OpenRouter data show Chinese AI models have overtaken U.S. rivals in global token usage for the first time, signaling a real-world deployment shift as overseas developers expand their adoption.

OpenRouter, the world’s largest AI model API platform, serves mainly international developers—about 47% from the U.S. and 6% from China—highlighting the data’s global relevance.

Chinese models now hold multiple positions in the world’s most-used large language models, underscoring growing overseas adoption.

MoE architecture enables on-demand activation of only relevant expert networks, cutting inference memory by up to 60% and boosting throughput up to 19x, reducing costs.

Market implications point to explosive growth in token usage, with institutions like JPMorgan Chase forecasting hundreds of times more tokens through 2030 as tokens shift from mere internet traffic to production fuel.

Among the top models by weekly tokens, MiniMax’s M2.5 leads with roughly 1.7 trillion tokens, followed by Google’s Gemini 3 Flash Preview at about 997 billion.

Chinese models boast a cost advantage, with input tokens around $0.30 per million and output tokens between $1.1 and $2.55 per million, versus higher costs for competing Western models, aided by MoE and vertical integration.

A broader move toward AI agents and deeper workflow integration is driving higher token usage as AI becomes more embedded in productivity and complex tasks.

Experts view token consumption as a key proxy for real-world deployment, stressing the importance of scale and deployment reach alongside capability.

Notable Chinese models contributing to the top rankings include MiniMax M2.5, Dark Side of the Moon Kimi K2.5, Zhipu GLM-5, and DeepSeek V3.2, together accounting for about 85.7% of Top 5 volume in the week.

Moonshot AI’s Kimi K2.5 and Zhipu AI’s GLM-5 each exceeded 600 billion tokens in developer usage, placing them in the global top tier.

China’s DeepSeek V3.2 ranks third with roughly 798 billion tokens, indicating strong cross-border adoption of Chinese-developed models.