China Is Flooding the AI Market: Why Is It Giving Models Away for Free?

Chinese large language models (LLMs) dominate the open-weight space. These models are freely available to download, run, and fine-tune, often under very permissive licenses. Some are also open-source, meaning that the code and training data needed to reproduce them are publicly available. According to benchmarks from LiveBench.ai as of September 20, 2025, these models outperform or match leading proprietary models from the US while costing only a fraction of the price.

These models are impressive and compete with leading closed models from the US on common benchmarks, while costing far less. One might wonder why Chinese technology companies are making AI models freely available, especially as the Chinese government supports open models as part of its AI strategy. In July, China published its Global AI Governance Action Plan, which emphasizes "international public good," "cooperation," and "openness." That sounds appealing until one realizes that China maintains one of the strictest and most heavily censored parts of the internet.

Why Is China Promoting Openness in AI?

Some of the stated objectives are straightforward. Open models facilitate the domestic adoption of AI technologies by increasing collaboration among Chinese corporations, the government, and academia. These models also accelerate iteration and allow researchers or military entities to fine-tune high-quality models to build capabilities that would otherwise be out of reach.

People unfamiliar with this may be concerned about data protection and security when using Chinese models. However, using these models does not require any interaction with Chinese servers or sending data to them. The best open models are hosted by US-based providers such as DeepInfra or Cerebras, which process queries in US data centers. Chinese companies have no visibility into this usage and generate no revenue from it.

One prominent hypothesis argues that if global developers adopt Chinese AI models, China could set standards that influence AI development worldwide. In the past, choosing open alternatives such as Android over iOS or Linux over Windows meant rewriting almost everything from scratch. But LLMs are different: they process and generate natural language, making them perhaps the most inherently interoperable technology ever invented. With tools such as OpenRouter, switching models is often as simple as changing a single parameter in the code. The same prompt can be processed by different LLMs. There is no lock-in here.

Comparison with China's Approach

Flooding the market with free, open LLMs resembles how China achieved dominance in manufacturing. Innovation in the US is driven by investors based on expected returns. Innovation in China is largely directed from the top down by the government according to national priorities. The Chinese Communist Party often prioritizes national capabilities over profits.

This is evident in China's approach to manufacturing. Over recent decades, China has provided massive government subsidies to manufacturers, producing and selling inexpensive products around the world, often at a loss. These enormous subsidies have enabled China to dominate key industries, including 80% of the solar energy supply chain and more than 70% of the world's electric vehicles. Prioritizing national capacity over profit has transformed China into the world's only manufacturing superpower, accounting for more than 30% of global manufacturing output—more than the next nine largest manufacturers combined.

China's subsidized manufacturing, unleashed on global markets, destroyed US manufacturing capacity, including sectors critical to defense. China now acquires advanced weapons systems five to six times faster than the US and has more than 230 times the shipbuilding capacity. It has thereby created an enormous military advantage by strengthening its wartime production capabilities while undermining those of the US.

Impact on American AI Labs

Similarly, open models reduce the pricing power of AI labs, making it harder to justify massive investments in research and development. OpenAI spent years and billions of dollars producing GPT-5, only to receive mixed reviews, including disappointment from some users and a revolt from others. Meta spent billions on AI infrastructure and training, but Llama 4 received a lukewarm response and failed to outperform cheaper, smaller open Chinese models. Why spend billions training a model when a better Chinese alternative may arrive tomorrow—released for free and hosted at minimal cost by competing US providers?

Open AI models are great for AI users, including startups, brands, and even AI inference providers—just as cheaper manufacturing is good for consumers and sellers of inexpensive products. AI startups such as Lightningrod.ai benefit from being able to train on these impressive open models. But—whether intentionally or not—lower AI inference prices make things more difficult for American companies developing foundational AI, just as cheap, subsidized manufacturing undermined American factories.

Outsourcing brought short-term savings but created long-term dependencies. When the US moved manufacturing overseas, it did not merely lose factories—it stopped developing the skills, knowledge, and supplier networks needed for production. Essential capabilities such as machine-tool capacity were cut in half. Even if bringing manufacturing back made economic sense, the industrial ecosystem needed to support it no longer exists. Freely available open models could have a similar, though smaller, impact on AI by pushing more engineers to specialize in fine-tuning models and fewer to build models from scratch.

The Response in Silicon Valley and Washington

Silicon Valley is concerned. Sam Altman has openly expressed concerns about China, particularly its open models. In a recent Washington Post commentary, he warned that "a democratic vision for artificial intelligence must prevail over an authoritarian one," citing Putin's prediction that whoever leads in AI will become the "ruler of the world," as well as China's ambition to become the global leader in AI by 2030. Anthropic's Dario Amodei described chip export controls as existentially important for slowing China's progress in AI. Marc Andreessen called DeepSeek AI's "Sputnik moment."

Washington has also taken notice of the threat and is encouraging American AI companies to release their own "leading open models founded on American values." OpenAI recently did just that by releasing the open GPT-oss model, acknowledging that Chinese open models motivated the move.

But this response could play directly into China's hands. A race to the bottom benefits Chinese labs because they are less dependent on potential profitability and returns on investment to secure funding.

Could It Work?

On the one hand, almost certainly not—China will not be able to derail our AI industry as effectively as it did manufacturing. Too many powerful interests, both public and private, understand what is at stake in this race. And too many technology giants in the AI race have sufficiently deep pockets and massive unrelated revenue streams to fund continued progress, even with thin—or negative—AI margins. American labs still have a significant lead at the AI frontier.

On the other hand... It is already having some impact. Without Chinese open models, inference prices from frontier labs in the US would probably be higher than they are today. Businesses and consumers would have less exposure to Chinese models if using them meant sending data to Chinese servers, so new models released in the US would appear more impressive. The AI market would be less competitive, but returns for US labs would be higher. The whispers about an AI "bubble" might not be so loud.

China's low-cost manufacturing gave it the capacity to surpass the US in everything from consumer goods to military equipment and created dependencies that could prove fatal in a conflict. Flooding the market with free, high-quality AI models could do the same, albeit on a smaller scale: building China's AI capabilities while making it harder to maintain those of the US. It worked once. Why not try it again?

Source: bturtel.substack.com