China’s artificial intelligence sector has a new bottleneck, and it isn’t export-controlled chips. The country is rapidly running out of high-quality Chinese-language training data. For months, the public debate about China’s AI ambitions has centered on US export controls and advanced semiconductors. Chinese experts now warn that data scarcity could prove just as limiting, and unlike chips, it has no hardware workaround.
Chinese is only 1.3% of global web content, according to the web technology tracker W3Techs. English accounts for nearly half of all web content, Spanish for 6%, and Japanese for 5%. That statistic may seem abstract, but it becomes concrete when an AI model needs to answer a question, summarize a legal document, or hold a natural conversation in Chinese. With less source material, models are prone to linguistic awkwardness, factual inaccuracies, and limited cultural nuance.
Key facts
- Chinese is just 1.3% of global web content, compared with 49% for English.
- Epoch AI estimates publicly available high-quality text could be exhausted within six years.
- OpenAI co-founder Andrej Karpathy has warned of a data wall by 2030.
- WeChat and Douyin do not share external data with AI developers.
- China’s National Data Administration plans to build validated AI training datasets by 2028.
- Huaxia Publishing House has added an AI training ban to a new book translation.
The problem is global but hits China harder. Epoch AI, an international research group, estimates the worldwide supply of high-quality, publicly available text could be fully exhausted within six years. OpenAI co-founder Andrej Karpathy has warned of a data wall by the end of the decade. Chinese developers already pay more per useful token than Western counterparts because their models must work harder with less native-language material. Every training run needs massive amounts of text, and if that text is repetitive, low-quality, or poorly sourced, the resulting model suffers.
China’s digital ecosystem makes the shortage worse. Platforms like WeChat and Douyin have enormous amounts of user-generated content, but they are walled gardens. They do not share data with third-party developers, leaving AI labs to train on lower-quality sources like old web crawls, academic papers, and government documents. The most valuable modern Chinese language data, the kind that reflects how people actually speak and write today, remains locked inside platforms that have no commercial incentive to release it.
Beijing’s data infrastructure plan
Beijing is responding by treating data as strategic infrastructure. In June, the National Data Administration unveiled a nationwide plan to build validated AI training datasets by 2028. The initiative covers manufacturing, energy, healthcare, finance, agriculture, autonomous driving, and embodied AI. The plan is not just about collecting more Chinese text. It aims to create standardized, high-quality datasets with clear ownership, quality control, and access rules.
Yu Xiaohui, president of the state-affiliated China Academy of Information and Communications Technology, put it plainly: Competition in the AI era is not only about models and computing power, but also about high-quality data supply systems. His remark reflects a growing consensus in Chinese policy circles that data is a national resource on par with energy or rare earths.
Tsinghua University computer scientist Sun Maosong has urged authorities to digitize historical archives, ancient manuscripts, scientific literature, and regional dialects. This is a reminder that China has a deep cultural heritage that could become a unique training advantage. Classical Chinese texts, local operas, minority languages, and centuries of philosophical literature are largely absent from public web datasets. Unlocking those materials would require enormous digitization efforts, but the payoff could be models that understand Chinese civilization better than any foreign competitor.
Publishers push back
Not everyone wants to be digitized. Huaxia Publishing House recently added a warning to a new translation: It is prohibited to use the content of this book for artificial intelligence training. Violators will be held legally responsible. That is a significant shift. Chinese publishers have historically been passive about digital rights, but AI has changed the calculus. If a model can reproduce text from a copyrighted book, publishers see direct revenue loss. Similar dynamics are playing out globally, with authors and news outlets seeking licensing deals or suing AI companies.
The publishing resistance creates a paradox. Governments and policymakers want more high-quality data for AI, while rights holders are tightening access. The data wall is not just a technical limit. It is a legal and economic limit. Even if a company has the technical capacity to scan millions of books, it cannot do so without permission, or it risks litigation.
How the US compares
The United States has one notable example of aggressive data acquisition: Anthropic’s Project Panama. According to reports, the project bought and destroyed millions of physical books to scan them, presumably to build a corpus without violating copyright laws. The strategy worked because physical copies could be legally purchased, scanned, and then discarded. It is a stark illustration of how far AI companies will go to secure scarce text.
China cannot easily replicate such projects. The country’s book market is vast, but digitization efforts are fragmented. Many ancient manuscripts and regional texts have never been digitized. Universities and libraries lack the resources to scan everything. And the publishing industry, as Huaxia’s warning shows, is starting to lock the door.
The data shortage also explains part of the gap between American and Chinese AI models. American labs have access to an enormous pool of English text, including books, scientific papers, and high-quality web pages. Chinese labs must make do with a much smaller native-language pool, and they often have to translate or generate synthetic data to compensate. Translation is useful, but it cannot capture the idiomatic freshness of native text. Synthetic data can help, but it can also amplify biases and errors.
What the future holds
China’s plan to build national datasets by 2028 is ambitious, but the timeline may be too slow. Global AI progress is accelerating, and every month of data scarcity puts Chinese developers further behind. The country has a clear advantage in population and digital scale, but that advantage is neutralized by proprietary platforms and rights restrictions. WeChat could theoretically provide a goldmine of conversational Chinese data, but its parent company, Tencent, has no obligation to share it. Douyin, the Chinese sibling of TikTok, is similarly closed.
There are possible workarounds. Some Chinese AI labs are exploring synthetic data generation, where models create new training examples based on existing ones. Others are investing in optical character recognition to digitize old books and newspapers. Some are even crowdsourcing language data from users. But these efforts remain smaller than the scale of the problem.
The international dimension is also important. If Chinese AI developers cannot access foreign data due to language barriers and legal restrictions, they may focus on Chinese-specific applications. That could lead to a split in AI development: one ecosystem for English-speaking markets, and another for Chinese-speaking markets. This split would be costly for both sides, but it may be inevitable if data does not flow across borders.
For now, China faces a hard reality: chips were the first bottleneck, and data may be the second. The country has the engineers, the computing power, and the political will to build advanced AI. What it lacks is an abundant supply of native-language words. As the global data wall approaches, every country will need to decide how to preserve, digitize, and open high-quality linguistic resources. China has begun that debate, but the clock is ticking.