U.S.-China Relations
NextFin News - China has proposed adding advanced artificial-intelligence technologies to its export-control list, a move that would let Beijing restrict overseas transfers of AI models, model weights, chip designs and computing clusters in a direct echo of Washington's own technology-denial campaign. The draft proposal, released October 16, 2026, covers 23 AI technologies — including large language models, autonomous agents, chip designs and computing clusters — and would place model weights and source code for systems above certain performance thresholds under licensing control. The question the market has not fully priced: after four years of trying to wall off Chinese AI, the United States may now face a world in which the open-weight models that power everything from startups' prototypes to enterprise agents can be pulled back behind a licensing gate on Beijing's say-so.
The Proposal: 23 AI Technologies Under a Licensing Gate
The draft measures mark the most concrete step yet in a policy shift that has been building since June 2026, when Chinese authorities began closed-door consultations with domestic AI laboratories — including Alibaba Group, ByteDance and Zhipu AI — over how to limit overseas access to the country's most advanced models. The options discussed ranged from restricting open-weight downloads and criminalizing unauthorized model distillation to blocking foreign fabrication of Chinese chip designs and curbing cross-border transfers of training data.
What changed between those consultations and the October 16 proposal is the form. Rather than applying ad hoc pressure on individual firms, Beijing is moving to codify control through its export-control catalogue — the same legal instrument that governs China's dual-use technology transfers and has anchored its responses in the semiconductor dispute. Under the draft, restricted technologies would include the model weights and source code behind products that exceed defined performance thresholds, turning the intangible assets of AI development — the trained parameters, the code, the designs — into licensable exports.
The scope matters. Large language models are the obvious target, but the inclusion of autonomous agents and computing clusters signals that regulators are thinking about AI as a stack, not a product. An agent is a model plus the ability to act on its own across multiple steps; a cluster is the hardware that trains and runs it. Controlling all three layers at once would give Beijing leverage over the entire value chain, from research to deployment. The draft also contemplates restrictions on overseas acquisitions of Chinese AI businesses holding strategically important technology, extending the export-control logic into the investment channel.
One important caveat: this is a proposal, not a final rule. Chinese regulators are still gathering feedback from the companies that would be affected, and the performance thresholds that determine which models fall inside the gate have not been published. The earliest real signal will not be a dramatic overnight ban — it will be how the major Chinese labs ship their next frontier releases.
Why Beijing Is Flipping the Script
For four years, the story of AI export controls ran in one direction. Washington restricted what China could access: advanced chips first, then chipmaking equipment, then, increasingly, the models themselves. The Biden administration's AI Diffusion Rule created a global licensing requirement for advanced AI chips and model weights; the Trump administration rescinded it in May 2025 but promised a replacement, and in January 2026 the Commerce Department shifted H200-class chip exports to China from a presumption of denial to a case-by-case review tied to strict supply, testing and end-user certifications — a calibrated valve, not an open door.
China's countermove is asymmetrical. It cannot deny Chinese firms access to Nvidia's best silicon — Washington controls that lever. What it can do is deny the world access to the one area where Chinese labs have built genuine global pull: open-weight models. The numbers behind that pull are large enough to make the threat credible. Qwen, Alibaba's model family, recorded about 2.045 billion downloads on the Hugging Face Hub in 2026, compared with about 418 million for Google's models and 227 million for Meta's, according to a Hugging Face state-of-open-models report. On the Hub alone, developers have created more than 151,000 models derived from Qwen — 2.6 times the number derived from Meta models and 4.7 times the repositories associated with the Llama family.
The usage data tells the same story. OpenRouter's token-routing figures show Chinese open-weight models moving from negligible to a majority share of all tokens processed between late 2024 and mid-2026, an inflection that traces directly to DeepSeek-R1's January 2025 release and the subsequent flood of Qwen, Kimi, GLM and MiniMax models. DeepSeek's share of token usage on Vercel rose from under 1 percent to roughly 17 percent by May 2026, even as its revenue share stayed near 1 percent — the clearest illustration of how cheap the open tier has made intelligence.
This is not a market China can afford to surrender lightly. But the calculus in Beijing is that the strategic value of keeping frontier capability domestic outweighs the soft-power value of open releases. If advanced AI is a critical national asset requiring controls, then China's most capable future models belong at home.
The Mechanism: Why Controlling Weights Is Different From Controlling Chips
Chip controls work because fabrication is concentrated and traceable. A high-end AI accelerator is a physical object that passes through customs, carries a serial number, and is made by a handful of foundries — TSMC, Samsung, Intel — all of which sit inside the reach of U.S. foreign-direct-product rules. Weights are different. They are files. They can be copied, split, encrypted, mirrored on a hundred servers, and carried across borders on a hard drive. Once released, they are effectively un-recallable.
That asymmetry shapes what the Chinese proposal can and cannot do. A licensing regime can constrain future releases — the next Qwen, the next DeepSeek, the next Kimi — and it can impose legal liability on the labs that distribute them. It cannot un-release what is already out. As of August 2026, the open downloads on the global market span DeepSeek V4, GLM-5.3, Kimi K3, MiniMax M3, and the smaller Qwen and Ernie families; Baidu's Ernie 5.1, ByteDance's Doubao, and the multimodal flagship Qwen3.8-Max remain API-only. Any rule Beijing writes will therefore draw a line between the already-distributed past and the controlled future, and enforcement will fall hardest on the frontier models that have not yet shipped.
The second mechanism is distillation. Even if weights are gated, a sufficiently capable API-accessible model can be used to train a smaller domestic copy — the technique U.S. officials have called adversarial distillation and addressed in an April 2026 White House memorandum. China's reported interest in criminalizing unauthorized distillation is the mirror image of that concern: an attempt to stop foreign actors from compressing Chinese capability into locally runnable models. The legal and technical difficulty is defining what counts as distillation versus independent training, a line that no regulator has yet drawn cleanly.
The third mechanism is the hardware stack that travels with Chinese models. U.S. guidance has treated the use of certain Huawei Ascend chips anywhere in the world as a presumptive violation, and May 2026 guidance extended license requirements for advanced computing items to entities with Chinese ultimate parents wherever they are located. A Chinese export-control regime that covers chip designs would complete the mirror: the same model, the same weights, the same silicon, governed by two opposing licensing authorities.
Second-Order Effect: Two Licensing Regimes, One Fractured Stack
The first-order effect of the proposal is obvious: Chinese labs lose some freedom to publish. The second-order effect is what should worry global developers and enterprises: the AI software stack is splitting into two licensing jurisdictions, and neither side is moving toward openness.
On the U.S. side, the Commerce Department has already shown it will control access to frontier models as such. In June 2026, Anthropic suspended foreign access to its most capable systems — Claude Fable 5 and Claude Mythos 5 — to comply with a Commerce Department export-control directive that barred access by any foreign national, the first time Washington controlled a model rather than its weights as an item. On the Chinese side, this proposal would create a parallel gate for Chinese-origin models, weights and designs. A developer in Berlin, Bangalore or Buenos Aires building on open-weight models would need to ask two questions instead of one: is this model U.S.-controlled, and is it China-controlled?
The compliance cost of that duplication falls disproportionately on smaller players. Large enterprises can afford export-control counsel and dual-stack architectures. Startups and independent developers — the very constituency that open-weight models unlocked — absorb the uncertainty or exit. That is the irony at the center of this escalation: both superpowers are using national-security logic to close the ecosystem that made the current AI boom possible, and the open layer that sits between their two walled gardens is the thing being paved over.
There is also a market-structure consequence. If frontier Chinese models become licensable exports rather than public releases, the competitive pressure they exert on U.S. labs diminishes. DeepSeek's January 2025 release proved a Chinese lab could match Western reasoning quality at a fraction of the cost, and the subsequent flood of Qwen, Kimi, GLM and MiniMax releases kept that pressure on. Gate the next generation, and U.S. frontier labs face less price competition — a quiet benefit to incumbents that no regulator on either side has stated aloud.
The Counter-Thesis: China Needs the World More Than the World Needs Chinese Models
The strongest argument against reading this proposal as a market-moving regime shift is simple: China's AI ambitions still depend on the global ecosystem it is threatening to fracture. Chinese labs train on stockpiled Nvidia A100 and H800 chips acquired before the tightest controls; their most advanced domestic accelerators still lag the frontier; and their global influence runs through the very platforms — Hugging Face, GitHub, international cloud marketplaces — that a licensing regime would complicate. Industry participants in China have raised concerns that strict controls could weaken the international reach and competitiveness of Chinese AI products.
More pointedly, the open-weight strategy is what made DeepSeek, Qwen, GLM and Kimi global players in the first place. Walking it back risks trading real, measurable influence — billions of downloads, hundreds of thousands of derivative models, majority token share on routing platforms — for a symbolic assertion of control. Analysts who view the proposal as mostly signaling argue that Beijing will calibrate the final thresholds to keep the popular mid-tier models flowing while gating only the true frontier systems, preserving influence without surrendering the crown jewels.
That counter-thesis is plausible, and it is the base case for why this should not be read as an overnight ban. But it misses the directional point. Even a calibrated gate changes the default from "publish unless prohibited" to "license before publishing," and that change in default is what shapes lab behavior. The strongest evidence that the counter-thesis is wrong would be the next frontier release from Alibaba, Zhipu or DeepSeek appearing globally, immediately, with downloadable weights, on the same cadence as 2025 and early 2026. If that happens, the gate is mostly theater. If the next generation ships API-only, China-region-only, or not at all, the regime shift is real.
What to Watch Next
Three signals will separate signaling from substance. First, the final wording of the export-control catalogue and the performance thresholds that define the gate — the difference between a rule that captures only the largest frontier systems and one that sweeps in mid-tier models is the difference between a narrow fence and a wall. Second, the behavior of the major labs at their next release cycle: global and open, or domestic and gated. Third, the U.S. response — Washington has spent years building the architecture of AI export controls, and a Chinese mirror regime gives it both a pretext and a pressure point for tightening its own rules further.
The time horizons point in different directions. In the short term, expect uncertainty and possibly a rush of releases as labs front-run the final rule. In the medium term, expect bifurcation: Chinese frontier models distributed through licensed channels, U.S. frontier models controlled through Commerce Department access rules, and the open middle thinning out. In the long term, this is structural, not cyclical — a fragmentation of the AI development stack into two licensing jurisdictions that will not re-merge on its own, because the driver is not a temporary policy dispute but a durable shift in how both superpowers classify advanced AI: as a tradable technology, or as a national asset.
Base case: Beijing finalizes a calibrated rule that gates only frontier-tier models while keeping mid-tier open weights flowing, preserving global influence while retaining a national-security lever. Upside case for openness: industry pushback during the consultation period produces thresholds high enough that most current models remain unaffected, and the next release cycle proceeds largely as before. Downside case: the thresholds land low, distillation restrictions carry criminal liability, and the first gated frontier release triggers reciprocal U.S. tightening — at which point the two licensing regimes harden into a durable split.
The draft proposal covers 23 AI technologies, including large language models, agents, chip designs and computing clusters. Restricted technologies would include model weights and source code for products exceeding certain performance thresholds.
The bottom line: China's proposal is the clearest signal yet that the era of open, borderless AI development is closing on both sides of the Pacific. The market has priced a technology cold war in chips. It has not fully priced a cold war in the models themselves — and that is where the second-order risk sits.
Explore more exclusive insights at nextfin.ai.
