China Tightens AI Exports: Keeping the "Crown" at Home, Selling Access Abroad

Have you ever thought that one day, the most valuable asset a nation owns wouldn't be gold or oil, but the lines of code and weights of an artificial intelligence model? It sounds abstract, but that is precisely the battle taking place right now.

China has just released a draft that analysts call a turning point in how Beijing manages its "AI crown." Rather than imposing outright bans, it draws a clear boundary: keeping core technology domestically while selling access abroad. That is more than just an administrative regulation. It is a new business philosophy for the AI era.

Join me, SilverZ, in dissecting this draft from its roots to understand why it not only impacts Chinese companies but also reshapes how the entire world approaches artificial intelligence.

The October 16 Draft: When "Renting Access" Replaces "Selling Technology"

On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology published a revised draft catalog of technologies prohibited or restricted from export. The draft adds 40 technologies, with 23 belonging to the AI sector—the largest share—alongside 12 photovoltaic technologies and 5 biomedical technologies.

Yet the crucial point lies not in the numbers. The crucial point lies in how the draft distinguishes between two types of transactions:

• Foreign customers using Chinese AI services: Permitted via websites and cloud platforms.
• Foreign companies owning the model, data, or expertise to independently operate or reproduce the model: Subject to government approval.

This is the boundary analysts term "keeping IP and R&D at home, selling operations and access abroad." Alfredo Montufar-Helu, managing director at Beijing-based consultancy Ankura, noted that this structure "is the model that the capital market has already been converging toward." In other words, the state is not imposing anything entirely foreign. They are merely formalizing an already emerging trend.

Three Tests: Which Models Are "Framed"?

The draft does not vaguely state that "all AI models are controlled." It establishes three specific criteria to determine if a model falls under the affected scope:

• Computing volume for training: A threshold of 10 to the 26th power of calculations (10^26 FLOPs). This is a colossal level of computation that only the most advanced models reach.
• Model size: A minimum of 2 trillion parameters. This figure puts the world's largest models squarely in the crosshairs.
• Performance ranking: Ranking in the global top 10 across certain benchmarks for performance, usage level, or efficiency.

These three criteria show that the draft targets the most powerful models, not the entire AI industry. However, legal experts have pointed out ambiguities. Law firm Han Kun questioned how to count the size of models that only activate partial components for each task. Furthermore, a provision regarding chip design is believed to potentially reach further than even the most advanced products because it lacks a minimum performance requirement.

The draft also includes AI agents—systems executing tasks with minimal human direction—and coordination technology for more than 10,000 AI chips simultaneously. These are core "weapons" of an autonomous AI economy.

Practical Impact: DeepSeek, Moonshot, Z.AI, and the "Two Versions" Dilemma

You might be thinking: "So I can still use DeepSeek and call APIs normally?"

The short answer is: Yes, but with conditions. AI services via APIs, websites, mobile applications, and SaaS platforms remain outside the restricted scope. End users can still access Chinese models via the cloud.

Yet the issue lies at a deeper layer. Companies like DeepSeek, Moonshot, and Z.AI provide models to foreign clients through websites or cloud platforms. Some Chinese developers have even shifted to using overseas computing infrastructure to access the most powerful chips—those that the US has banned from exporting to China.

And this is the most headache-inducing part. Researchers at China's leading AI labs are seeking clarification: Will they be allowed to train or operate the most advanced models on foreign-owned overseas servers? In the worst-case scenario, developers might have to maintain two separate versions—one domestic, one overseas—of the exact same downloadable model.

Just imagine the consequences: doubled costs, doubled procedural complexity, and the tangible risk of divergence between the two versions. That is the price of an ill-defined legal boundary.

Beyond Models: Foreign Ownership and the Lesson of Meta and Manus

The draft does not stop at exporting models. It also reaches into corporate transactions.

According to analysis by Fangda Partners, registration requirements effective October 16 also cover deals that bring Chinese technology under foreign ownership or control—including transactions currently in progress.

And this is not a theoretical concern. Earlier this year, Beijing ordered Meta to divest from its over $2 billion acquisition of Manus, a Chinese-founded AI startup. That move highlighted the real risks foreign investors face when acquiring AI assets with Chinese roots.

For data centers, the draft provides a simplified approval mechanism for Chinese companies building and operating facilities abroad themselves. However, it does not clarify technology transfers to foreign clients for building equivalent systems—leaving yet another gap that legal professionals are waiting to be filled.

Counterbalancing the US: Controlling Inflows and Outflows

What makes this draft unique is that it is designed as a deliberate counterweight to US measures.

Yang Jingru, a partner at law firm Jincheng Tongda & Neal, described this new approach with a thought-provoking statement: "The US controls the inflow of computing power, hardware, and capital. China is beginning to systematically manage the outflow of models and technology assets."

This is a calculated symmetry. The US blocks chips, capital, and advanced technology from entering China. China retaliates by tightening the outflow of models, data, and expertise. The tech war is no longer a war over products. It is a war over who controls the flow of knowledge.

Montufar-Helu noted that security concerns—including keeping powerful cyber-attack capabilities and sensitive data onshore—come paired with the commercial goal of maintaining valuable technological advantages. This is not merely about security; it is about long-term competitive advantage.

Timeline and Unresolved Unknowns

The draft is open for public comment until November 15, 2026. According to a source familiar with the matter, China expects to adopt the proposal by the end of the year and issue implementation guidelines thereafter.

Yet many questions remain unanswered:

• How will open-weight models be affected? The draft could complicate transferring model weights from China abroad if regulatory approval is required.
• Which ranking lists will be used to determine the "global top 10"? The draft does not specify.
• How will transactions currently in process be handled?
• Will licensing for open-weight AI technology truly be "fast-tracked" as China's Ministry of Commerce claims, or will it become a de facto barrier?

These unknowns leave businesses and investors waiting in a state of limbo—a position no one desires when trying to map out long-term strategies.

Lessons for Vietnam: When "Access" Is No Longer "Ownership"

Even if this draft does not directly impact Vietnam yet, it brings strategic lessons that cannot be ignored.

1. The Difference Between "Using" and "Mastering"

China's draft draws a clear line: you can use my model via API, but you cannot own the weights, training data, or expertise to reproduce it yourself. This serves as a lesson for Vietnamese enterprises relying on foreign APIs. Using an AI service is one thing. Building internal AI capabilities is entirely different—and that is what creates sustainable advantage.

2. Data and Expertise Are Strategic Assets

China protects more than just models. They protect training data and operational expertise. For Vietnamese businesses, this serves as a reminder that user data, internal processes, and specialized knowledge are assets that must be managed and protected like intellectual property, rather than just "raw inputs."

3. Preparing for a Fragmented AI World

As the US controls inflows and China controls outflows, the AI world risks splitting into two ecosystems with distinct standards, models, and barriers. Vietnamese enterprises—with their geographical advantages and relations with both sides—must prepare for a world where choosing an AI vendor is not merely a technical decision, but a geopolitical one.

4. Building Legal and Compliance Capabilities

The controversies surrounding the draft—from computation thresholds and parameter counting methods to the scope of chip design rules—demonstrate that the rules of the AI game are being written as it is played. Vietnamese enterprises looking to collaborate with Chinese partners or utilize their models need legal and compliance capabilities to adapt to continuous shifts.

When the "Crown" Is Kept, But the "Kingdom" Remains Open

I have watched the global tech industry go through numerous battles over the past two decades. Yet I have never seen a conflict quite this sophisticated. This is not a total embargo war. It is a war of drawing boundaries—between what is permitted for sale and what must be kept, between access and ownership, between commerce and security.

China is not closing its doors. They are opening them with control. They still let the world use their AI—via APIs, via the cloud, via apps. But they retain the core part: model weights, training data, and reproduction expertise. It is a smart strategy, and one worth studying by every nation and business.

And the biggest question is not "Will China succeed?" The question is: In a world where AI knowledge is fractured by invisible walls, will you—as an individual, a business, or a nation—own your own "crown," or forever remain a renter of access?

What are your thoughts on how China is simultaneously opening up AI services while tightly closing the door on core technology ownership? Is it a wise strategy to protect national advantage, or a barrier that will cause Chinese AI companies to lose global cooperation and expansion opportunities? If you are a Vietnamese business considering building AI capabilities, which path would you choose—using foreign services to move fast, or investing in internal capabilities to go far? Share your views below, because in this AI game, those who understand the boundary between "using" and "owning" will never be left behind!