--- title: "Chinese AI models rose from 6–13% to 57–67% of OpenRouter tokens" summary: "Chinese models' share of tokens on the OpenRouter gateway by customer region, February 2025 to September 2026, checked against OpenRouter's own weekly open-weight series." url: https://www.thetechbag.com/blog/chinese-open-weight-models-gateway-traffic format: Trend desk: "AI in the enterprise" published: 2026-09-28 facts_checked: 2026-09-28 author: TechBag (The TechBag Journal) tags: ["OpenRouter","Vercel","open-weight models","DeepSeek","Z.ai","GLM-5.3","Qwen","AI gateway","token pricing","data residency","model licences","DPDP Act"] cover: https://www.thetechbag.com/blog/covers/chinese-open-weight-models-gateway-traffic.png --- # Chinese AI models rose from 6–13% to 57–67% of OpenRouter tokens *Chinese models' share of tokens on the OpenRouter gateway by customer region, February 2025 to September 2026, checked against OpenRouter's own weekly open-weight series.* Trend · AI in the enterprise · By TechBag · Published 2026-09-28 · Facts checked 2026-09-28 · Canonical: https://www.thetechbag.com/blog/chinese-open-weight-models-gateway-traffic The short version - On OpenRouter, an AI-model gateway, Chinese models carried 57–67% of the tokens its business customers used in the week of 14 September, against 6–13% in February 2025, CNBC reported from OpenRouter’s data.[1] - OpenRouter’s own public series agrees on the direction: open-weight models rose from 26% of its tokens in November 2025 to 70% in September, and named Chinese labs to at least 57%.[3][4] - The people who run the gateways name price first. Chinese models took about half or more of tokens but a quarter to a third of spend, which leaves buyers with three questions: what it costs, where it runs, and under which licence.[1] Chinese AI models went from a small slice of the traffic on two large model gateways to most of it in 2026. A gateway puts many companies’ models behind one account and one programming interface: a developer picks a model, and the gateway passes the request to a company that runs it. OpenRouter is one; Vercel’s AI gateway is another. Usage is counted in tokens, the word-fragments a model reads and writes. Many of the Chinese models in that traffic are **open-weight**: the trained model (its “weights”) is published for anyone to download and run, under a licence. That is narrower than open source; CNBC describes such models as making “different parts of an AI model available for developers to inspect, use and sometimes modify”, unlike closed models whose inner workings stay proprietary.[2] ## The line: from a tenth to half or more CNBC reported on 26 September that Chinese models accounted for 57–67% of OpenRouter tokens in the week of 14 September, up from 6–13% in February. The range covers three groups of OpenRouter’s business customers: the US, Europe, and what OpenRouter calls the “Global South”, 82 countries in Central and South America, Africa and Asia. The chart CNBC published with the article shows that the February figure is February 2025, and it ends at the week of 7 September 2026, when the share was 68% in the Global South, 50% in the US and 48% in Europe. About half of OpenRouter’s tokens are used by US companies.[1] On Vercel’s gateway, the Chinese share rose from 11% in January to 55% in August.[1] OpenRouter’s own public weekly series, which we pulled, show the same climb, and the total under it grew from about 5.7 trillion tokens a week at the start of November 2025 to about 132 trillion in the week of 21 September.[3][4] **Figure: On OpenRouter's own figures, open-weight models went from 26% to 70% of tokens, and Chinese labs to at least 57%** *Share of weekly tokens on OpenRouter, %, token-weighted by month, November 2025 to September 2026 (weeks beginning 7, 14 and 21 September)* | | Open-weight | Chinese labs | | --- | --- | --- | | Nov 2025 | 26.1 | 16.8 | | Dec 2025 | 35.5 | 19.1 | | Jan 2026 | 36.5 | 22.8 | | Feb 2026 | 52 | 36.3 | | Mar 2026 | 41 | 45.7 | | Apr 2026 | 46.7 | 39 | | May 2026 | 57.6 | 43.1 | | Jun 2026 | 67.4 | 49.4 | | Jul 2026 | 74 | 56.8 | | Aug 2026 | 70.5 | 53.8 | | Sep 2026 | 69.6 | 57.3 | Source: OpenRouter rankings: open-weight vs closed-weight tokens by week, and tokens by model author by week; TechBag calculation · as of 27 September 2026 · Chinese labs = DeepSeek, Z.ai, Tencent, Xiaomi, Qwen, MiniMax, Moonshot AI and StepFun when named among a week's top nine authors, so a lower bound; stealth models excluded TechBag countWhat the gateway's numbers do and do not show In the week of 14 September, the named Chinese labs held 59.5% of OpenRouter’s tokens by our count, inside CNBC’s range; a week later DeepSeek alone held 25.1%, Z.ai 16.0%, Tencent 9.3% and Xiaomi 7.3%. “Chinese” and “open-weight” are not the same set: an open-weight model can come from any country, and the Chinese line counts only the top nine makers each week. OpenRouter says its rankings show traffic routed through OpenRouter and measure neither users nor spend; a model that writes longer answers counts for more tokens.[5] This is one gateway’s customers, developers who build on many models, not the whole market: companies that buy directly from OpenAI, Anthropic, Google or a cloud provider do not appear in it. One number does not line up. In July CNBC put the US weekly peak since 8 February 2026 at 46%; its September series puts several June weeks for the US above 55%.[2][1] Neither article explains why, so we quote CNBC’s September figures as reported. ## What drives it: price first, then licence and quality **Price.** “Chinese models are becoming capable enough for more tasks at a much lower cost. Once a model meets the quality bar for the job, that price difference becomes compelling,” Harpreet Arora, head of agentic infrastructure at Vercel, told CNBC; he added that companies still want US frontier models for more complicated tasks.[1] In July, Justin Summerville of OpenRouter put open Chinese models at “60% to 90% cheaper” than the leading Anthropic and OpenAI models.[2] The spend data agree: in the week of 7 September, Chinese models took 48–68% of tokens by region but 26–33% of what customers spent.[1] **Figure: Chinese models took about half or more of OpenRouter tokens by early September, but a quarter to a third of the spend** *Chinese models' share of OpenRouter tokens and of spend, %, by customer region, weeks of 10 February 2025 and 7 September 2026* | | Tokens, Feb 2025 | Tokens, Sep 2026 | Spend, Feb 2025 | Spend, Sep 2026 | | --- | --- | --- | --- | --- | | Global South | 12.9 | 68.3 | 7.8 | 33.3 | | US | 6.4 | 50.1 | 7.2 | 25.6 | | Europe | 9 | 48.3 | 7.9 | 31.7 | Source: OpenRouter data published by CNBC (charts in the article of 26 September 2026) · as of 26 September 2026 · Global South = 82 countries in Central and South America, Africa and Asia, as OpenRouter defines it TechBag countWhat that gap implies per token Dividing each side’s share of spend by its share of tokens, a Chinese-model token cost on average about a quarter of a non-Chinese one for Global South customers, about a third for US customers and about a half for European customers that week. It is an average across every model and task, not a price list. **Licence and control.** “We’re seeing companies increasingly motivated to turn to cheaper AI stacks they can control and adapt themselves,” Yacine Jernite, head of machine learning at Hugging Face, told CNBC in July.[2] DeepSeek published V4.1 Flash on 10 September under the MIT licence.[6] Z.ai’s GLM-5.3 licence allows commercial use but requires a firm running a model-as-a-service business with more than US$10bn of revenue in 12 months to pass Z.ai’s security review first.[7] MiniMax and Moonshot AI publish their current models under licences of their own.[8] **Quality.** Peter Walker, head of insights at OpenRouter, said Chinese open models released this year “can credibly perform in advanced agentic use cases, especially in regards to coding, in a way that was just not true in late 2025.”[1] Kyle Chan of Brookings put them “six to nine months” behind in July.[2] The other side of the argument is in the same reports: Daniel Remler of the Center for a New American Security called Chinese AI “real economic and security risks for the United States”, and two US House committees are investigating its adoption.[1] ## What it means when a team picks a model - **Cost.** The spend gap is the case for testing a cheaper model per task; the same reports say teams still send harder work to frontier models.[1] Price per token is not cost per task: a model that needs more tokens or more retries to finish can cost more. - **Where the data goes.** Where a model was built and where a request is processed are separate questions. On a gateway, one open-weight model can be served by several hosting companies; OpenRouter lets a customer exclude named hosts, route only to hosts that do not store prompts, or require zero data retention.[9] Those settings are off unless someone turns them on. - **Governance.** Keep the model’s name and version, its licence, the host and the host’s retention terms on file; each can change on its own, and so can the rules, as Washington’s inquiry shows.[1] If you are buying this in India We found no Indian figure: CNBC did not say whether India is in the “Global South” group, and neither OpenRouter series splits by country. The Indian rule that bites is about the data, not the model’s origin. Section 16 of the Digital Personal Data Protection Act, 2023 lets the Central Government restrict transfers of personal data to countries it notifies,[10] and the DPDP Rules, notified on 13 November 2025, bring the main duties into force on 13 May 2027.[11] A prompt that carries customer data goes wherever its host runs, so ask the gateway for the host’s country before the data leaves. **What to watch.** OpenRouter’s rankings update daily, and the week beginning 28 September closes on 4 October: that is the next point on both of its lines.[5] The open question is whether the US share holds near half after OpenAI and Anthropic both announced cheaper models in the week of 22 September,[1] and what the House committees conclude. How we did this Figures as of 28 September 2026 Regional token and spend shares are OpenRouter’s data as CNBC published them on 26 September, in the article’s text and the two charts embedded in it (last updated 25 September); we read the February 2025 and 7 September 2026 weeks from those charts. CNBC’s text gives the week of 14 September as 57–67%; its chart ends a week earlier. Our re-pull uses OpenRouter’s weekly rankings series for weeks beginning 3 November 2025 to 21 September 2026 (Monday-start weeks, UTC). Monthly values add the tokens of every week beginning in that month. “Chinese labs” counts only makers named among a week’s top nine; the rest sit in “Others”, so the line is a floor. Stealth (unnamed) models are left out of both lines. OpenRouter can revise past weeks. The implied per-token comparison is our arithmetic on CNBC’s regional shares. Views on risk and quality are attributed to the people who gave them; we take no side on them. Spotted an error? Write to us; corrections are published at the foot of the piece. Sources 1. [CNBC, Kai Nicol-Schwarz: Chinese AI models surge in global popularity, and Washington is worried (with its OpenRouter and Vercel charts)](https://www.cnbc.com/2026/09/26/china-ai-global-adoption.html) — 26 September 2026 2. [CNBC, Kai Nicol-Schwarz: Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge](https://www.cnbc.com/2026/07/07/chinese-ai-models-costs-us-openai-anthropic.html) — 7 July 2026 3. [OpenRouter rankings: share of tokens processed by open-weight vs closed-weight models, by week](https://openrouter.ai/rankings) — weekly series, weeks beginning 3 November 2025 to 21 September 2026 4. [OpenRouter rankings: tokens by model author, by week](https://openrouter.ai/api/frontend/v1/rankings/author-market-share) — weekly series, weeks beginning 3 November 2025 to 21 September 2026 5. [OpenRouter rankings: notes on what the rankings measure](https://openrouter.ai/rankings) — undated page 6. [Hugging Face, deepseek-ai/DeepSeek-V4.1-Flash model card (MIT licence)](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) — 10 September 2026 7. [Hugging Face, zai-org/GLM-5.3: GLM-5.3 License](https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE) — August 2026 8. [Hugging Face model cards: MiniMax-M3 (minimax-community licence) and Kimi-K3 (kimi-k3 licence)](https://huggingface.co/moonshotai/Kimi-K3) — June 2026 9. [OpenRouter documentation: provider routing](https://openrouter.ai/docs/guides/routing/provider-selection) — vendor documentation, undated 10. [Digital Personal Data Protection Act, 2023 (section 16, transfer outside India)](https://www.meity.gov.in/data-protection-framework) — 11 August 2023 11. [MeitY, DPDP Rules, 2025 (main duties from 13 May 2027)](https://www.meity.gov.in/data-protection-framework) — 13 November 2025 Related on TechBag: the [DLP and insider-risk guide](https://www.thetechbag.com/browse/data-security-privacy/dlp-insider-risk), [how AI agents are billed](https://www.thetechbag.com/blog/how-ai-agents-are-billed-credits-actions), and [how MCP lets an agent run commands](https://www.thetechbag.com/blog/mcp-explained-ai-agents-run-commands). **Disclosure.** TechBag is a software reseller: it earns a margin when a reader buys through it, including from vendors named here. The analysis is TechBag’s own; no vendor paid for or saw this piece before publication. Facts are labelled reported, TechBag count or our view. Go deeper on TechBag - [DLP and insider-risk guide](https://www.thetechbag.com/browse/data-security-privacy/dlp-insider-risk) — the tools that control what staff send to AI services - [How AI agents are billed](https://www.thetechbag.com/blog/how-ai-agents-are-billed-credits-actions) — credits, actions and tokens: what a model choice does to the bill - [MCP explained](https://www.thetechbag.com/blog/mcp-explained-ai-agents-run-commands) — the agent side of the same governance questions