Why China Is Opening Its AI Models While US Labs Guard Their Most Advanced Systems

Free Models Are Challenging the Business of AI

Chinese artificial intelligence companies are challenging US technology leaders with a proposition that is reshaping the competition: powerful models that other developers can obtain, adapt and build into their own products. As of August 4, 2025, DeepSeek, Alibaba's Qwen and Zhipu's GLM-4.5 have moved to the forefront of global rankings, putting pressure on OpenAI, Google and Meta.

DeepSeek illustrates the scale of the disruption. Its R1 and V2 models, released with open weights and permissive licenses, outperformed many Western counterparts on key benchmarks. The reported cost of training R1 was under $6 million, compared with the hundreds of millions spent on models such as OpenAI's GPT-4. Those figures have become central to the argument that advanced AI can spread without every developer paying for access to a closely guarded system.

US technology stocks suffered a selloff of about $1 trillion as investors worried that cheaper, widely accessible AI could weaken the businesses built around proprietary models. Developers in the United States and Europe adopted DeepSeek's technology, while major American companies faced fresh pressure to reconsider their strategies.

The divide is not absolute. Meta and Google release open models, and the White House has made open source AI a national priority. Chinese companies, meanwhile, operate within a system of strong state guidance and strict internet censorship. The contest is over how much technology to share, how to make money from it, and whose systems become the foundation for AI around the world.

What Developers Actually Receive

Open source AI makes model software and, crucially, model weights available to the public. Weights are the numerical settings that a model learns during training. They help determine how it responds to an instruction, answers a question or produces other output.

Access to those weights allows researchers, startups and multinational companies to use, modify and build upon a model. Under permissive licenses, they can do so without restrictive licensing terms or costly access fees. The practical difference is between obtaining technology that can be adapted directly and paying a provider to use a system that remains under its control.

Open source software already underpins familiar technologies, including the Linux operating system and the Android mobile platform. In AI, the same collaborative approach lets developers in different places improve a model, adapt it to particular tasks and share their work with others.

Researchers and smaller businesses gain a lower barrier to entry. Wider access can also support transparency and efforts to make systems more robust and trustworthy. The same openness brings risks: powerful capabilities can be misused, security becomes harder to manage, and technology that companies once sold at a premium can become widely available at much lower cost.

China's Strategy Has Deep Roots

China's interest in open source technology predates the current AI competition. In the early 2000s, government initiatives such as the Red Flag Linux project sought to reduce dependence on US software. Open source later became part of a broader push for technological self reliance.

US restrictions on access to advanced chips and software have given that objective new urgency. Chinese companies face pressure to develop technology they can improve and deploy despite limits on the foreign tools available to them. Sharing models can help a wider domestic community work from the same foundation.

The Chinese government supports research, talent development and infrastructure. Companies including Alibaba, Baidu and Tencent, alongside startups such as DeepSeek and Zhipu, lead model development and releases. Universities and researchers contribute to an ecosystem that combines state direction, private business and academic collaboration.

Platforms such as Gitee and organisations such as the OpenAtom Foundation support domestic open source communities. Chinese developers are also among the most active contributors to global platforms such as GitHub. China's strategy therefore includes both building local institutions and participating in international development networks.

The policy has evolved across several periods:

• Early 2000s: Initiatives including Red Flag Linux sought to reduce reliance on US software.
• Over subsequent years: Open source became more closely tied to technological self reliance, particularly as US export controls tightened.
• During the year leading to August 4, 2025: Chinese AI companies advanced rapidly through releases that challenged leading US models.
• By August 4, 2025: Washington had also elevated open source AI to a national priority, even as leading American labs continued to protect their most advanced systems.

DeepSeek Turns Efficiency Into Competitive Pressure

DeepSeek's appeal combined performance, access and the reported cost of development. Its R1 and V2 models arrived with open weights and permissive licenses, giving developers a route to advanced AI without relying entirely on paid access to a Western provider.

The reported training figure of under $6 million for R1 stood in sharp contrast to the hundreds of millions associated with models such as GPT-4. That comparison helped drive interest in whether expensive development and tightly controlled distribution would remain the dominant business model.

Financial markets reacted to the prospect that advanced AI could become a cheaper and more widely available product. The selloff of about $1 trillion in US technology stocks reflected anxiety that proprietary systems would face growing price competition and that the value of exclusive access could erode.

Meta chief executive Mark Zuckerberg reportedly ordered a complete overhaul of the company's AI strategy after DeepSeek's success. The disruption also helped ignite a talent war in Silicon Valley as US companies scrambled to catch up.

Nvidia chief executive Jensen Huang praised the quality of Chinese models, describing them as:

world-class

Meta chief AI scientist Yann LeCun framed DeepSeek's rise as a success for the development approach itself, calling it an:

open-source triumph

Those reactions capture two connected pressures on US companies: Chinese models are becoming serious technical competitors, and open distribution lets their advances spread quickly beyond the companies that created them.

Why Leading US Labs Still Keep Their Best Models Closed

OpenAI, Anthropic and Google DeepMind tightly guard their latest systems. They cite concerns about misuse, security and the possibility that adversaries could repurpose powerful AI. Keeping control over access gives a provider more authority over how its technology is distributed.

Commercial incentives also matter. Proprietary models can generate revenue through subscriptions, application programming interfaces and enterprise contracts. An application programming interface, or API, lets a business send requests to a provider's model and receive its responses without obtaining the model itself.

Subscriptions and contracts give companies a clear route to earning money from their technology. Freely distributing a powerful model can increase adoption, but it also puts pressure on charging customers for capabilities they may be able to obtain elsewhere.

Meta's Llama and Google's Gemma show that US participation in open AI continues. Their releases, however, often carry restrictions or sit at a technical distance from the companies' most profitable systems. An open model release and a company's leading commercial service can therefore occupy different places in the same business strategy.

US policymakers also view access through the lens of national security. They worry that powerful models could assist hostile actors or weaken America's technological lead. Recent US AI action plans stress strategic control, export restrictions and cooperation with trusted partners as Washington seeks to counter Chinese influence.

Washington Is Also Reconsidering Openness

The rise of Chinese models has strengthened the argument within the United States that leadership requires more than exclusive control of the most capable systems. Some policymakers and industry leaders argue that America must lead in open source AI as well, both to reflect democratic values and to retain global influence.

The White House has elevated open source AI to a national priority. That shift recognises the importance of collaborative development and the ability of widely shared technology to attract developers around the world.

American companies and policymakers still face competing demands. Businesses need revenue, security officials want controls, and developers benefit from access and international cooperation. The tension among those goals remains unresolved.

The competition is therefore more complicated than an open China facing a closed United States. Both countries recognise the value of broad adoption, while companies on both sides must decide which capabilities to share and which to reserve for commercial services.

Free Models Carry Influence Beyond China

Chinese models including DeepSeek, Qwen, Kimi K2 and GLM-4.5 now lead global rankings and offer high performance at a fraction of the cost of Western alternatives. Chinese and US companies are integrating them into products, reducing business costs and widening access to advanced AI.

Qwen3-Coder and DeepSeek R1 also serve a diplomatic purpose. By releasing powerful models freely, China presents itself as a supporter of accessible and equitable AI. The approach appeals to developing countries and the Global South, where the cost of obtaining advanced technology can shape who gets to use it.

China's message connects access with digital sovereignty: the ability of countries to make decisions about the technology and digital systems they depend on. At international forums, Beijing promotes an inclusive, multilateral approach to AI governance anchored in the United Nations.

That position is winning goodwill in emerging economies by offering an alternative to technology ecosystems dominated by US companies, with fewer political conditions. Broader use of Chinese models can also give China a greater role in shaping technical standards and the way AI is deployed.

At the World Artificial Intelligence Conference in Shanghai, Premier Li Qiang criticised technological monopolies and called for international cooperation to overcome bottlenecks, including access to chips. His message linked the expansion of AI access with the physical resources needed to build and operate it.

Chip Restrictions Put Efficiency at the Centre

US export controls on advanced chips have pushed Chinese companies to find ways to train and deploy models with fewer resources. DeepSeek and Zhipu have pursued efficiency, sometimes using domestically produced hardware or creative workarounds to resource constraints.

Training is the process through which a model learns patterns from data. Deployment is the stage at which people or businesses use it. Reducing the resources required at either stage can make a model more practical for organisations that cannot afford the largest computing systems.

China's model development also benefits from extensive Chinese language datasets. Those resources give its models an advantage in Chinese language tasks, while integrated technology ecosystems in cities such as Shenzhen support rapid prototyping and deployment.

These strengths help AI spread across manufacturing, healthcare, public services and consumer applications. A useful model becomes more valuable to businesses when it can be adapted and put to work quickly, rather than remaining a research result.

China's generative AI algorithm registration database provides an advanced and transparent element of its regulatory infrastructure, supporting responsible development and expansion. The system sits alongside the country's broader effort to move AI from development into widespread use.

Open Technology Still Faces Questions of Trust

China's strict internet censorship regime creates tensions for its open AI strategy. Developers and customers must consider how content controls affect model behaviour and whether a system can be adapted to different expectations and standards around the world.

Data security, intellectual property and transparency are also concerns. Some countries have restricted the use of Chinese AI tools, particularly in sensitive sectors. For businesses choosing a model, low cost and strong performance are only part of the decision.

International enterprises have particular reasons to examine systems such as GLM-4.5. Its appeal includes multilingual processing and workflows driven by AI agents, systems that carry out connected tasks rather than simply answer a single question. Those uses can bring a model closer to important business information and decisions.

Companies therefore face demands for stricter evaluation frameworks, transparent governance and compliance with the rules that apply to their operations. National security concerns can also affect whether a model is suitable for a particular organisation or sector.

China faces a further tension between building an independent technology ecosystem and keeping its developers connected to global collaboration. Open source development benefits from diverse contributors and exchange across borders. Excessive state control or ideological constraints could limit that exchange and weaken competitiveness over time.

The Contest Is About Rules as Well as Models

China's state guided approach and America's more market driven model are shaping rival visions of AI governance. Beijing advocates a framework based on consensus, the United Nations, sovereignty and equitable access. Washington seeks a global AI alliance built around American infrastructure, values and strategic interests.

For developing countries, the immediate issue is often practical: which models they can afford, adapt and use within their own institutions. Chinese releases offer accessible technology, while US strategy stresses partnerships with trusted countries and control over sensitive capabilities.

Both governments recognise AI's ability to transform economies and the risks that accompany it. Their strategies differ over the balance between sharing technology, protecting national interests and retaining influence over the systems others adopt.

The United States risks losing influence if it cannot respond to the spread of open models. China faces its own limits if censorship, distrust or isolation reduces the willingness of developers and governments to build on its technology. Neither approach is absolute, and both countries are adjusting under competitive pressure.

The central choices concern access, security and cooperation. How governments and companies balance those priorities will shape the technology people use, the businesses that can compete and the rules governing AI across borders.

Key Points

• As of August 4, 2025, DeepSeek, Qwen, Kimi K2 and GLM-4.5 were challenging US AI leaders through powerful, widely accessible models.
• DeepSeek R1 reportedly cost under $6 million to train, while models such as GPT-4 involved spending of hundreds of millions.
• Open weights and permissive licenses let developers adapt models directly, lowering barriers for researchers and businesses.
• China's strategy combines state support, private innovation, academic collaboration and a push for technological self reliance.
• Leading US labs protect their most advanced systems for commercial and security reasons, while Washington has also prioritised open source AI.
• Chip restrictions have encouraged efficiency, helping Chinese models spread into business and public services.
• Censorship, data security, intellectual property and national security remain central concerns for international adoption.
• The competition will help determine both who supplies AI and who shapes the rules governing its use.