Recent routing and spending data from Vercel, OpenRouter and Ramp point to growing use of Chinese AI models in production workflows, including among U.S. businesses.
Vercel’s Oct. 15 AI Gateway Production Index, based on traffic routed between production applications and AI labs, said open-weight models processed more than seven in ten gateway tokens in September and accounted for nearly one in four gateway dollars. The same report said nearly 70% of September gateway spend on Chinese models went through U.S.-based providers, up from less than 10% in December.
That distinction matters. The model can be developed by a Chinese lab while access, hosting and revenue flow through an American provider. Vercel said DeepSeek models processed 58% of all AI Gateway tokens in September for 6.6% of spend. Moonshot AI, the Beijing-based company behind Kimi, accounted for 3.6% of gateway tokens and 13.2% of spend, with U.S.-based providers serving nine in ten of those tokens. Z.ai, also based in Beijing and creator of GLM, accounted for 7.2% of all tokens and 4.5% of spend, with U.S.-based providers serving seven in ten.
OpenRouter’s rankings showed a similar production signal, though from a different slice of the market. Usage data through Oct. 15 ranked DeepSeek V4.1 Flash first by tokens processed over the trailing 30 days. GLM 5.3 Flash from Z.ai, Hy4 preview from Tencent, MiMo-V2.6-Flash from Xiaomi and other China-linked model families also appeared high on OpenRouter’s token tables. OpenAI and Anthropic still appeared in the top rankings, but the mix is no longer dominated by U.S. model families on router-based usage alone.
OpenRouter has also cautioned that its rankings measure traffic through its own API. They do not measure total industry usage, users, spend or model quality. The trend is strongest among AI-heavy teams using model gateways and routers, not across every American business.
Cost is the clearest pull
The economics are hard to ignore. OpenRouter’s June analysis listed DeepSeek V4 Flash, on its cheapest endpoint at the time, at $0.09 per million input tokens and $0.18 per million output tokens. It compared that with GPT-5.5 pricing of $5 per million input tokens and $30 per million output tokens at the time.
Those exact prices can change, but the broader incentive is clear. High-volume AI agents, code tools, research assistants and workflow automations can burn through tokens quickly. A model that is good enough for routine work can become attractive even if a more expensive frontier model remains stronger for high-stakes reasoning, brand-sensitive responses or complex tasks.
The Associated Press reported in July that Coinbase said it was switching to Chinese AI models to trim costs. AP also described U.S. developers and executives using Moonshot’s Kimi, Z.ai’s GLM and DeepSeek for routine work such as research, documents, calendar tasks and business lead generation.
Vercel’s data points to a related pattern. September gateway spend was about 13 times December’s level, while the average price per token fell 47%. In other words, cheaper models are not necessarily reducing AI spending. They may be allowing companies to run more AI work.
Chinese models are complementing, not simply replacing, U.S. labs
Ramp Economics Lab offers a useful counterweight to the router data. In June 2026, 5.8% of AI-spending businesses on Ramp used model-serving platforms, up from 4.5% in January. Ramp described those platforms as an imperfect proxy for open-source and Chinese model adoption.
That is growth, but it is not broad market takeover. Ramp also found that among businesses using model-serving platforms, 85.8% also used OpenAI, 93.2% used Anthropic, and 96.4% used at least one of the two.
The emerging pattern is multi-model adoption. Chinese and open-weight models are taking more workloads where price, latency, customization or routing flexibility matter. U.S. frontier labs still hold strong positions where companies value perceived quality, integrated products, procurement familiarity, enterprise controls and vendor trust.
Vercel’s October report points in the same direction. It said 68% of teams that put more money into open-weight models in September also put more money into closed-weight models. The model is becoming a replaceable component inside a larger AI system, rather than a single vendor decision made once and left alone.
Policy risk is now part of the model choice
U.S. lawmakers are already treating Chinese open-weight AI adoption as a security and supply-chain issue. In April, the House Committee on Homeland Security and the House Select Committee on China began a joint investigation involving Airbnb and Anysphere. In July, the committees said they were continuing the inquiry and had requested information from DoorDash.
The July committee update said the investigation covers U.S. companies integrating People’s Republic of China-developed open-weight AI models into consumer platforms, enterprise software and systems with access to sensitive commercial or personal information. The committees also acknowledged why companies may evaluate those models, including competitive capabilities, lower costs, greater customization and alternatives to reliance on a small number of proprietary providers.
The April letter to Airbnb said the committees were examining Airbnb’s reported use of Alibaba’s Qwen large language model in customer service operations. The letter asked for details including model names, versions, business functions, whether each model was accessed through an API, self-hosted infrastructure or a third-party inference provider, and how customer data or other sensitive inputs moved through the processing chain.
The committee letters contain allegations and risk concerns, not a final finding of misuse. But they show how model provenance has moved from an engineering detail to a governance, procurement and policy question.
China is adding its own export-control wrinkle
On Oct. 16, China’s Ministry of Commerce and Ministry of Science and Technology proposed revisions to the country’s catalogue of technologies prohibited or restricted from export. The proposal covers 40 technologies across artificial intelligence, biomedicine and photovoltaics. The AI section includes 23 technologies related to large models, agents, compute clusters and information processing.
The proposal is open for public comment from Oct. 16 to Nov. 15. The announcement says technologies already in the public domain through open sourcing or academic publication would not be included in the management scope. A temporary pre-export registration requirement took effect on Oct. 16 for transfers involving technologies on the proposed list, even though the revised catalogue itself is still under consultation.
A temporary pre-export registration requirement took effect on Oct. 16 for transfers involving technologies on the proposed list, even though the revised catalogue itself is still under consultation.
A ministry Q&A added an important operational distinction. Overseas local deployment of restricted AI technology may fall within technology export management, while providing API calls, app or web interactions, SaaS applications and lightweight local deployments using restricted large models are not within the big-model technology export management scope described by the ministry.
AP reported that the draft could restrict exports of AI models with parameter counts of 2 trillion or more, or models with certain leading capabilities. AP also reported that Moonshot’s Kimi K3 and Alibaba’s Qwen 3.8-Max are among Chinese models above 2 trillion parameters, while the proposal did not name specific models.
The practical questions are getting more specific
For companies building AI into products, the model decision is becoming less about nationality alone and more about the full route from user prompt to model response. A Chinese-developed model served by a U.S.-based provider creates a different operational profile than a direct API call to a China-based provider or a self-hosted deployment using downloaded weights.
The questions now include which workloads go to low-cost models, which stay with frontier models, where prompts and metadata are processed, which entity hosts inference, what jurisdiction applies, what the license permits, what security testing has been performed, and how quickly the system can fall back to another model if access, pricing or policy changes.
Get new small business insights by email
Practical ideas and useful articles to help you make better business decisions.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
