According to the latest data from OpenRouter, U.S. AI large language models have overtaken their Chinese counterparts in usage for the first time in nearly two months during the week ending April 19. This inflection point comes amid a global decline in overall AI model consumption, highlighting shifting competitive dynamics between the two AI superpowers.
U.S. Models Stage a Strong Rebound
For the week of April 13–19, U.S. AI models recorded a total usage of 4.908 trillion tokens, representing a 20.62% increase from the previous week. This marks the first time since mid-February that U.S. models have regained the lead over China. In preceding weeks, Chinese models—fueled by the explosive growth of open-source projects like DeepSeek—had consistently outpaced their U.S. rivals in token consumption.
Meanwhile, Chinese AI model usage experienced a notable decline, dropping to 4.441 trillion tokens, a 23.77% week-over-week decrease. Despite remaining at elevated levels, this decline ended China's two-month streak of dominance.
Global Total Falls, but China Retains Four Spots in Top Nine
Total global AI model usage for the week stood at 20.6 trillion tokens, down from the prior week. Nevertheless, Chinese models still claimed four of the top nine positions, demonstrating strong ecosystem resilience.
Specifically, DeepSeek V3.2 ranked second with 1.28 trillion tokens, trailing only an unnamed flagship model from OpenAI. The standout performer was MiMo-V2-Pro, which surged 90% week-over-week to 1.15 trillion tokens, climbing to fourth place. MiniMax also contributed two models to the top rankings, although its M2.7 variant saw a 19% decline in usage.
Market Interpretation and Outlook
Analysts attribute the U.S. rebound to recovering demand for enterprise API services and promotional campaigns by leading vendors. Conversely, China's usage drop may reflect post-Lunar New Year normalization and a shift toward more efficient vertical models.
Overall, the “token race” between the U.S. and China remains highly competitive. With new iterations of DeepSeek, MiMo, and other models on the horizon—alongside increased compute investments by American tech giants—the coming weeks will offer critical data points for tracking this rivalry.