# What the China AI Export Controls Draft Actually Says > Source: > Published: 2026-10-29 04:43:31+00:00 China drew a fence around its most valuable AI technology on October 16, 2026. The Ministry of Commerce and the Ministry of Science and Technology released a draft update to the *Catalogue of Technologies Prohibited or Restricted from Export*, and [Fortune reported](https://fortune.com/2026/10/16/china-restrictions-ai-chips-solar-power-exports-security/) that Beijing is putting restrictions on AI, chips and solar power exports as security concerns grow. The Economic Times and other outlets have since framed the package as a line around China’s AI crown jewels, and that framing is now the most searched version of the story. The numbers are specific. The draft lists 40 technologies, and 23 of them touch artificial intelligence. It reaches large AI models, AI agents, AI-chip design, [memory-chip](https://shattered.io/category/hardware/) technology and the software that ties giant computing clusters together. One headline threshold stands out: models with 2 trillion or more parameters, or with specified world-leading capabilities, would face export restrictions. Those China AI export controls are still a draft, but a temporary registration rule went live the same day. This matters beyond Beijing’s policy circles. Chinese open-weight models already sit inside Western developer stacks, a trend we tracked when we covered [the DeepSeek and Moonshot distillation claims](https://shattered.io/deepseek-moonshot-gpt-6-astra-distillation-two-models-2026/). A rule that decides which pieces of that stack can leave China, and under what approval, changes the math for every company that builds on them. Below, we separate what the reports confirm from what remains open, then look at market impact, comparisons, history and what could happen next. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What the China AI Export Controls Draft Actually Says The draft is an update to an existing instrument rather than a new law. The catalogue is the list Chinese authorities use to decide which technologies cannot leave the country, and which can leave only with permission. By adding AI, chip and solar items, regulators extend a tool built for older industries to the newest ones. According to reports, the proposal covers 40 technologies in three areas: artificial intelligence, chips and solar power. Twenty-three are AI-related. Within that AI group, the draft names large AI models, AI agents, AI-chip design, memory-chip technology and computing-cluster technologies. That is a full stack, from the model at the top to the silicon and interconnect at the bottom. Two features set this draft apart from a typical export list. First, it uses capability and scale thresholds, such as a parameter count, rather than only naming product categories. Second, it ties restrictions to how a model was trained, with separate treatment for models built on “core data” and “important data.” Both ideas push the control from the product to the ingredients. The Straits Times [described the plan](https://www.straitstimes.com/business/china-plans-ai-export-curbs-to-exert-control-over-key-tech) as a bid to exert control over key technology. That reading fits the structure. Rather than banning Chinese AI abroad, the draft focuses on the pieces that let someone else reproduce it. ## The 40 Technologies and the 23 AI Items The table below maps the AI and chip categories in the draft to the treatment reports describe. Where the reports give a threshold, it appears in the middle column. Where they do not, the cell says so instead of guessing. | Category | Threshold or trigger reported | Proposed treatment | |---|---|---| | Large AI models | 2 trillion or more parameters, or specified world-leading capabilities | Export restrictions | | Models trained on “core data” | Training used China’s core data | Proposed export ban | | Models trained on “important data” | Training used important data | Proposed export restrictions | | AI agents for defense, critical information infrastructure or core government affairs | Intended use in those areas | Proposed export ban | | Computing-cluster connection and scheduling technology | Clusters with more than 10,000 accelerators | Covered by the draft | | AI-chip design | Manufacturing process of 14 nm or below | Covered by the draft | | Advanced chip packaging | 2.5D or 3D packaging | Covered by the draft | | Memory-chip technology | No threshold in the reports reviewed | Covered by the draft | The pattern is clear even without the fine print. Each row targets something a foreign party would need to run frontier AI without relying on China: the model, the data lineage, the agent, the cluster software, the chip design and the packaging. Remove any one of those inputs and the foreign copy gets harder to build. Note what the table does not show. Reports do not specify how regulators would count parameters for mixture-of-experts designs, or how they would test for “world-leading” capability. Those definitions will decide who is actually caught, and the draft leaves them to the consultation and the final text. ## The 2 Trillion Parameter Line for Large AI Models The parameter threshold is the line most readers will remember. A model with 2 trillion or more parameters, or one with specified world-leading capabilities, would fall under proposed export restrictions. The wording gives regulators two paths into the same rule: a size test and a performance test. Size tests age quickly. A parameter count says little about quality, since smaller models trained on better data regularly beat larger ones. By adding a capability limb, the drafters acknowledged that. A model that stays under 2 trillion parameters can still be caught if regulators judge it world-leading. For developers outside China, the practical question is which models sit near the line. The reports reviewed do not name any Chinese model as covered. They also do not publish parameter counts for the best-known Chinese releases. Anyone claiming that a specific model is now restricted is running ahead of the evidence. We compared several of these open-weight families in our [DeepSeek vs Qwen vs Mistral breakdown](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/). That piece shows how quickly Chinese labs have narrowed the gap with closed Western systems. The draft arrives as that gap shrinks, which is part of why it reads as strategic rather than purely technical. ## Core Data and Important Data: Training Sets Become Controlled Inputs The data provisions may prove more consequential than the size threshold. Under the draft, AI models trained using China’s “core data” would face a proposed export ban. Models trained using “important data” would face proposed export restrictions. The distinction is a ban versus a licensing regime. Chinese data law already sorts information into tiers, and these two labels come from that system. Linking export controls to those tiers means a model’s training history becomes a compliance fact. A lab that wants to ship a model abroad would need to show what went into it. That creates a new burden. Training pipelines at large labs mix web crawls, licensed corpora, synthetic data and domain sets. Tracing every shard back to a classification tier is hard work. Labs that kept thin records will struggle, and labs with clean data lineage will move faster. The reports do not say how regulators would verify training data claims, or who would audit them. Nor do they define the boundaries of “important data” beyond the label. Until the final text appears, legal teams at any firm touching Chinese models should treat those boundaries as unknown. ## AI Agents Aimed at Defense and Critical Systems AI agents get their own entry. The draft proposes export bans on agents intended for national defense, critical information infrastructure or core government affairs. The test here is intended use, not raw capability. An agent that books travel is untouched. An agent built to operate inside a power grid operator’s systems or a government workflow is not. The choice reflects how the industry has moved. Over the past year, labs shifted their focus from chat to systems that take actions, call tools and run multi-step jobs. Regulators noticed. An agent embedded in critical infrastructure is a different risk from a chatbot, and the draft treats it that way. Security teams have been asking similar questions on the defensive side. Our piece on [zero trust for AI agents](https://shattered.io/zero-trust-for-ai-agents-three-tiers-after-login-toledano-2026/) lays out why access after login is the hard part. China’s draft approaches the same worry from the opposite direction, by limiting where its own agents may go. ## Computing Clusters Above 10,000 Accelerators The infrastructure item is narrow and pointed. The draft covers technology for connecting and scheduling computing clusters that contain more than 10,000 accelerators. That is the know-how that turns a warehouse of chips into one training machine, including the networking, orchestration and fault handling. Cluster software is where much of the practical advantage in AI training sits. Buying chips is one problem. Keeping ten thousand of them busy, synchronized and recoverable after failures is another. By controlling that layer, China restricts the export of operating experience, not only hardware. The hardware race behind this is visible in the numbers Western vendors publish. We laid out the memory and capacity gap between rival accelerators in [our AMD versus Nvidia AI chip comparison](https://shattered.io/amd-vs-nvidia-ai-accelerators-mi455x-vs-vera-rubin-432gb-vs-288gb-2026/). Chinese firms operate under different supply limits, so scheduling software that squeezes more out of each chip carries extra value at home. ## Chip Design at 14 nm and Advanced Packaging The semiconductor items tie the AI rules to China’s chip ambitions. Proposed controls include AI-chip design using a manufacturing process of 14 nm or below, plus advanced 2.5D or 3D packaging. Memory-chip technology appears on the list too. ### Why packaging belongs on an AI list Advanced packaging stacks or tiles chips so that memory sits close to compute. For AI accelerators, that proximity often matters as much as the transistor node. A design that cannot be packaged well cannot feed its cores. Putting 2.5D and 3D packaging on the list shows regulators understand where the bottleneck lives. ### What a 14 nm line signals The 14 nm figure covers designs for that process and anything more advanced. It does not say that China can manufacture at that node at scale. It says that design knowledge at or below that level is considered strategic. Foreign firms that license Chinese chip design work, or Chinese firms that sell it, now face a registration step. Memory is the third leg. Training large models demands fast, plentiful memory, and the draft’s mention of memory-chip technology suggests Beijing wants to keep any progress on that front at home. ## Pre-Export Registration Took Effect October 16 The draft is not the only thing that moved. A temporary pre-export registration requirement took effect on October 16, 2026. It covers the technologies in the draft while their classifications remain under consultation. In other words, companies cannot wait for the final text to start complying. The requirement applies to three kinds of transaction: transfers of ownership abroad, licenses for overseas use, and changes that place ownership or actual control under a foreign entity. The third category is the broadest. A funding round, merger or restructuring that hands a foreign investor control of a covered technology could trigger it. TechTimes [warned readers](https://www.techtimes.com/articles/329196/20261019/chinas-new-ai-export-controls-apply-now-if-you-licensed-deployed-covered-chinese-models.htm) that the controls apply now to anyone who licensed or deployed covered Chinese models. That headline captures the urgency. Legal exposure begins with the interim rule, not with a future final catalogue. | Date | Event | Status | |---|---|---| | October 16, 2026 | Draft update to the export catalogue released by the Ministry of Commerce and the Ministry of Science and Technology | Confirmed | | October 16, 2026 | Temporary pre-export registration takes effect | Confirmed | | October 29, 2026 | Comment period still open, registration rule in force | Ongoing | | November 15, 2026 | Public comment period closes | Confirmed | | After consultation | Final catalogue text and classifications | No date confirmed | With the consultation closing on November 15, 2026, stakeholders have roughly two and a half weeks from today to file feedback. Industry groups, universities and foreign chambers of commerce typically use that window to push for narrower definitions. ## Websites and Cloud Access Stay Open, Core Transfers Need Approval The most reassuring piece for ordinary users concerns access. Under the draft rules, overseas users could generally still reach Chinese AI through websites and cloud platforms. What requires government approval is the transfer of the underlying models, data, expertise or deployment technology in some cases. That split between service and substance is the center of the policy. A person in Berlin or Boston who uses a Chinese chatbot or calls a Chinese model through an API is not the target. A company that wants to download weights, license the training recipe or stand up its own copy of a covered model is. Our earlier report on [China’s open-weight strategy and the DHS order on DeepSeek](https://shattered.io/china-open-weight-ai-strategy-dhs-deepseek-order-oct-20-2026/) described how open weights became both a growth tool and a security flashpoint. This draft complicates the first half of that story. If some transfers need approval, open release looks less like a default and more like a decision. The reports do not say whether routine open-weight publication of models below the thresholds would need approval. Developers should read that silence as unresolved rather than as permission. ## What Is Confirmed and What Is Not Coverage of this story has run hot, and some claims have outrun the evidence. The table below separates facts that multiple outlets support from items that remain unproven. | Claim | Status | Basis | |---|---|---| | Draft released October 16, 2026 by two ministries | Confirmed | Multiple outlets | | 40 technologies, 23 AI-related | Confirmed | Multiple outlets | | 2 trillion parameter threshold for large models | Confirmed as proposed | Draft as reported | | Interim pre-export registration in force | Confirmed | Reports on the interim rule | | Comments open until November 15, 2026 | Confirmed | Multiple outlets | | A named Chinese AI company, model or product is restricted | Unconfirmed | No such finding in the reports reviewed | | On-record quote from a named official or company | Unconfirmed | None found | | Final text adopted unchanged | Unconfirmed | Draft stage only | Three gaps deserve emphasis. No named Chinese AI company, model, executive or product has been shown to be restricted under the draft. We found no qualifying on-record quote from an official or company representative, which is why this article paraphrases the policy instead of quoting people. And nobody can say yet whether the final classifications will match the draft. Publications with deeper policy focus are still working through the text. The Diplomat [argues that China wants its AI to go global on its own terms](https://thediplomat.com/2026/10/china-wants-its-ai-to-go-global-on-its-own-terms/), which is a useful lens for the service-versus-technology split. ChinaTalk has also [examined the new Commerce Ministry rules](https://www.chinatalk.media/p/mofcoms-new-ai-rules) and what they mean for the future of Chinese AI. ## Market Impact: Who Feels It First Market reaction is hard to measure this early, and no verified price or revenue figure ties directly to the draft. What can be said is which groups face the earliest friction. The analysis below is our judgment based on how the rules are structured, not reporting on company decisions. ### Chinese model developers Labs that publish open weights face the clearest change. Each release now needs a quick internal check against the thresholds, the data tiers and the registration rule. Large labs have legal teams for this. Smaller teams may slow releases or shift to hosted-only access until the final text lands. ### Foreign companies built on Chinese models Enterprises that fine-tune or self-host Chinese weights should map their licenses now. If a covered model reaches them through an overseas license, the interim rule may touch the arrangement. Companies that consume the same models through a cloud API sit on safer ground, according to the reported access carve-out. ### Chip, cloud and investor communities Investors in Chinese AI and chip design startups should watch the control-change trigger. A deal that moves actual control to a foreign entity falls inside the registration requirement. Cloud providers that host Chinese models for global customers will watch the wording around deployment technology, since that is where hosting and transfer blur. Beyond that, the move may help domestic Chinese infrastructure vendors. If cluster software and packaging know-how stay inside the country, local suppliers keep a protected position in a market that is growing quickly. ## How China’s Approach Compares With US Controls The comparison to US policy is tempting and partly useful. Washington has leaned on controls over advanced chips and the equipment that makes them, aiming to slow rivals’ access to compute. China’s draft works on the other side of the pipe. It restricts what leaves China: models, data, agents, cluster software and chip design. The US approach restricts inbound capability to a rival. The Chinese approach restricts outbound know-how from its own firms. One protects a lead by denying inputs. The other protects an emerging position by keeping the finished product’s recipe at home. Both treat AI as a strategic asset, and both rely on licensing to keep flexibility. A second contrast sits in the debate over open models. In the US, the argument centers on whether to restrict powerful models for safety reasons. Our coverage of [Anthropic’s letter urging courts to block risky AI](https://shattered.io/anthropic-newsom-letter-courts-block-risky-ai-models-oct-15-2026/) shows one thread of that discussion. China’s draft, by contrast, frames the issue as national security and data sovereignty, with tiers for defense and government use. Neither side has a clean rule for open weights. Once a model’s weights are public, no export regime can pull them back. The Chinese draft tries to catch that moment earlier, at the point of release or transfer, which is why the registration step matters so much. ## Historical Context: A Catalogue Built for Technology Transfer The catalogue itself is not new. It predates the AI boom and has been revised before to reflect fresh priorities. That history explains the draft’s style. The list gives regulators a legal hook for reviewing technology deals, and each update widens the hook to cover new industries. What changes in this version is the subject matter. Earlier revisions dealt mainly with manufacturing and industrial processes. This one reaches into model architecture, training data and agent behavior. That shift mirrors how value has moved. A decade ago the crown jewels were factories. Now they are weights, data and the systems that run them. The wider backdrop is a year of rising friction over AI technology across borders. Our reporting on the [debate over whether China’s open AI is a weapon or armour](https://shattered.io/china-open-weight-ai-weapon-or-armour-cybersecurity-debate-2026/) traced how security officials in several countries began to treat model weights as sensitive. Beijing’s draft is the mirror image, treating its own weights the same way. The takeaway: export control logic, once reserved for missiles and machine tools, now covers software that fits on a hard drive. Enforcement is the open question, because code is easier to copy than a lathe. ## Five Predictions for the Consultation and Beyond These are our judgments, not reported facts. The draft is a draft, and final outcomes depend on choices not yet made. - **Definitions will tighten.** Expect the final text to clarify how parameters are counted and how “world-leading” capability is measured, since vague tests invite uneven enforcement. - **Registration becomes the real chokepoint.** Because it is already in force, the interim rule will shape deals and licenses well before the final catalogue appears. - **Hosted access gains share.** Foreign buyers who want certainty will lean toward API and cloud access to Chinese models instead of self-hosting weights, since that path is the one the draft leaves broadly open. - **Labs publish data lineage.** Companies that can document clean training data will gain a speed advantage, and documentation will turn into a selling point. - **Named-company guidance arrives case by case.** Nothing in the reports names a restricted firm today. If restrictions reach specific products, they will likely surface through individual approvals and denials rather than the catalogue text. Any of these could miss. A tougher final text, a softer one or a long delay would all change the picture. The November 15 deadline is the first checkpoint, and the publication of the final catalogue is the second. ## What Developers and Companies Should Do Now Practical steps are modest and cheap compared with the cost of a compliance surprise. Start with an inventory. List every Chinese model, dataset, chip design or cluster tool your organization licenses, hosts or fine-tunes, and note whether it arrives through download, license or API. Next, flag any arrangement that involves a transfer of ownership abroad, an overseas license or a shift in control to a foreign entity. Those are the three triggers for the interim registration rule. Ask counsel in China and in your home market to review them together, since the rule runs in one jurisdiction while your exposure may sit in another. Then decide whether to comment. The consultation runs until November 15, 2026. Trade bodies and individual firms can file feedback on thresholds, definitions and the treatment of open-weight releases. Silence leaves the drafters to guess at industry needs. Finally, keep a fallback. If a covered model becomes harder to obtain, a substitute from another region should be ready to test. Our [AI and machine learning coverage](https://shattered.io/category/ai-machine-learning/) tracks the alternatives as they ship. ## Frequently Asked Questions **What did China announce on October 16, 2026?** The Ministry of Commerce and the Ministry of Science and Technology released a draft update to the Catalogue of Technologies Prohibited or Restricted from Export. It lists 40 technologies across AI, chips and solar power, and 23 of them are AI-related. **Which AI models would face export restrictions?** The draft proposes restrictions for models with 2 trillion or more parameters or with specified world-leading capabilities. Models trained on core data face a proposed ban, and models trained on important data face proposed restrictions. **Can people outside China still use Chinese AI?** According to reports on the draft, overseas users could generally still reach Chinese AI through websites and cloud platforms. Some transfers of underlying models, data, expertise or deployment technology would require government approval. **Is the pre-export registration rule already in force?** Yes. A temporary pre-export registration requirement took effect on October 16, 2026. It covers the technologies in the draft while classifications stay under consultation. **What triggers registration?** The requirement applies to transfers of ownership abroad, licenses for overseas use, and changes that put ownership or actual control under a foreign entity. **Are DeepSeek, Qwen, Kimi or other named models banned?** The reports reviewed do not establish that any named Chinese AI company, model, executive or product has been restricted under the draft. Treat claims to the contrary as unproven. **When can the public comment?** Public comments are open until November 15, 2026. **Will the final rules match the draft?** Nobody can confirm that yet. The final text, effective classifications and any changes after consultation have not been published.