China’s Next AI Bottleneck May Be Data, Not Chips

For years, the discussion around China’s AI ambitions has focused heavily on computing power. US restrictions on advanced semiconductors have made access to high-end AI chips one of the most visible obstacles facing Chinese technology companies.

Now, another constraint is coming into focus: data.

Chinese AI researchers are warning that the country faces a shortage of high-quality Chinese-language material for training increasingly capable AI models. According to the South China Morning Post, the problem could become a major bottleneck that cannot be solved simply by finding alternatives to restricted hardware.

The issue is also much larger than China. Estimates suggest that the global supply of high-quality, publicly available human-generated text could be exhausted within the next six years. As AI models consume more of the accessible internet, developers need to find new sources of reliable information.

Chinese Content Is Underrepresented Online

China’s challenge is particularly noticeable because of the composition of the public web.

Chinese accounts for only around 1.3% of website content, compared with approximately 49% for English. This does not mean that China lacks digital information. Platforms such as WeChat and Douyin contain enormous volumes of Chinese-language content, but much of it is confined to closed ecosystems and is not freely available for AI training.

That creates an unusual situation. China has more than a billion internet users and some of the world’s largest digital platforms, yet the amount of accessible Chinese-language material suitable for training frontier models remains relatively limited.

Quality adds another constraint. Training a more capable model requires more than simply collecting billions of additional words. Developers need diverse, accurate, useful, and properly sourced information.

AI Is Approaching a Data Wall

The problem is not unique to Chinese AI companies.

AI researcher Andrej Karpathy has previously warned of a coming “data wall” toward the end of this decade. Once models have consumed most of the high-quality material available online, simply making them larger may yield diminishing returns.

Developers are already experimenting with alternatives. These include synthetic data generated by other AI systems, licensed datasets, private information, digitized books and archives, and more specialized industry data.

However, each option creates new questions.

Synthetic data can increase volume, but repeatedly training AI on AI-generated material risks reinforcing mistakes or reducing diversity. Private and copyrighted datasets create licensing and ownership challenges. Meanwhile, specialized information may be highly valuable but difficult to collect and standardize.

The AI race is therefore becoming, in part, a race for better information.

E-commerce Has Its Own High-Quality Data Problem

This matters directly for e-commerce because AI systems increasingly need more than general knowledge from the web.

A shopping assistant answering “Which laptop should I buy for video editing?” needs detailed information about processors, memory, displays, connectivity, dimensions, compatibility, availability, and many other characteristics.

General web text cannot reliably provide that level of structured knowledge across millions of products.

The same applies to AI-powered search, recommendation engines, automated product enrichment, customer service, and emerging shopping agents. Their usefulness depends on access to accurate product information that models can understand and compare.

As general-purpose training data becomes harder to obtain, specialized datasets can become more valuable.

Better Data May Matter More Than More Data

The Chinese AI data shortage highlights an important limitation of the current AI boom.

More computing power can train larger models, but processors cannot create reliable knowledge that does not exist in the training material. Once easily accessible information becomes scarce, the quality, structure, provenance, and diversity of data become increasingly important.

For e-commerce, this makes product data more than content used to populate a product page. It can become part of the knowledge infrastructure used by search engines, recommendation systems, AI assistants, and autonomous shopping agents.

The next improvements in AI may therefore depend not only on building larger models or faster chips, but on giving those systems better information to work with.