Chinese AI models gain global ground on developer platforms

On September 26, data reported by CNBC showed Chinese artificial intelligence models had become a majority of usage on two major developer platforms, a sharp change from the start of 2026. On OpenRouter, Chinese models represented between 57% and 67% of tokens used in the week of September 14, compared with 6% to 13% in February. Vercel said their share reached 55% in August, up from 11% in January. The figures point to rapid international uptake, particularly for software development and other tasks handled by AI agents.

The numbers do not establish that Chinese providers lead every part of the market. US frontier models still command more overall spending and remain preferred by some companies for complex work. But the shift matters because OpenRouter and Vercel provide businesses with ways to access and compare models from multiple suppliers. Token use is an indicator of activity on those platforms, rather than a complete measure of global AI revenues, users or technical leadership.

Lower prices meet improving capabilities

Platform executives attributed much of the adoption to a practical calculation: models only need to meet the quality threshold for a given task before their price becomes decisive. Harpreet Arora, Vercel’s head of agentic infrastructure, told CNBC that Chinese models are becoming capable enough for more work at substantially lower cost. Peter Walker, OpenRouter’s head of insights, said Chinese open-source models released this year can now credibly handle advanced agentic use cases, especially coding, in ways that were not true in late 2025.

Chinese companies including DeepSeek, Z.ai and Alibaba have released models with significant gains in coding and related tasks. Their progress has narrowed the practical gap for organisations that do not require the most capable system available. CNN reported that developers weighing AI tools may see a choice between a model that is more advanced but far more expensive and another that is cheaper, customisable and sufficiently capable for routine business needs.

• On OpenRouter, Chinese models accounted for 57% to 67% of tokens in the week of September 14, versus 6% to 13% in February.
• On Vercel, their share rose from 11% in January to 55% in August.
• OpenRouter said companies in its defined “Global South” used more than two-thirds, or 67%, of their tokens on Chinese models in recent weeks.
• About half of the tokens on OpenRouter are used by companies in the United States, according to the platform.

Open models broaden the appeal

China’s appeal is linked not only to price but also to the availability of open models, which can be downloaded, adapted and run by users rather than accessed solely through a paid proprietary service. CNN reported that Chinese developers have leaned into open systems partly in response to US export controls that restrict access to the most advanced chips, as well as to smaller domestic capital markets. That approach gives customers more flexibility to fine-tune systems for local languages, industries and internal workflows.

The strongest demand on OpenRouter has come from businesses in the countries the platform groups as the Global South: 82 nations across Central and South America, Africa and Asia. Daniel Remler of the Center for a New American Security said Southeast Asia could see substantial uptake because of close economic and cultural links with China and expanding digital infrastructure. He also pointed to demand from entrepreneurs and governments seeking inexpensive, open models in cities from Lagos to São Paulo and Jakarta.

Usage figures have important limits

The available data should be read carefully. OpenRouter’s figures cover companies in the United States, Europe and its own Global South classification, while Vercel did not specify the geographical distribution of its data. Neither dataset represents every cloud provider, consumer application or enterprise deployment. A rising token share can reflect many factors, including pricing, product routing decisions and experimentation by developers. It does not by itself show which models generate the most profit or perform best across all benchmarks.

US models continue to lead most benchmarks, according to CNBC’s account, and businesses still turn to frontier US systems for certain difficult tasks. OpenAI and Anthropic also announced cheaper models earlier this week. Anthropic product executive Dianne Penn said the company was seeking to make responses more efficient, using fewer tokens depending on the selected effort setting. That suggests the cost competition is likely to intensify rather than settle around a single group of providers.

A commercial shift with geopolitical consequences

The growth of Chinese AI use is drawing scrutiny in Washington. Two US House committees are investigating the effects of growing adoption of Chinese models, amid concerns about security, technology competition and Beijing’s international influence. The United States has sought to maintain an advantage through export controls limiting Chinese companies’ access to the most advanced chips. US officials are also concerned about potential remote access to Nvidia chips through overseas data centres and allegations involving distillation, in which newer models imitate more established systems.

Those allegations remain contested. CNN reported that OpenAI and Anthropic have claimed Chinese labs trained models on outputs from US systems, while Beijing has repeatedly rejected accusations of distillation. DeepSeek and Moonshot did not respond to CNN’s request for comment. The same report noted uncertainty over how much Chinese developers may rely on access to US computing capacity abroad or on distillation, making definitive assessments difficult from public information alone.

Openness may not be permanent

For overseas customers, the durability of China’s open-model strategy is a central uncertainty. Analysts cited by CNN said Beijing may impose tighter controls as its most capable models improve, especially where cyber capabilities create security concerns. China’s models may therefore develop through a mix of open and closed releases rather than remain freely available without conditions. For companies in Europe and elsewhere, the immediate lesson is less about choosing a permanent camp than about testing models against cost, data handling, reliability and local compliance requirements. The rapid rise in platform usage shows that Chinese AI is already becoming part of that procurement calculation.