China may not own the world’s best artificial intelligence model for long stretches of time. It may not need to. Its stronger wager is that the model used by a factory, ministry or start-up will count for more than the model that tops a benchmark. Beijing is therefore backing systems that are capable, inexpensive and available for others to modify. The United States sells access to intelligence. China increasingly distributes the machinery.
The distinction has begun to alter the AI contest. Stanford University’s 2026 AI Index found that the performance gap between the leading American and Chinese models had narrowed to 2.7% by March. American laboratories still produced more top-tier models. China led in AI publications, citations, patents and industrial robot installations. The frontier remains largely American; the ground beneath it is becoming Chinese.
Alibaba’s Qwen, DeepSeek, Moonshot AI’s Kimi and models developed by Z.ai are now among the principal building blocks of the open-weight market. Downloads of Chinese models overtook those of American models in 2025. Qwen and DeepSeek account for much of the increase, but the breadth of the Chinese field is more consequential than the success of any one laboratory. Baidu, once committed to closed models, has also released weights from some of its systems.
This is not technological charity. An open model attracts developers, derivative models, cloud customers and applications. Each adoption strengthens the tools, documentation and technical habits surrounding it. The provider may not collect a fee each time the model is downloaded, but Chinese cloud platforms, chips and engineering services can follow. Android showed how software given away at one layer could secure commercial power elsewhere. Beijing has studied such precedents.
READ I Women’s jobs face greater disruption from artificial intelligence
China’s open-weight AI advantage
American companies entered the generative AI boom with the strongest models, the most advanced chips and access to enormous capital. They chose a business model consistent with those advantages. OpenAI, Anthropic and Google sell controlled access through subscriptions and application programming interfaces. The customer receives a service, while the provider retains the model weights and can change prices, conditions or availability.
That arrangement suits companies that can pay dollar-denominated charges and accept dependence on an American supplier. Governments are less comfortable. A ministry handling tax records or medical data may want the model on its own servers. A business in Indonesia, Brazil or Kenya may prefer a system that can be adapted to local languages without sending every query abroad. Chinese open-weight models address these objections at a price that few domestic alternatives can match.
US export controls helped produce this outcome. Denied unrestricted access to Nvidia’s most advanced chips, Chinese laboratories had to use computing power more frugally. DeepSeek’s success in early 2025 demonstrated that efficiency in training and inference could narrow part of the hardware disadvantage. Chinese developers subsequently competed through sparse architectures, smaller specialised models and lower token prices.
Washington succeeded in raising China’s cost of obtaining frontier chips. It also gave Chinese firms an incentive to build models that run on cheaper hardware. That quality travels well.
The Chinese state has supplied much of the supporting structure. Local governments have subsidised computing centres. State-owned enterprises purchase domestic technology. Universities and public laboratories provide researchers, while venture funds absorb risks that private capital might reject. RAND describes this as a “full stack” industrial policy covering research, talent, compute and applications.
The results include waste. Some local governments built computing centres unsuited to current workloads; reports in 2025 suggested high idle capacity at several facilities. Subsidies have financed duplication and weak firms. China’s advantage does not arise from flawless planning. It comes from the willingness to tolerate costly experiments until a few produce scale.
READ I Algorithmic transparency key to digital democracy
AI Plus takes the contest into factories
China’s policy differs most sharply from America’s in the place assigned to industry. The State Council’s “AI Plus” programme directs the technology towards scientific research, manufacturing, consumer products, public services and government. The 15th Five-Year Plan for 2026–30 adds AI agents, multimodal systems, embodied AI and swarm intelligence to that programme. Beijing expects AI-related industries to exceed 10 trillion yuan by 2030.
The number should be treated as an official ambition, not a forecast. The industrial connection is nevertheless substantial. China already has the world’s largest manufacturing base and installs more industrial robots than any other country. Its companies produce drones, electric vehicles, batteries, machine tools and telecommunications equipment. AI can be trained and tested inside these systems, generating data from machines operating outside a laboratory.
A March 2026 report by the US-China Economic and Security Review Commission described two reinforcing loops. Open models encourage wider industrial use. Factories and robots then generate specialised data that improve the models used in production. The loop is difficult to reproduce in an economy that has surrendered much of its manufacturing depth.
American frontier laboratories are trying to build increasingly general and autonomous systems. China is attempting to place adequate intelligence inside millions of machines. The first course may produce the larger scientific breakthrough. The second could generate earlier gains in productivity and exports.
Energy also favours China in a less obvious way. Training and running AI systems require dependable electricity, transmission capacity and cooling infrastructure. China can build power plants, grids and data centres faster than most democracies. Its dominance in solar modules, batteries and power equipment gives it control over much of the physical supply chain behind new computing capacity. Restrictions on advanced chips still hurt. A shortage of electricity is less likely to do so.
READ I Women’s jobs face greater disruption from artificial intelligence
China turns AI access into diplomacy
Beijing is carrying its domestic model abroad. At the 2026 World Artificial Intelligence Conference, Xi Jinping offered 5,000 AI training and seminar places to developing countries over five years. China is also proposing cooperation centres with BRICS, ASEAN, the African Union and Latin American countries. Its Global AI Governance Action Plan calls for greater access to infrastructure, open-source technology and training.
The offer is well matched to the complaints of developing countries. Most have supplied data and users to the digital economy but own little compute, few large models and no global platforms. Western arguments about AI safety can sound like an attempt to preserve a technological hierarchy. China offers usable models and engineers.
There is an echo of the Belt and Road Initiative, though AI travels more easily than a railway or power station. A government can download a model before signing a financing agreement. Chinese companies can provide cloud capacity, fine-tuning, cybersecurity tools and training afterwards. Dependence develops through compatibility and accumulated technical choices rather than sovereign debt.
Beijing also understands the value of attendance. Chinese officials and technology executives have become prominent at UN discussions on standards and AI governance. China has previously influenced international work on facial recognition, surveillance technology and automated vehicles. When senior American officials and frontier laboratories stay away, China does not have to defeat an opposing proposal. It has to provide the draft.
Standards are rarely neutral once products are built around them. Technical definitions decide which systems can be certified, which data must be retained and whose equipment can connect. A country that supplies both the model and the standard gains an advantage that does not appear on benchmark tables.
Chinese AI carries political conditions
Concerns about Chinese models cannot be dismissed as American protectionism. China’s 2023 interim rules require public generative AI services to uphold “core socialist values”. They prohibit content that subverts state power, damages national unity, encourages separatism or harms the country’s image. Tests of DeepSeek have found evasions and official formulations on Tiananmen Square, Xinjiang, Taiwan and Liu Xiaobo.
The claim that every downloaded Chinese model is an immutable propaganda instrument goes too far. Open weights can be fine-tuned. Developers outside China can change system prompts, remove some safeguards and run the model without connecting to the original provider’s servers. Chinese law governs firms and services within Beijing’s reach; it does not acquire automatic jurisdiction over every foreign deployment.
The risk survives in a less theatrical form. Political choices can enter the training data, post-training process and evaluation criteria. Some are difficult to identify because they appear only when a user asks an indirect question. A foreign developer may modify the visible refusals while retaining factual distortions or omissions learned during training. Model weights are open; the underlying datasets and training decisions usually are not.
Chinese models also create familiar security questions. Code, update channels, cloud services and model dependencies can expose data or create access for an outside supplier. These risks vary greatly between a model downloaded and inspected locally and a service accessed through a Chinese cloud. Procurement rules that treat the two as identical will be analytically weak and commercially convenient for domestic incumbents.
Governments should require disclosure of the model’s origin, training and hosting arrangements when AI is used in public administration or critical infrastructure. Independent teams should test political bias, cybersecurity, data leakage and performance in local languages. Sensitive systems should be capable of running within national or trusted regional infrastructure. A label saying “made in China” is no substitute for such work.
The rest of the world needs an open alternative
A broad ban on Chinese models would protect American AI companies more effectively than it protects users. It would also push countries seeking technological autonomy towards Beijing. Mark Zuckerberg and several American technology companies have argued against restrictions on open-weight systems. Their commercial interests are evident, but the strategic point is sound.
The United States cannot answer a low-cost Chinese product only with export controls and expensive proprietary services. It needs competitive open models, cheaper access to compute and dependable terms for foreign users. Universities and smaller firms require support because the economics of open development do not favour laboratories spending tens of billions of dollars on closed systems. Safety controls should be tied to demonstrated capability and deployment, not nationality alone.
Europe has the regulatory authority to set procurement and testing standards but lacks enough competitive models and cloud capacity. India has a larger opening. Its language diversity, software workforce and digital public infrastructure make it a plausible developer of models for countries whose requirements resemble its own. Yet dependence on imported chips, foreign clouds and adapted foreign models will persist unless public compute and research funding are matched by disciplined procurement.
Most countries will use Chinese and American technology together. Siemens already works with models from both systems. Singapore’s national AI programme has built on Qwen. Developers choose models task by task, often without much concern for the flag attached to them. Governments will find it harder to remain casual when those choices enter courts, hospitals, schools or military supply chains.
China may still fail to become the undisputed global leader in AI. US chip design, frontier research and private capital remain formidable. Beijing is pursuing a more attainable position: principal supplier of the intelligence that much of the world can afford, alter and install. Leadership acquired through daily use can outlast leadership on a leaderboard.