Chinese Models Capture Up to 46% of U.S. Companies' Token Consumption, and the Numbers Add Up
DeepSeek and GLM 5.2 are available at a fraction of the cost of American frontiers and attract corporations that pay per token. CNBC presents the chart that OpenAI and Anthropic preferred not to see published.
The share of tokens consumed by American companies in Chinese language models via OpenRouter surged past 30% for the entire week since February 8, reaching peaks of 46% in recent weeks, according to data published by CNBC on July 7. The chart provides the missing argument for a thesis that has been circulating among infrastructure analysts: the unit cost of American frontiers has risen faster than the willingness of CIOs to pay for them.
DeepSeek continues to charge about 3% of the price per token of GPT 5.5, according to CNBC's own survey. GLM 5.2, from the Chinese company Z.ai, has a base price of $0.95 per million input tokens and $3.00 per million output tokens, against a range exceeding $4.00 per million for American frontier models. Chinese open models are priced between 60% and 90% lower than their equivalent Western competitors in a similar timeframe.
Benchmark for Agent and Volume Growing on a Vertical Curve
GLM 5.2 reached less than one percentage point of Claude Opus 4.8 in an agentic benchmark monitored by Vercel, at approximately one-fifth of the price. Vercel also recorded the fastest growth of any model on its platform this year: a daily token volume multiplier of 27 times and an 80-fold increase in its customer base during the first full week after launch. This behavior can only be sustained if the model is genuinely entering production pipelines rather than merely undergoing exploratory testing.
A second piece of data, this time from DeepSeek, helps complete the picture. The company began directly marketing to American clients through resellers listed on OpenRouter, which reduced the compliance friction perceived by CIOs who previously required an intermediary layer in Europe. The result has been a consistent 30% weekly share since February, with spikes to 46% during feature release weeks.
Where the Effect Hits: United States, Germany, and India
In the United States, platform providers like Vercel, Fireworks, and Together AI have become bridges for Chinese models, bypassing reputational barriers. Code, service, and RAG startups that cover OpenAI's monthly bill see a 60% savings and do not hesitate to run A/B tests. In Germany, where CIOs at banks like Deutsche Bank and Commerzbank are under explicit pressure to contain compute costs by 2026, GLM 5.2 and DeepSeek V4 are emerging as second engine options for prompt routing, particularly for summarization and classification tasks. In India, TCS, Infosys, and Wipro combined operate over 300,000 seats of Microsoft 365 Copilot, but their internal engineering practices have been piloting Chinese models, driven by a dual objective: reducing token costs and freeing themselves from strategic dependence on a single vendor. All three are situated in regions sensitive to data control, pushing pilots toward on-premises deployments on their own GPUs or within sovereign clouds.
The part that CNBC did not write, and which sourcing analysts echo behind the scenes, is that prompt routing has become an engineering discipline: separating complex calls for Opus or GPT-5 and directing routine volume to inexpensive Chinese models. This is different from merely switching suppliers. It represents an architecture that assumes two permanent engines.
The Steelman Argument from the American Side
The opposing view deserves mention. Analysts from Morgan Stanley and Sam Altman himself argue, with some justification, that the price per token understates the total cost. A cheaper Chinese model requires more rounds of prompt engineering, yields less consistent quality jumps in Portuguese and German, and carries a layer of geopolitical risk that has already materialized in corporate decisions for complete blockages. Anthropic, for its part, stated in May that it is on track for $47 billion in current revenue by 2026 and to profitability by 2029, which suggests ongoing demand for the premium segment. The counterargument does not invalidate the CNBC chart but contextualizes it: part of the decline in tokens is due to substitution, and part is new volume that only exists because it became cheaper.
The question for the next quarter is not whether to choose between American and Chinese. It is about designing the routing architecture that can withstand both suppliers changing their prices in the same month.