Chinese AI Models Now Power 30% of US Companies' Weekly AI Usage
For years, the assumption in Silicon Valley was that American companies would default to American AI models like OpenAI, Anthropic, Google, almost as a matter of course. That assumption is quietly breaking down. New data cited by CNBC, sourced from AI routing platform OpenRouter, shows Chinese AI models now account for more than 30% of weekly token usage among US companies, with usage peaking as high as 46% at certain points since early February. A year ago, that figure sat at roughly 11%.
"Token usage" refers to the volume of text processed by AI models when companies run queries, generate content, or power AI features inside their products. OpenRouter, which acts as a kind of marketplace that lets developers route their AI requests to different models depending on cost, speed, or capability, has visibility into which models companies are actually choosing when given the option and not just which ones they say they prefer.
That makes this data particularly telling. It's not a survey asking companies which AI model they trust most; it's a direct measurement of where computing dollars and workloads are actually flowing.
The reasoning behind this shift comes down to a combination of cost, capability, and ease of integration. Chinese AI labs, including companies like DeepSeek and Zhipu AI (through its GLM series of models), have been aggressively pricing their models well below what comparable American frontier models charge, while closing much of the capability gap that once made Chinese models a clear step behind in raw performance.
For a developer or company optimizing for cost efficiency, the calculation has become less about brand loyalty and more about which model delivers acceptable quality at the lowest price per token. When a Chinese model can handle a given coding or content generation task at a fraction of the cost of a premium American model, many teams are making the switch, at least for a portion of their workloads.
It's important to be precise about what this data does and doesn't show. A 30% weekly usage share doesn't mean US companies are abandoning American AI providers wholesale. What it more likely reflects is the rise of multi-model routing as standard practice, companies increasingly split their AI workloads across several providers, sending simpler or high-volume tasks to cheaper models while reserving premium, higher-capability models for tasks that genuinely need them.
This shift toward routing flexibility matters strategically. It means competitive advantage in the AI model market is shifting away from pure brand power and toward performance, price, latency, and availability, a much more commoditized, competitive dynamic than the earlier "pick one lab and stick with it" era of AI adoption.
This trend arrives at a moment when China's AI labs have been explicit about their ambitions to match Western frontier labs. Zhipu AI's founder has publicly stated the company's GLM series will match Anthropic's most capable models before the end of the year, and the company has already launched agentic coding tools priced substantially below comparable American offerings. DeepSeek, meanwhile, has been working on its own inference chips specifically to reduce dependence on Nvidia and Huawei hardware, a move aimed at insulating its cost advantages from supply chain and export control disruptions.
For American AI labs, this data should serve as a wake-up call that competitive moats built purely on model capability may not hold if a rival offers "good enough" performance at a dramatically lower price. The gap that once clearly separated top-tier US models from Chinese alternatives has narrowed enough that price and practicality are now winning arguments in real procurement decisions, not just theoretical debates among AI researchers.
Whether this usage share continues climbing will depend heavily on a few unresolved factors: how aggressively US labs respond on pricing, whether new export control frameworks restrict how freely American companies can route sensitive workloads to Chinese-origin models, and whether Chinese labs can sustain their pace of capability improvements. What's clear already is that the era of American AI dominance being simply assumed, rather than competed for, appears to be over.