China's Next Export Ban Is Its Own AI

For four years, “export controls” in Artificial Intelligence (AI) meant one thing: Washington deciding which chips China was allowed to buy. This week the direction of travel reversed. Reuters reported that China’s Ministry of Commerce spent the past month meeting with Alibaba, ByteDance, and the startup Z.ai about restricting overseas access to the country’s most capable AI models, including versions that have not shipped yet.

Nothing has been decided, no draft rule exists, and the officials involved gave Reuters no timeline. But the meetings themselves mark a threshold: Beijing is now treating model weights, the billions of numerical parameters that make a model run, the way it treats rare-earth refining techniques and rocket engines. And the timing is brutal for one specific group. As of this week, eight of the ten most-used AI models on OpenRouter, a neutral marketplace that routes companies’ AI traffic across hundreds of models, are Chinese, and the number one spot belongs to a model made by Xiaomi, a phone company. American companies did not stumble into that dependency. They bought it, eagerly, at a discount of roughly 100-to-1.

Key Takeaways

• Chinese models carried between 30% and 46% of the tokens US enterprises routed through OpenRouter every week since early February 2026, up from 4.5% in the first half of 2025.
• The crossover came in the week of February 9 to 15, 2026, when Chinese models processed 4.12 trillion tokens on OpenRouter and edged past American ones for the first time.
• The price gap this dependency is built on runs roughly 100-to-1 on a typical workload, and discounts that size are policies, not prices.
• Beijing has put a technology this popular under export control before, over a single weekend in 2020. What that precedent can and cannot repeat against open model weights is the question the coverage has skipped.

Eight of the Ten Most-Used Models Are Chinese

Start with the dependency, because most coverage of the Reuters story skipped it. OpenRouter is a routing marketplace that lets developers send requests to hundreds of AI models through one Application Programming Interface (API), which makes its public traffic data one of the few neutral windows into what companies actually run. A CNBC analysis of that data found Chinese models accounted for at least 30% of US enterprise token volume in every week since early February 2026, peaking at 46%. A year earlier, in the first half of 2025, the figure was 4.5%.

The crossover moment came in the week of February 9 to 15, 2026, when Chinese models processed 4.12 trillion tokens on the platform and edged past American ones for the first time. A token is the unit AI models bill by, roughly three-quarters of a word.

The live leaderboard makes the intro’s eight-of-ten figure concrete. As of the week of July 11, 2026, Xiaomi’s MiMo-V2.5 sits at number one with 5.4 trillion tokens this week, followed by DeepSeek V4 Flash, Tencent’s Hy3, MiniMax M3, Z.ai’s GLM 5.2, DeepSeek V4 Pro, and Stepfun’s Step 3.7 Flash. The only US entries in the top ten are two versions of Anthropic’s Claude Opus and an open model from Nvidia. Add the volumes and the Chinese models in that top ten carry about 27 trillion weekly tokens against roughly 6 trillion for the American ones.

Silicon Valley noticed before Washington did. Models from DeepSeek and Moonshot AI are, in the words of one report on the Journal’s findings, “highly popular in Silicon Valley, becoming the core of daily operations for companies of all sizes there.”

The 100-to-1 Subsidy

Why did American companies hand nearly half their token traffic to models from a strategic rival? Because the price sheet reads like a misprint. OpenRouter’s published pricing on July 11, 2026 lists DeepSeek V4 Flash, the second most-used model on the platform, at $0.08 per million input tokens and $0.15 per million output tokens. Anthropic’s Claude Opus 4.8 lists at $5.00 and $25.00, and the flagship sol tier of OpenAI’s GPT-5.6, added to the platform’s price sheet this same week, lists at $5.00 and $30.00.

Take an illustrative production workload of 750 million input tokens and 250 million output tokens a month and run it down the price sheet:

The same job costs $97.50 on the Chinese flagship’s fast tier and $10,000 on the American one:

$$\frac{750 \times 5.00 + 250 \times 25.00}{750 \times 0.08 + 250 \times 0.15} = \frac{10{,}000}{97.50} \approx 103$$

And that anchor is the conservative one. Anthropic’s newest model, Claude Fable 5, lists at $10.00 in and $50.00 out, which prices the identical workload at $20,000 a month, a gap of roughly 205-to-1 against the Chinese fast tier.

The frontier US models are better, and that matters for the hardest tasks. But research group Epoch AI, which tracks model capability over time, finds Chinese models have lagged the US frontier by an average of seven months since 2023, with the gap ranging between four and fourteen months. For a company summarizing support tickets or extracting fields from invoices, seven months behind the frontier at one percent of the price is not a compromise. It is the obvious call. The site’s earlier look at DeepSeek and the efficiency gap traced how that economics took shape; what is new in 2026 is the scale of the American reliance on it.

That is what makes this week’s news land differently. The discount was never just a price. It was a policy, and policies can be revoked.

Beijing Has Run This Play Before

If restricting the export of software your own champions give away sounds far-fetched, recall August 28, 2020. That day China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology revised the country’s Catalogue of Technologies Subject to Export Bans or Restrictions, adding five AI technologies to the restricted list, among them “data analysis-based personalized recommendation pushes.” The category was written broadly, but observers immediately connected it to one company: ByteDance, then under a Trump-administration deadline to sell TikTok.

The effect was that TikTok’s recommendation algorithm, the asset any buyer actually wanted, could no longer leave China without an export license from Beijing. The arrangement ByteDance had forged with Oracle and Walmart “fell apart after the commerce ministry and the Ministry of Science and Technology revised the export control list, adding two recommendation technologies relevant to TikTok’s operations,” as the South China Morning Post later put it, and five years on the algorithm remained the sticking point in the deal talks that followed. Beijing had discovered that an export catalogue entry is a veto it can write for itself, quietly, over a weekend.

Model weights are a natural next entry. The same instrument, a technology catalogue revision, would not need a new law or a public confrontation. One panel of Chinese legal scholars has already floated a tiered scheme to Reuters: a light filing requirement for basic tools, security reviews for stronger models, and a domestic-only lockdown for the most sensitive systems. Also on the table, per the same reporting: treating theft of proprietary AI technology as a national-security crime, and limits on which investors can fund Chinese AI labs.

A Curtain That Only Falls Forward

Here is where the story stops being a simple rug-pull threat, and the objection every open-source developer is typing right now deserves to win some ground. You cannot recall a published model. DeepSeek V4, GLM 5.2, Kimi K2.6, and Qwen’s current line are open-weight: their parameter files are on Hugging Face and mirrored on corporate servers across the planet. No catalogue revision reaches into an American data center and deletes a file. Time’s reporting on the Reuters story drew the same conclusion: because circulating weights cannot be retrieved or retrofitted with safeguards, any restrictions would in practice apply only to future models.

So the honest version of the threat is not “your AI vendor disappears tomorrow.” It is quieter and slower. Companies that standardized on Chinese models would keep the models they have and lose the upgrade path. The weights on disk stop improving the day the curtain drops, while the frontier keeps moving. Epoch’s seven-month average gap is measured between a moving US frontier and a moving Chinese one; freeze the Chinese side at its last public release and the gap stops being seven months and starts being forever.

The likely response is already visible in outline: a “frozen fork” economy. Enterprises and cloud providers would fine-tune, distill, and squeeze the last released generation of Chinese weights for years, the way industries kept flying old airframes under sanctions. That works until the task changes, a security hole needs the original lab, or a competitor on a live frontier model simply does the job better.

And it is worth stating the boring explanation plainly, because it is probably the true one. This does not require a Politburo master plan to punish America. Chinese labs have watched their most valuable intellectual property walk out the door for free while Z.ai’s GLM 5.2 claims parity with US frontier systems on tasks as sensitive as software vulnerability discovery. A state applying its existing export-control framework to its newest strategic technology, with labs themselves asking for protection from theft, is the least surprising event in trade policy. The 2020 catalogue already lists voice recognition and recommendation engines; nobody should be shocked if model weights join them.

Washington Is Pulling the Same Cord

The strangest feature of this story is that both governments want the same wire cut. While Beijing debates limiting what leaves China, Washington has been moving to limit what enters corporate America, with proposals to push Chinese models out of US enterprises colliding with the awkward fact that freely downloadable weights are nearly impossible to ban in practice.

The US side has already crossed its own version of the threshold. This summer Washington forced Anthropic’s newest frontier model offline within days of its launch, a saga covered in the site’s report on the suspension, and Time notes that access restrictions on Anthropic’s Mythos and OpenAI’s GPT-5.6 were justified by their skill at finding software vulnerabilities. Both capitals, in other words, have concluded within the same year that frontier model access is a national-security lever. The disagreement is only about whose hand is on it.

For the American companies in the middle, the incentives now point in an uncomfortable direction. If Beijing restricts future releases, the 100-to-1 discount evaporates for every new capability, and pricing power flows straight back to OpenAI and Anthropic, whose margins the cheap Chinese tier has been compressing. The dynamic behind why cheaper AI keeps producing bigger bills would get a geopolitical accelerant. Alibaba’s incentive cuts the other way: free Qwen adoption is a funnel for the cloud business built around it, and a lockdown would amputate the overseas half of that funnel. That tension, between labs that profit from openness and a state that increasingly sees openness as leakage, is the fault line these meetings exposed.

Watch the Catalogue, Not the Announcement

The counterargument that deserves the most respect is the simplest: “weighs” is not “will.” Reuters’ own sources stressed that nothing is decided, curbs might touch only future models, and no timeline exists. China has strong reasons not to pull the trigger, since global adoption of its models is the best soft-power win its tech sector has ever had. The scenario where these meetings end in a filing requirement and nothing else is entirely live.

But the mechanism to go further already exists, it has been used against a much bigger political football than an API, and the paper trail is public. China’s export catalogue is a published document; its 2020 revision added five AI entries in a single stroke. The tell will not be a press conference. It will be a new line item in a PDF, or a release note that quietly says GLM-6 and DeepSeek V5 are domestic-deployment only.

Until then, every US company running its stack on $0.08 tokens is enjoying the cheapest intelligence money can buy, subsidized by a rival government’s openness strategy, revocable by catalogue amendment. The models already downloaded are theirs forever. The next ones were never promised.