Chinese AI Models Now Process 60% of US Business Tokens on OpenRouter

Nobody planned this. No CTO sat in a boardroom and declared "we are going to route the majority of our AI workloads through Chinese models for strategic reasons." What happened instead is what always happens when a commodity market meets a cost curve: buyers went where the price was lowest, and the lowest price was not American.

On OpenRouter, the popular marketplace where developers and businesses route AI workloads across competing models, Chinese open-weight models now process roughly 60% of token volume from American companies. All five of the top models by weekly usage are Chinese. Moonshot's Kimi K3 and DeepSeek variants hold the leading positions. The frontier models from OpenAI, Anthropic, and Google are still there, still used for hard problems... but the volume tells a different story than the headlines.

The Two-Tier Market

What has emerged is a clearly stratified market that the "AI arms race" narrative did not predict. At the top tier, US frontier labs still lead on raw capability. For complex reasoning, code generation, creative tasks, and anything requiring state-of-the-art intelligence, OpenAI's and Anthropic's best models remain the choice. That has not changed.

But the second tier, the commodity layer of AI inference where businesses run customer service bots, content classification, data extraction, summarization, translation, and thousands of other "good enough" tasks, has gone to the lowest-cost provider. And the lowest-cost providers are Chinese open-weight models that run on commodity hardware, cost a fraction per token, and perform well enough for 80% of real-world use cases.

This is the pattern from every technology commoditization cycle. The premium tier holds on capability. The volume tier goes to cost. And the volume tier is where the majority of the money is.

What This Means for US AI Companies

The immediate implication is uncomfortable for anyone modeling AI revenue at US hyperscalers. If the commodity layer of inference, the high-volume, lower-margin workloads that were supposed to generate scale revenue from enterprise adoption, is increasingly running through Chinese models, then the revenue projections that justify current AI capex spending need re-examination.

This does not mean OpenAI, Anthropic, or Google are in trouble. It means their addressable market may be narrower than the market has priced. The "every business will use our API for everything" thesis is being replaced by a more nuanced reality: businesses will use expensive frontier models for hard tasks and cheap commodity models for everything else. The split matters enormously for revenue forecasting.

The Open-Weight Advantage

A critical piece of this story is the open-weight model. Chinese models dominating OpenRouter are not proprietary cloud APIs like OpenAI's. They are open-weight releases that anyone can download, fine-tune, and deploy on their own infrastructure. This creates a fundamentally different competitive structure.

When the model weights are open, the competition shifts from "whose API is best" to "who can run inference cheapest." And on that axis, the combination of efficient Chinese architectures, lower development costs, and the ability to deploy on commodity hardware creates a structural cost advantage that proprietary US models cannot easily match without releasing their own weights.

For builders and operators, this creates a practical question: if Chinese open-weight models handle 80% of your workload at 20% of the cost, what is your actual AI budget, and how much of it should go to premium frontier access versus commodity volume? The answer depends on your use case, your risk tolerance for model provenance, and your assessment of whether open-weight models from Chinese labs carry any data-handling or security considerations your compliance team cares about.

What to Watch

Three threads to follow from here:

1. Hyperscaler earnings commentary next week. Amazon, Meta, and Microsoft all report. Listen for any acknowledgment that commodity inference pricing is under pressure from open-weight competitors. The absence of that acknowledgment is itself a signal.
2. Export control response. The US policy apparatus has not yet adjusted to the reality that Chinese AI competitiveness is not primarily a compute story anymore. It is an efficiency and cost story. Watch for any policy shifts targeting model distribution rather than chip distribution.
3. Enterprise adoption patterns. OpenRouter is one marketplace. The broader question is whether this pattern holds across direct API usage, cloud deployments, and enterprise contracts. If enterprise IT departments start routing through Chinese models at scale, the revenue implications for US AI companies compound.

The AI moat narrative assumed that intelligence was the only axis of competition. The market is adding a second axis: cost. On that axis, different players are winning.