Your China Brand Problem Just Went Global: Chinese Open-Weight Models Now Answer Most of the World's Routed AI Traffic (2026)

In June 2025, US models — OpenAI, Google, and Anthropic combined — accounted for roughly 70% of tokens routed through OpenRouter, the largest neutral marketplace for AI model access. By June 2026, that share had fallen to about 30%. DeepSeek is now OpenRouter's single largest vendor at 17.6% of weekly routed tokens, with Alibaba's Qwen next at 13.9%.

For brand marketers, the instinct is to file this under "AI infrastructure news" and move on. That would be a mistake. The models that decide whether your brand appears in a Doubao answer in Shanghai are increasingly the same models running inside Western SaaS products, agent frameworks, and internal enterprise tools. Whatever those weights believe about your brand — accurate or not, flattering or not — no longer stays inside China's borders.

That reframes the entire business case for China GEO. It stops being a market-entry line item and becomes a question about your brand's representation in the global model layer.

The number that changed

The crossover happened faster than almost anyone forecast. During the week of February 9–15, 2026, Chinese models passed US models on OpenRouter for the first time, at 4.12 trillion weekly tokens. By June 2026, weekly usage of the top nine models showed Chinese models at roughly 18 trillion tokens against about 5.5 trillion for US models. Chinese domestic reporting citing OpenRouter data put total weekly Chinese-model calls at 40.48 trillion tokens for the week of August 17–23, 2026 — the seventeenth consecutive week leading globally.

The driver is not nationalism or novelty. It is price. Open-weight Chinese models run 60% to 90% cheaper than flagship Anthropic and OpenAI offerings, and the capability gap has narrowed to the point where the discount is hard to justify passing up for high-volume, low-margin workloads: summarization, classification, extraction, first-draft generation, routing, and the long tail of background inference that no end user ever sees labeled.

That last category matters most for brands. A product that says "powered by AI" almost never tells the user which model produced the answer. A Western marketing tool that recommends competitor sets, a procurement assistant that shortlists vendors, a travel app that suggests hotels — any of these may be running Qwen or DeepSeek weights behind an unbranded API call.

What actually travels: weights, not citations

Here is the distinction most GEO coverage misses, and it determines what you should actually do about this.

Brand visibility inside a Chinese AI app is produced by two layers stacked on top of each other:

The retrieval layer. When a user asks Doubao about skincare, the app queries a live index — Douyin content, SMZDM reviews, Baidu results, Xiaohongshu-adjacent sources, brand and OTA pages — and feeds fresh documents into the prompt. This is the layer most GEO work targets, and it is the layer that responds fastest to intervention.

The parametric layer. Underneath sits what the model absorbed during pretraining: which brands exist, which category they belong to, which tier they occupy, what adjectives cluster around them, whether they are described as premium or mid-market, domestic or imported, trustworthy or contested. This is baked into the weights and cannot be edited by publishing a press release next Tuesday.

When Qwen or DeepSeek weights get deployed on OpenRouter, downloaded from Hugging Face, or self-hosted inside a European company's VPC, only the second layer travels. There is no Douyin index attached. No SMZDM. No Baidu. The Chinese retrieval infrastructure stays in China.

What ships worldwide is the parametric brand prior — the model's baseline, no-context belief about your brand, learned largely from Chinese-language web text.

This cuts in two directions, and both are worth taking seriously:

Most international brands doing China GEO work today are optimizing the top row. The bottom row is where the new exposure sits.

Why "zero-score" brands should be most worried

Our tracking of international brands across six Chinese models — Doubao, Kimi, DeepSeek, Qwen, Wenxin, and Hunyuan — has consistently found a hard bimodal split. Category leaders cluster near 100. A meaningful set of well-known global brands score at or near zero: recognized in their home markets, essentially absent from Chinese model outputs.

The standard reading of a zero score has been "you have no China presence, and Chinese consumers won't find you." That is still true. But the second-order reading is newer and worse: a near-zero parametric prior in a model family that now handles a large share of global routed inference means the weakness is no longer geographically contained.

Concretely, if Qwen weights hold a thin, stale, or category-confused representation of your brand, that representation is what a US-built agent gets when it calls Qwen with no retrieval attached and asks it to name the leading brands in your category. Your competitor with deep Chinese-language coverage — often a domestic Chinese brand, sometimes a Western rival that started China content work three years earlier — gets named instead. Nobody in your organization sees this happen, because the interface says "AI recommendations," not "Qwen3."

There is a related governance dimension. Coverage of Chinese-model adoption in Western enterprises has flagged data-residency and content-provenance concerns, and some organizations are running procurement reviews on exactly this basis. Brand teams should expect the question "what do Chinese models say about us?" to arrive from legal or comms before it arrives from marketing.

Where the priors come from

If parametric priors are the exposure, the practical question is what forms them. Three inputs dominate, and none of them are your owned website:

Chinese-language encyclopedic and reference text. Baidu Baike, Wikipedia's Chinese edition, and Sogou Baike carry disproportionate weight in shaping the model's baseline entity description. We have documented before that these three sources frequently disagree with each other on basic brand facts — founding year, ownership, category, market positioning — and models trained on all three inherit the disagreement.

High-volume Chinese-language commercial and editorial content. Aggregated review platforms, e-commerce category pages, trade press, and vertical media. Volume and consistency matter more than any single placement.

Long-lived user-generated discussion. Forum and social content that persists long enough to be crawled repeatedly. Ephemeral campaign content contributes almost nothing here.

The uncomfortable implication is that parametric priors move on a training-cycle timescale, not a campaign timescale. Content published today can influence the weights of a model released next year. That is a slower loop than most brand teams are structured to run — and precisely why the brands that started early hold a lead that is difficult to buy back.

The measurement gap

Nearly all commercial GEO measurement, ours included, tests brands through consumer app interfaces — the full retrieval-plus-generation stack. That is correct for measuring what Chinese consumers experience.

It does not measure the parametric layer. To see that, you have to query the raw model with retrieval disabled and ask category questions cold. The gap between the two scores is diagnostic: a brand scoring 85 in-app and 20 raw is renting its visibility from the retrieval layer. A brand scoring 60 in-app and 55 raw owns it.

We do not yet publish this split, and to our knowledge no vendor does. Brand teams can approximate it today by running the open-weight versions of Qwen and DeepSeek through any inference provider with web search turned off and comparing the outputs to their in-app scores.

Takeaway for Brand Marketers

1. Stop scoping China GEO as a market-entry project. If Chinese open-weight models handle a large and growing share of global routed inference, your Chinese-language brand footprint is now an input to worldwide AI outputs. Budget it accordingly — this is a reasonable argument to bring to a CMO who has previously declined China GEO on market-size grounds.

2. Split your measurement. Ask your vendor, or run yourself, a retrieval-off test against raw Qwen and DeepSeek weights. The delta against your in-app score tells you whether your visibility is owned or rented.

3. Prioritize durable, factual, Chinese-language reference content over campaign bursts. Encyclopedic consistency, category pages, and long-lived editorial coverage are what enter training corpora. A three-week campaign will not.

4. Fix entity disagreements first. Where Baidu Baike, Chinese Wikipedia, and Sogou Baike contradict each other about your brand, every future model generation inherits the confusion. This is the highest-leverage, lowest-cost item on the list.

5. Treat a zero score as a global risk, not a China risk. If your brand is absent from Chinese model outputs, assume that absence now propagates into products your customers use in markets where you do have share.

The unglamorous version of this: the cheapest inference in the world is currently Chinese, the market has noticed, and the models doing that work carry a view of your brand that was formed by Chinese-language text you probably never commissioned. That view is now in production, globally, and it updates on a timescale measured in model releases.

Related: See our brand visibility scores across all six Chinese AI models, and our earlier analysis of why brand facts diverge across Chinese AI models.