?How China Is Changing the Rules of AI Competition
From the Superiority of American Models to the Geopolitical Power of China’s Open Models
Mohammad Ali Falahi, MA in International Relations
Introduction
AI competition between China and the United States is usually portrayed as a linear race in which the United States stands at the technological frontier and China seeks to close the gap. This picture obscures the strategic transformation of 2026. The issue is not merely which country builds the most powerful model, but who defines the metrics of AI power. China is seeking to shift the competition from absolute superiority in frontier models toward diffusion, localization, cost reduction, industrial integration, and the creation of infrastructural dependencies. Within this framework, an open model is not merely an engineering choice; it is an instrument of industrial policy, technology diplomacy, and the production of influence in the international system.
The Stanford AI Index 2026 shows that the performance gap between the best American and Chinese models has narrowed sharply. Since early 2025, models from the two countries have exchanged the top position several times, and in March 2026 the best American model’s lead was reported at only 2.7 percent. Nevertheless, the United States still leads in the number of leading models, private investment, data centers, and influential inventions. U.S. private AI investment reached $285.9 billion in 2025, compared with $12.4 billion in China. The convergence in model performance therefore does not mean structural parity between the two ecosystems (Stanford Institute for Human-Centered Artificial Intelligence, 2026).
Changing the Unit of Measurement of Power
The American advantage rests on capital, advanced chips, computing capacity, and companies that push the scientific frontier of general-purpose models. Yet this model has a geopolitical limitation. The more models are closed, expensive, and dependent on proprietary cloud services, the less independently governments, universities, and small firms can use them. Superiority at the top of the technological hierarchy does not necessarily translate into dominance of the global market. China has intervened in the gap between technical superiority and global diffusion. Beijing regards AI as a source of national power when it leaves the laboratory and enters factories, power grids, hospitals, transportation systems, local governments, and supply chains. Brookings describes competition between the two countries as a set of simultaneous contests in computing, models, adoption, integration, and deployment. The United States leads at the technological frontier, but China is creating a different advantage through greater efficiency, the diffusion of open models, and the integration of AI with the real economy (Chan, 2026).
Consequently, the decisive metric is not simply the smartest model, but the model that becomes public economic infrastructure. For many developing countries, an inexpensive, locally hostable, and customizable model is more attractive than a higher-performing model that requires continuous payment and the transfer of data to foreign servers. Competition is gradually shifting from the ability to produce intelligence to the ability to distribute, adapt, and institutionalize it.
The Open Model as Industrial Policy
Many Chinese models are in fact released with “open weights”: trained parameters are published, while the training data and all details of the training process are not made available. Nevertheless, publishing the weights allows developers to run the model on their own infrastructure and fine-tune it for a specific language or industry. Technology is thereby transformed from an external service into an asset that can be appropriated and localized, and this gives it geopolitical value. The 2026 Work Report of the Chinese government prioritized support for open-source AI communities, the construction of an open ecosystem, expansion of the “AI Pulse” initiative, development of computing clusters, and applications of intelligent agents. The open model has moved beyond the strategy of individual companies and become part of a national development program. The goal is to create a cycle in which inexpensive models increase adoption, widespread use generates experience, and that experience feeds back into improvements in domestic software, chips, and industrial solutions (State Council of the People’s Republic of China, 2026).
This cycle is also a response to U.S. export controls. Washington has sought to slow the development of China’s frontier models by restricting its access to advanced chips. However, in January 2026, case-by-case export licenses for chips such as NVIDIA H200 and AMD MI325X were accepted under security conditions. This change reveals an internal tension in U.S. policy. Strict bans may push China toward faster development of an independent hardware and software supply chain, while controlled access can also increase the competitor’s capacity. The United States simultaneously seeks to preserve its computing lead and keep the global ecosystem dependent on American technology (Bureau of Industry and Security, 2026).
Export controls raise the cost of AI development in China in the short term, but in the long term they may create stronger incentives for the integration of domestic chips, local software frameworks, and low-compute models. This situation shows that technology-containment policy does not necessarily remove a competitor from the field; sometimes it changes the competitor’s innovation model. Instead of replicating the costly path of American companies, China has focused on computing optimization, reducing inference costs, and adapting models to more constrained hardware.
Token Economics and the Architecture of Dependency
China’s strategy can be described as the political economy of cheap tokens. The primary objective is not necessarily to maximize profit from each user, but to increase the volume of use, make the model the default choice for developers, and capture the application layers. Mercator reports that since late 2025 Chinese models have surpassed American models in Hugging Face downloads, and seven of the ten highest-performing open-weight models have been Chinese. Low prices, computational efficiency, and the ability to host models independently give these models an advantage in cost-sensitive environments (Chang, 2026).
A low-cost model is a gateway into an ecosystem. Development tools, cloud services, compatible chips, and training emerge around it. A country that adapts a Chinese model for digital government or industry may become locked into that ecosystem’s technical architecture. Dependency is created not only through data storage; standards and update paths can also constrain future choices. This pattern is connected to the concept of infrastructural power. Infrastructural power does not necessarily arise from direct control over users, but from the ability to determine the environment within which other actors make decisions. When companies, universities, and governments base their software, databases, and administrative processes on Chinese models, the cost of switching to another ecosystem rises. An open model appears to allow exit, but the network of tools and skills surrounding it can ensure the persistence of dependency.
China’s AI Diplomacy
In 2026, Beijing explicitly incorporated open models into the language of foreign policy. At the World Artificial Intelligence Conference in Shanghai, Xi Jinping emphasized openness, cooperation, and sharing, and described building the capacity of Global South countries as necessary to prevent a new AI divide. He announced training for 5,000 people from developing countries, the creation of application cooperation centers with regional organizations, and the provision of AI-based weather-warning systems to 30 countries (Xi, 2026).
These measures turn AI into an instrument of capacity-building diplomacy. China seeks to present itself not only as a technology vendor but also as a provider of digital public goods. This narrative is attractive to countries concerned about the concentration of advanced models in a few American companies, the high cost of cloud services, and the possibility that access could be restricted during crises. Open models allow Beijing to connect technological equity, digital sovereignty, and the right to development with its industrial interests. Geopolitics is constructed here through the way technology is disseminated. This narrative is not without contradictions. China’s domestic ecosystem is subject to content oversight and data-security requirements, and open model weights do not guarantee transparency regarding training data or safety mechanisms. Recipient countries must distinguish between genuine localization and hidden dependency. Adopting Chinese models may increase independence from American companies, but it does not necessarily lead to full technological autonomy.
America’s Difficult Response
The American challenge is not the immediate loss of technical superiority. The problem is that the business model of American companies does not necessarily align with the geopolitical objective of broad diffusion. Frontier laboratories tend to favor closed models, cloud services, and proprietary control in order to generate returns on investment. China can treat part of the cost of diffusion as a strategic investment in industrial influence and standard-setting. Washington’s effective response therefore cannot be limited to tighter export controls. The United States needs competitive open-weight models, accessible computing infrastructure for allies, and development packages that keep American technology attractive in terms of cost, language, and data sovereignty. Otherwise, the United States may possess the best models while surrendering a large share of the global market for AI applications to the Chinese ecosystem.
Conclusion
China’s main innovation is not the creation of a particular model, but a change in the unit by which AI power is measured. If power is measured solely by the score of the best model, the United States still has a significant advantage. But if power means presence in industrial processes, control over development tools, the shaping of standards, and becoming the cognitive infrastructure of other countries, the outcome of the competition remains uncertain. China is moving the arena of competition from the laboratory to the ecosystem and from ownership of technology to the architecture of dependency. The future order may be two-layered: leading American models may perform complex, high-value tasks, while Chinese models capture a large share of routine processing, local applications, and low-cost services. In such an order, the United States may produce the world’s most advanced brains, while China builds the common language of the world’s machines. The strategic danger for Washington is not falling behind on a single technical metric, but seeing its scientific superiority become an expensive island in an ocean of applications based on Chinese models.
The decisive question is no longer who reaches artificial general intelligence first. The question is who will define the rules governing access, modifiability, cost of use, and the right to host AI. Future power will belong to the country that turns others not merely into consumers of its models, but into developers within its ecosystem. China is building precisely this possibility through open models, and for that reason AI competition is becoming a struggle over the organization of global technological dependency.
References
Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index Report. Stanford University.
Bureau of Industry and Security. (2026, January 13). Department of Commerce Revises License Review Policy for Semiconductors Exported to China. U.S. Department of Commerce.
State Council of the People’s Republic of China. (2026, March 13). Report on the Work of the Government.
Chan, K. (2026, April 16). Competing AI Strategies for the US and China. Brookings Institution.
Chang, W. (2026, June 3). China’s AI Competition Strategy: Wide Dispersion, Cheap Tokens. Mercator Institute for China Studies.
Xi, J. (2026, July 17). Joining Hands to Build a Just and Equitable System for Global AI Governance. Xinhua News Agency.