Chinese-origin AI models now command approximately 61 per cent of all tokens consumed on OpenRouter — the largest neutral LLM router, processing over 20 trillion tokens weekly — according to data compiled through mid-2026. The shift represents one of the most dramatic market reversals in the AI industry's short history.
Xiaomi's MiMo model leads unexpectedly, capturing roughly 21 per cent of routed tokens and 22 per cent of all coding traffic — a surprising outcome from a company better known for smartphones than frontier AI. DeepSeek ranks second at approximately 17.6 per cent of tokens, surpassing every US laboratory including Anthropic at 15.4 per cent. Alibaba's Qwen anchors the broader ecosystem as the most-downloaded open model globally, with over 1 billion cumulative downloads and approximately 40 per cent of new LLM derivatives on Hugging Face now based on Qwen.
The pricing advantage is stark. DeepSeek-V4-Pro lists at approximately $0.435 per million input tokens — roughly three times cheaper than Gemini 3.1 Pro and approximately twelve times below GPT-5.5 at comparable performance levels. This cost arbitrage is driving adoption at scale, particularly for cost-sensitive coding and enterprise workloads.
The displacement of Western models has been swift. Meta's Llama, which defined open-weight AI in 2023–24, fell below 1 per cent of routed volume after complete displacement by Chinese alternatives. Google's share collapsed from roughly 37 per cent to 13 per cent as developers shifted cost-sensitive workloads eastward.
At the national level, China's daily token consumption reached 140 trillion by March 2026, according to the National Data Administration — a more than thousand-fold increase from 100 billion in early 2024. The first quarter of 2026 saw consumption rise more than 40 per cent compared to end-of-2025 levels. Xiaomi's head of large models predicted token consumption could increase a further hundred-fold in 2026 alone, driven by agent-based applications.
The pattern has been described as 'token exports' — Chinese AI companies positioning themselves as major suppliers of computational tokens to global markets, leveraging cost advantages to compete internationally while US frontier labs command premium revenue per token but lose volume share.
For context engineers, the data challenges a persistent assumption that US labs dominate global AI usage. While Anthropic, OpenAI and Google lead in premium enterprise and frontier capabilities, Chinese models now capture the majority of actual consumption. Developers building multi-model architectures increasingly route cost-sensitive workloads to Chinese providers — a trend that raises both opportunity and data sovereignty questions.
