies in countries of concern. The Treasury Department’s October 2024 final rule, scheduled to take effect on January 2, 2025, created notification requirements and prohibitions for defined transactions involving semiconductors and microelectronics, quantum information technologies, and certain AI systems. It was not a prohibition on all investment in Chinese technology companies; its scope depended on covered activities and transaction characteristics.
AI’s broader economic effects also depend on adoption outside the technology industry. Productivity improvements require usable data, compatible software, employee training, organizational redesign, and effective evaluation. A strong research sector does not automatically produce widespread gains in manufacturing, healthcare administration, or small businesses. Differences in industrial structure and labor costs can make the same technical capability economically attractive in one setting but less so in another.
Competition over AI is therefore partly competition over complementary institutions and assets. Intellectual property, access to finance, procurement practices, electricity supply, and competition law influence which firms can enter markets and capture value. Public concern about job displacement belongs within this picture, but exposure of tasks to automation is not equivalent to measured job loss. Actual effects depend on demand, business decisions, and the creation of new tasks as well as the technical capacity to automate existing ones.
Data, domestic regulation, and political institutions
AI systems operate within legal and political environments that shape what data they may use and what services they may provide. China and the United States differ in governmental structure, approaches to speech, and mechanisms of regulatory oversight. Yet neither country’s AI governance can be reduced to a single principle. Both employ combinations of privacy rules, consumer protection, security requirements, sector-specific regulation, and administrative supervision. The practical effects vary by application, institution, and type of information.
In China, the Cybersecurity Law took effect in 2017, while the Data Security Law and Personal Information Protection Law took effect in 2021. Together with implementing measures, these laws establish obligations concerning network security, data handling, personal information, and some cross-border transfers. The Personal Information Protection Law includes rules on automated decision-making, including requirements concerning transparency and fairness. These provisions affect AI developers because collecting data, training models, and delivering services may involve different legal purposes and responsibilities.
China also introduced measures specifically addressing algorithmic recommendation services, deep-synthesis services, and generative AI. The Interim Measures for the Management of Generative Artificial Intelligence Services took effect on August 15, 2023. They principally concern generative AI services offered to the public within China and address matters including lawful data sources, personal information, and content management. Their scope should not be confused with a uniform approval requirement for every internal research experiment or every business use of machine learning.
The United States has historically relied on a more fragmented framework, including sectoral privacy laws, state legislation, civil-rights protections, and enforcement against unfair or deceptive practices. The Federal Trade Commission has applied existing legal authorities to AI-related conduct. State measures, including the California Consumer Privacy Act, add obligations concerning personal information. In January 2023, the National Institute of Standards and Technology released its voluntary AI Risk Management Framework. President Biden’s October 2023 Executive Order 14110 directed federal agencies to undertake additional work on AI safety, security, and other policy issues.
These arrangements influence bilateral commerce through compliance costs and restrictions on data flows. A company providing the same service in both countries may need different hosting, documentation, data-processing arrangements, and output controls. Data privacy and cybersecurity requirements can overlap without serving identical purposes: protecting a person’s information is different from restricting a foreign entity’s access to a sensitive dataset. Likewise, security review, commercial confidentiality, and content regulation are distinct legal mechanisms even when they affect the same service.
Training-data disputes introduce further complexity. Copyright, contractual restrictions, and rules governing personal data can apply differently to collection, model training, and generated outputs. Technical access to material does not by itself establish legal permission to use it. These unresolved questions affect both domestic markets and international cooperation, particularly when research teams or companies seek to combine information originating under different legal systems.
Military applications and strategic stability
AI has military relevance because it can assist with sensing, analysis, logistics, communications, and decision support. Potential uses include recognizing objects in imagery, maintaining equipment, coordinating uncrewed platforms, and processing intelligence. Many underlying technologies are dual-use technologies: the same advances in computer vision or navigation may serve industrial and military purposes. This overlap helps explain why governments scrutinize commercial chips, software, research relationships, and investment rather than treating military AI as a separate, easily bounded sector.
American defense initiatives have included Project Maven, announced in 2017 to apply machine learning to the analysis of imagery and video, and the establishment of the Chief Digital and Artificial Intelligence Office in 2022. Chinese official military publications and defense documents have discussed the development of “intelligentized” warfare and the growing role of AI-related capabilities. Public statements and organizational initiatives, however, do not establish the operational performance of particular systems. Military effectiveness depends on integration, training, communications, maintenance, and the ability to function under adversarial conditions.
The distinction between prediction and action is especially important. An algorithm that identifies a possible military object does not independently determine whether attacking it would comply with international humanitarian law. Distinction, proportionality, and precautions in attack require context that may be absent from a statistical classification. Systems can also fail because operating conditions differ from training conditions, sensors are degraded, or an adversary deliberately manipulates inputs. High accuracy on a curated benchmark does not establish battlefield reliability.
AI can affect strategic stability without producing fully autonomous weapons. Faster information processing may improve warning and situational awareness, but it may also compress the time available for judgment. Decision-makers can become excessively reliant on apparently precise outputs, while adversaries may misinterpret automated actions. A security dilemma arises when measures adopted by one state to improve its security are perceived by another as threatening, encouraging reciprocal measures. Secrecy surrounding military AI makes it difficult to distinguish tested capabilities from aspirations and to evaluate intentions.
The relationship with nuclear weapons is particularly sensitive. AI-assisted early warning, intelligence analysis, cyber operations, or tracking systems could influence assessments of vulnerability even without direct automation of launch decisions. A model-generated assessment may carry serious consequences if uncertainty is hidden or poorly communicated. Maintaining human authority does not by itself resolve every risk; the quality of information, organizational procedures, and time available for deliberation remain important.
At their November 2024 meeting in Lima, Biden and Xi Jinping affirmed the need to maintain human control over decisions to use nuclear weapons. This was a specific diplomatic statement, not a comprehensive agreement governing military AI. Broader negotiations face difficulties familiar from arms control: defining covered systems, verifying compliance, managing classified information, and distinguishing civilian infrastructure from military capability.
Research networks, talent mobility, and openness
AI research remains substantially international even when national competition becomes more pronounced. Researchers share techniques through conferences, journals, repositories, and software libraries. Peer review and public testing can expose weaknesses that individual laboratories miss, while preprints allow rapid dissemination before formal publication. Because many AI results are implemented in software, knowledge can cross borders more quickly than physical equipment. This circulation makes a clean division into nationally self-contained research systems difficult.
Educational mobility has been an important channel connecting China and the United States. Chinese students form a substantial part of the international research population at American universities, while researchers of Chinese origin work across universities, established firms, and start-ups. Nationality, birthplace, institutional affiliation, and citizenship describe different things and should not be treated as interchangeable evidence of political allegiance or technology transfer. A scientist’s publications may contribute simultaneously to an employer, a multinational research community, and applications in several countries.
Security concerns have nevertheless changed the conditions of exchange. In 2018, the U.S. Department of Justice launched the China Initiative, which included work on economic espionage and research-security matters. In February 2022, the department announced that it would replace that initiative with a broader strategy for countering nation-state threats. Separately, Presidential Proclamation 10043, issued in 2020, restricted entry for certain Chinese students and researchers associated with entities supporting China’s military-civil fusion strategy. It was not a general prohibition on Chinese students attending American universities.
The issues grouped under research security are legally distinct. Theft of a trade secret, failure to disclose research support, breach of a contract, and unauthorized export of controlled technology involve different requirements and forms of evidence. Treating them as equivalent obscures both genuine security risks and the responsibilities of universities. Institutions must also consider academic freedom, nondiscrimination, and the value of international collaboration when implementing government requirements.
Openness itself has several dimensions. Publishing a scientific paper does not necessarily reveal training data, model weights, engineering details, or the resources needed to reproduce a result. Conversely, restricting a final model does not erase the shared mathematical and computational foundations on which it rests. Commercial secrecy and high training costs can limit reproducibility even in the absence of government restrictions.
Restrictions on mobility and collaboration may therefore have mixed effects. They can reduce particular channels of access but also redirect researchers, encourage the development of alternative institutions, and weaken professional networks. Their consequences are difficult to isolate because migration decisions also reflect employment opportunities, family circumstances, funding, and domestic conditions. In bilateral relations, scientific exchange functions both as a source of capability and as a channel of communication whose value cannot be measured solely through publication counts.
International institutions, standards, and third countries
China–United States AI relations are shaped by governments and institutions beyond the two countries. Semiconductor supply chains involve multiple jurisdictions, while digital services operate in markets with different regulatory requirements. The European Union, for example, adopted the AI Act in 2024, establishing a risk-based legal framework with provisions taking effect in stages. European regulation affects American and Chinese suppliers seeking access to its market and demonstrates that international AI governance is not simply the export of one of two national models.
Technical standards provide another arena for influence. Standards can define testing procedures, documentation, interoperability, and risk-management practices. Organizations including the International Organization for Standardization, the International Electrotechnical Commission, and the International Telecommunication Union bring together governmental and technical participants in different institutional arrangements. Participation can confer influence, but an adopted standard does not necessarily bind every country or determine how a technology will be used. Its significance depends on incorporation into procurement, regulation, certification, and commercial practice.
Multilateral statements have also established areas of shared vocabulary. In November 2021, UNESCO adopted its Recommendation on the Ethics of Artificial Intelligence, addressing subjects including transparency, fairness, environmental effects, and human oversight. In March 2024, the United Nations General Assembly adopted a U.S.-led consensus resolution on safe, secure, and trustworthy AI systems for sustainable development. In July 2024, it adopted a China-led consensus resolution on enhancing international cooperation in AI capacity-building. Neither resolution created a binding global AI regulatory authority.
These initiatives reflect overlapping but distinct priorities. Safety-focused discussions often emphasize risks from highly capable systems, misuse, and failures in important applications. Capacity-building discussions emphasize access to infrastructure, skills, and development opportunities. These concerns are not mutually exclusive: a country may seek access to advanced tools while also wanting protection from unreliable systems, fraud, or external dependence. The digital divide includes differences in computing access, language coverage, connectivity, and institutional capacity, not merely the presence or absence of a chatbot.
Third countries therefore have interests that cannot be reduced to choosing between China and the United States. They may use American-designed chips, Chinese telecommunications equipment, locally developed applications, and services hosted in several jurisdictions. Their choices can be influenced by financing, energy availability, procurement conditions, language support, and rules concerning sovereignty and information access.
AI also connects bilateral competition with shared international problems. Applications in climate modeling, agricultural monitoring, and scientific discovery may support sustainable development, but benefits depend on validation, local data, and usable institutions. International cooperation becomes especially difficult when a useful civilian capability also has sensitive security applications. Standards and development programs can facilitate limited coordination without eliminating these underlying conflicts.
Diplomatic risk management and unresolved questions
The emergence of advanced general-purpose models created a limited diplomatic agenda concerning AI safety. Safety concerns include unreliable outputs, vulnerabilities, harmful misuse, and the possibility that more capable systems could behave in difficult-to-predict ways. Some risks can be studied through technical testing; others depend on deployment context, institutional incentives, and the conduct of users. The existence of shared concerns does not mean that China and the United States agree on their relative importance or on acceptable regulatory responses.
In November 2023, representatives from both countries participated in the AI Safety Summit at Bletchley Park in the United Kingdom. The resulting Bletchley Declaration recognized opportunities from AI and the need to understand and manage risks associated with advanced systems. Following the November 2023 Biden–Xi meeting near San Francisco, the two governments established an intergovernmental dialogue on AI. Their first meeting under that arrangement took place in Geneva in May 2024. It provided a formal channel for exchanging views, but did not produce a comprehensive bilateral regulatory agreement.
Risk management is complicated by the difference between describing a danger and verifying a commitment. A pledge concerning model testing requires decisions about which systems are covered, what evaluations count, and how results are shared. Training-compute thresholds offer a relatively concrete criterion, but computing expenditure does not map perfectly onto capability. Smaller systems can become more effective through better methods, while a large system may perform unevenly across tasks. Static thresholds can therefore become outdated as technology changes.
Disclosure presents a related problem. Sharing information about model vulnerabilities can improve collective understanding, but it can also expose security weaknesses, commercially valuable details, or sensitive research. Independent testing may require access that developers or governments are unwilling to grant. Agreement on broad principles is thus easier than agreement on inspection, incident reporting, or restrictions that impose unequal costs.
Analogies with the Cold War illuminate some issues but have limits. AI systems are reproducible software artifacts supported by widely used civilian infrastructure, rather than a single category of readily countable weapon. Their capabilities can change through updates, connections to external tools, and new deployment arrangements. At the same time, the history of crisis communication and negotiated restraint shows why adversaries may maintain technical discussions without resolving larger political disputes.
Several empirical questions remained open at the end of 2024: how much export controls would slow particular capabilities; whether efficiency improvements would reduce the importance of restricted hardware; how widely AI would raise productivity; and which evaluation methods would reliably identify serious hazards. Evidence for these questions requires more than national rankings or announced investment totals. Relevant indicators include reproducible system performance, actual hardware access, operational reliability, adoption outside leading technology firms, and the durability of diplomatic communication during periods of political tension.