← All insights
The AI-First Web

Chinese AI Models Process 45% of the World's Tokens. A Year Ago It Was 2%.

DeepSeek-V4-Flash tops OpenRouter's global rankings at 3.43 trillion tokens per week. MiniMax, Kimi, Qwen follow. The AI model market followed the same cost-driven adoption curve as WordPress. The concentration risks may follow too.

· 6 min read
Share on X LinkedIn

The Shift Nobody Expected

One year ago, Chinese-origin AI models accounted for less than 2% of all token traffic on OpenRouter — the world's largest AI model routing platform. In June 2026, that number exceeds 45%. Nearly half of all AI inference — the reasoning, coding, writing, and analyzing that AI models do — now runs through Chinese models.

DeepSeek-V4-Flash leads the global usage rankings with 3.43 trillion tokens per week. MiniMax M2.5 and M3, Kimi K2.5, and Alibaba's Qwen3.7-Plus round out the top Chinese models. The five fastest-growing models on OpenRouter all offer either free access or pricing below $1 per million tokens.

45%+
Chinese model token share
Source: OpenRouter rankings analysis, May-June 2026. Was <2% one year ago.
3.43 trillion tokens
DeepSeek-V4-Flash weekly volume
Source: OpenRouter global rankings, late May 2026.

The WordPress Pattern

This should sound familiar. WordPress captured 43% of the web (W3Techs) through the same mechanism: free or cheap, good enough, widely available, with abundant tutorials and community support. The result was an enormous concentration of web infrastructure on a single platform — with 18,005 CVEs as the consequence.

Chinese AI models are winning volume through aggressive pricing and competitive capability. The pattern is identical: cost-driven adoption concentrates market share. Whether the risks of that concentration — single points of failure, data sovereignty concerns, geopolitical dependency — will mirror the WordPress story is an open question. But the trajectory is unmistakable.

The Data Sovereignty Question

When 45% of the world's AI inference runs through Chinese models, the data flowing through those models — prompts, context, code, business logic — traverses infrastructure governed by Chinese data regulations. For organizations in the US, EU, or other jurisdictions with data sovereignty requirements, this creates a compliance dimension that most haven't evaluated.

OpenRouter provides data policy controls that let organizations restrict which providers and models receive their prompts. But the default behavior — cheapest capable model wins — routes increasing volumes of Western business data through Chinese infrastructure. The market chose cost over sovereignty, just as it chose WordPress over security.

What This Means for the AI-First Web

WebPulse's thesis is that the web is being rebuilt for AI consumption. The AI models consuming the web are increasingly Chinese. The AI agents browsing your site, parsing your content, and extracting your data are increasingly powered by DeepSeek, MiniMax, and Kimi. The geopolitics of AI model adoption is becoming the geopolitics of web content consumption.

Framework choice matters here too. Sites with clean structured output (high AI-Readiness scores) are consumed more efficiently by any model — Chinese or Western. But the question of who processes your content, where the tokens flow, and what jurisdiction governs the inference is a strategic question that sits alongside framework choice, not beneath it.

Share this insight
More insights