China May Be Ending the Open-Model Era

Two weeks ago, we wrote about GLM-5.2, the new open-weight model from Chinese AI startup Z.ai that came remarkably close to the performance of the best American models at a fraction of the cost.

Our argument was that the AI race was beginning to shift. The winner might not be whoever built the smartest model, but whoever delivered the most intelligence for every dollar spent.

It turns out we may have overlooked another increasingly important variable: whether the rest of the world will be allowed to use the model at all.

Reuters reported last week that Chinese authorities have been meeting with Alibaba, ByteDance and Z.ai about potentially restricting overseas access to the country’s most advanced AI models. The discussions reportedly include both closed and open-weight systems, as well as models that have not yet been released. Officials also discussed treating leaks or theft of proprietary AI technology as national security offences.

Nothing has been finalized. The restrictions may only apply to future models, and it remains unclear whether they will ever come into force. But the direction of travel matters here. For the past several years, China has played a very different role in the global AI ecosystem than the United States.

The leading American labs largely kept their strongest models closed. OpenAI, Anthropic and Google sold access through APIs and consumer products, but users could not download the underlying weights, run the models independently or guarantee they would always have access.

Chinese labs took the opposite approach. DeepSeek, Alibaba’s Qwen, Moonshot and Z.ai released increasingly capable open-weight models that developers around the world could download, customize and run themselves.

That openness was an extremely effective competitive strategy, especially when it came to getting power users (notably developers) onboard.

When you are behind, open source helps you catch up.

It gives developers a reason to try your model. Researchers improve the tooling around it. Startups build products on top of it. Every download expands distribution, every fine-tune creates another use case and potentially another enterprise contract, and every integration weakens the incumbent’s grip on the market.

China benefited from that adoption, but so did the West. American startups gained access to Chinese research without paying American frontier-model prices. Enterprises gained bargaining power against the closed labs. Developers got frontier-adjacent capabilities they could run on their own infrastructure.

GLM-5.2 was perhaps the clearest expression of that bargain. The rest of the world gained cheap intelligence. Z.ai gained global relevance.

But openness becomes less attractive once you have caught up.

At that point, the model is no longer just a product. It is intellectual property, economic leverage and potentially a national security asset. Giving it away may accelerate adoption, but it also gives foreign companies and governments access to capabilities you spent billions developing.

This is not a uniquely Chinese realization.

OpenAI began as a nonprofit research lab that said its researchers would be encouraged to publish papers, code and patents and collaborate freely with other institutions. By 2019, it was already experimenting with staged releases for GPT-2 because of misuse concerns. By the time GPT-4 arrived, access came through ChatGPT and an API rather than downloadable model weights.

The logic is straightforward: when you are behind, go open. When you reach the frontier, go closed.

China may now be arriving at the same conclusion.

ByteDance’s Doubao, China’s most widely used AI model, already relies on proprietary models. Alibaba and Z.ai still maintain major open-weight families, but Reuters reported that Chinese legal experts have discussed a tiered system where basic tools remain available, advanced models face security reviews and the most sensitive frontier systems are restricted to domestic use.

If that becomes policy, GLM-5.2 could look less like the beginning of a new open-source era and more like the end of one.

The current model is already available, which makes fully pulling it back difficult. Once weights have been downloaded and mirrored, they cannot be recalled like an API can be switched off. But future versions can be withheld. Overseas access can be blocked. The next GLM may simply never leave China which would have meaningful consequences.

Intelligence per dollar would get worse. Chinese models have provided a low-cost alternative to expensive American APIs. Removing the strongest of those alternatives would push more workloads back toward OpenAI, Anthropic and Google, giving the closed labs more pricing power.

It would also expose how dependent the Western open-source ecosystem has become on Chinese research. According to Reuters, Chinese developers currently hold the leading positions across open models for coding and agentic tasks, while thousands of American startups, companies and researchers rely on Chinese systems. Two weeks ago, GLM-5.2 looked like evidence that AI was becoming a commodity.

At the risk of overreacting to what may very well be China posturing, today it looks like evidence of the opposite.

The underlying technology may be getting cheaper and more abundant, but access to the best versions may become more political. The next phase of the AI race may still be about intelligence per dollar, but only after governments decide who is allowed to buy the intelligence in the first place.