Chinese AI Models Triple Market Share to 30% Globally

• Chinese AI models’ global adoption has nearly tripled from 13% to approximately 30% of total usage in 2025, according to OpenRouter’s 100 trillion token analysis
• Programming and roleplay dominate AI usage patterns, with open-source models capturing one-third of the market

Recent data reveals that Chinese AI models’ global adoption has reached unprecedented levels, with models from mainland China, Taiwan, and Hong Kong accounting for nearly 30% of total artificial intelligence usage by late 2025—a dramatic surge from just 13% at the start of the year.

The findings emerge from OpenRouter’s newly released State of AI report, which analysed over 100 trillion tokens of real-world large language model (LLM) interactions across 300+ models from 60+ providers.

The report provides the most comprehensive empirical study to date on how AI systems are actually being deployed in practice.

“Chinese open-source models steadily gained traction, reaching nearly 30% of total usage among all models in some weeks,” the report states, highlighting models from DeepSeek, Qwen (Alibaba), and Moonshot AI’s Kimi as key drivers of this expansion.

DeepSeek leads the open-source ecosystem with 14.37 trillion tokens processed during the study period, though its market dominance has fragmented as competitors emerged. Qwen ranked second with 5.59 trillion tokens, followed by Meta’s LLaMA at 3.96 trillion tokens.

The report identifies what researchers term a “Summer Inflexion” in mid-2025, when the open-source AI landscape shifted from near-monopoly to pluralistic competition.

“By late 2025, the competitive balance had shifted from near-monopoly to a pluralistic mix. No single model exceeds 25% of OSS tokens,” according to the analysis.

Simplified Chinese now accounts for nearly 5% of global token volume—the second-largest language after English, which dominates at 82.87%. This linguistic distribution reflects sustained engagement by users in bilingual or Chinese-first environments and underscores the growing importance of multilingual AI capabilities.

Asia’s overall share of AI inference spending more than doubled during the study period, rising from approximately 13% to 31% of global usage. Singapore emerged as the second-largest country by token volume (9.21%) after the United States (47.17%), with China ranking fourth at 6.01%.

The report reveals unexpected usage patterns that challenge conventional assumptions about AI deployment. Contrary to beliefs that productivity tasks dominate, creative roleplay accounts for over 50% of open-source model usage, while programming represents the second-largest category and the fastest-growing segment overall.

“Programming queries accounted for roughly 11% of total token volume in early 2025 and exceeded 50% in recent weeks,” the researchers note, reflecting AI’s deep integration into software development workflows.

Open-source models now represent approximately one-third of total AI usage, reaching an “equilibrium at roughly 30%,” according to the report.

This marks a fundamental shift in the AI ecosystem, where proprietary systems from OpenAI, Anthropic, and Google compete directly with community-driven alternatives.

The research introduces the concept of “agentic inference,” describing a shift from single-turn text generation to multi-step, tool-integrated reasoning workflows.

Reasoning-optimised models now account for more than 50% of all tokens processed—up from nearly zero in early 2025—indicating that AI systems are increasingly deployed as components in larger automated systems rather than for isolated queries.

“The median LLM request is no longer a simple question or isolated instruction. Instead, it is part of a structured, agent-like loop, invoking external tools, reasoning over state, and persisting across longer contexts,” the report explains.

Average prompt token lengths have grown nearly fourfold since early 2024, from around 1,500 to over 6,000 tokens, while completion lengths nearly tripled. Programming workloads drive this expansion, with code-related requests routinely exceeding 20,000 input tokens.

The study also identifies what researchers call the “Cinderella Glass Slipper effect”—a phenomenon where early user cohorts who find precise model-workload alignment demonstrate significantly higher retention rates than later adopters.

This pattern suggests that being first to solve a critical workload creates a lasting competitive advantage. Cost analysis reveals that demand remains relatively price-inelastic, with pricing power varying significantly across market segments.

Premium models from Anthropic and OpenAI command around US$2-35 per million tokens while maintaining high usage, while “efficient giants” like Google’s Gemini 2.0 Flash and DeepSeek V3 achieve similar scale at under US$0.40 per million tokens.

The report concludes that “LLMs must be globally useful—performing well across languages, contexts, and markets. The next phase of competition will hinge on cultural adaptability and multilingual capability, not just model scale.”

For the full State of AI report with detailed methodology and analysis, visit OpenRouter’s website.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology eventsclick here for more information.