On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology jointly issued Announcement No. 48, proposing major revisions to the Catalog of Technologies Prohibited and Restricted from Export. The revisions cover 40 technologies across three major areas—AI, biopharmaceuticals, and solar energy—with AI accounting for 23 items, spanning large models, AI agents, computing clusters, and information processing technologies. The public comment period on the draft closes on November 15, 2026.
On October 22, according to Reuters citing an account associated with official media, DeepSeek and MiniMax were the first to complete export registration under the Ministry of Commerce's new system, becoming the first batch of AI companies authorized to continue providing services to overseas users.
Where the Regulatory Boundary Is Drawn
The draft's technical triggers are quite precise: models with 2 trillion parameters, or those entering the global top ten in mainstream benchmarks, fall within the restricted export scope. Regulated items include model weights, checkpoints, source code, and architecture. Large models trained with computing power on the order of 10²⁶ FLOP that are trained on core data and then used to provide services are proposed for direct inclusion in the prohibited export category—the draft divides the 40 technologies into 14 prohibited and 26 restricted.
However, the Ministry of Commerce simultaneously specifies a series of circumstances not subject to jurisdiction:
- Providing services externally through API calls, official websites, mobile apps, or SaaS does not constitute technology export and is not restricted;
- Technologies already in the public domain through open source licenses or academic papers are entirely excluded from regulation;
- Building and operating computing centers overseas is eligible for "automatic licensing" facilitation measures;
- Technologies involved in overseas chip tape-out and manufacturing are likewise eligible for automatic licensing.
According to Taiwan tech media outlet INSIDE, a Ministry of Commerce spokesperson explicitly emphasized that "the vast majority of AI technologies can still be freely exported," and that restrictions target "the act of transferring the technology itself from within the country to abroad," not the reach of commercial services.
This boundary draws the policy's impact far narrower than it initially appears, while also providing the institutional basis for DeepSeek's and MiniMax's rapid clearance: both companies adopted open source release paths, their model weights long having entered the public domain, and under the draft framework they would already qualify for automatic licensing or even full exemption.
The "Open Source Passport": A Legal Channel
The cases of DeepSeek and MiniMax were quickly interpreted in the industry as "open source as a talisman." Once model weights are publicly released under an open source license, the technology enters the public domain in the legal sense, and national-level export controls cease to apply. In the draft's own words, such technologies "are no longer within the scope of regulation."
What does this mean for China's leading AI companies? Open source release is no longer merely a business or community operations strategy—it simultaneously constitutes a legal passport for going abroad. Against the geopolitical backdrop of US-China tech decoupling, the value of this pathway has exceeded purely technological dissemination logic.
From an institutional design perspective, the draft contains a structural asymmetry: open source release is irreversible—once a model enters the public domain, the government cannot retrieve it; but the regulatory framework itself can be modified. The current draft's open source exemption clause is transitional text established during the public comment period, and the final version has not yet been published. If a future revision narrows the exemption scope or sets a retroactive time threshold for determining what "already in the public domain" means, the legal status of existing open source models will face the risk of redefinition.
Beijing's Real Calculation: Layered Sovereignty, Not a Comprehensive Blockade
Reading this draft as a "Chinese version of the chip ban" is a common misjudgment. The core of US export controls against China is hardware scarcity—restricting the supply of advanced GPUs to physically limit computing power accumulation. China's draft operates at a completely different level: it regulates intellectual property and capability transfer, not access to computing resources.
This difference determines that the two regimes are fundamentally different in nature. GPUs are scarce goods—export controls can create real bottlenecks. Model weights are information—in the information age, controlling the flow of information carries marginal costs close to zero, making enforcement inherently more difficult.
What the draft truly aims to lock down is another category of assets: according to INSIDE, the draft explicitly lists "AI-driven malicious code generation and mutation technology" as prohibited from export and "AI-assisted vulnerability discovery technology" as restricted from export. These are the policy's core concerns—AI tools with military potential or cyberattack capabilities, not the overseas commercial deployment of general-purpose large language models.
UK tech media outlet The Register noted in its reporting that Beijing's arrangement simultaneously satisfies two objectives: on one hand, allowing "champion enterprises" like DeepSeek and MiniMax to continue expanding in overseas markets and maintain China's global AI influence; on the other, through establishing registration and licensing mechanisms, retaining policy room to apply the brakes on AI technologies that "represent important sovereign capabilities."
In other words, this is not a gate locking technology inside the country, but a layered sovereignty declaration: the vast majority of commercial AI technologies flow freely, while a small number of strategically significant capabilities have their export pace controlled by the state.
The Asymmetric Position of Open Source and Closed Source in the Regulatory Landscape
The draft inadvertently divides Chinese AI companies' competitive strategies into two fundamentally different situations.
Companies choosing the open source path—DeepSeek, MiniMax, and even Zhipu's GLM series—actually gain an implicit competitive advantage under the new framework: open source means compliance; global users can directly download, deploy, and build upon the models, with minimal regulatory friction. This aligns closely with their original business logic, yet under the policy's unintended support it becomes a compliance endorsement.
Companies choosing the closed source path face a completely different situation: every transfer of weights to overseas partners and every sharing of architecture design requires registration and approval processes. In an AI industry where competitive windows are measured in months, the time cost of licensing cycles is a real commercial price.
This asymmetry may inadvertently push China's AI industry toward open source—not because open source is commercially superior, but because it is less costly from a compliance standpoint.
Knock-On Effects for the Global South
Industry analysis suggests the potential spillover effects of these controls are particularly profound for AI ecosystems in Africa, Southeast Asia, and Latin America. Many research institutions and startups in these regions have long relied on open-weight models as the infrastructure for local AI development. If future revisions narrow the open source exemption or impose stricter limits on overseas deployment of closed-source frontier models, these regions' access to high-quality AI capabilities will directly contract.
The exemption for API and SaaS models leaves an alternative pathway for these user groups—calling Chinese AI services through network interfaces rather than deploying locally. But this means critical AI infrastructure must depend on overseas cloud connections, which is not a reliable option in regions with weak network infrastructure and also raises new concerns about data sovereignty.
Independent Assessment
China's draft AI export controls are essentially an act of sovereignty marking, not a blocking action. The actual regulatory intensity is far lower than the initial headlines suggested, but the policy toolbox left in place is quite complete: registration systems, licensing mechanisms, tiered catalogs—once established, these can be tightened at any time.
DeepSeek and MiniMax completing clearance within six days demonstrates the effectiveness of the open source pathway under the current framework. But it also means that the compliance position of Chinese AI companies in global markets is increasingly bound to their open source strategies—a new type of institutional dependency whose sustainability depends on how subsequent revisions define the scope of the "public domain."
What is most worth tracking long-term in the draft is not which models are listed in the catalog, but whether the open source exemption clause can maintain its current lenient wording in the official version. If the final version tightens here, the entire industry's compliance pathways will need recalibration.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接