Source: Fortune
Summary
Chinese AI models dominated the top five spots on OpenRouter in July, with Xiaomi’s MiMo V2.5 leading by token volume. Chinese models now account for over 60% of the platform’s traffic, surpassing 20 trillion tokens weekly. US models previously held 70% of traffic but now carry only 30%. Chinese models offer significantly lower costs, making them attractive for high-volume workloads. Alibaba’s Qwen family has surpassed Meta’s Llama in downloads, signaling a shift in developer preference.
Our Reading
The numbers tell one story.
Chinese models dominate OpenRouter traffic, outpacing US models by a wide margin.
US labs still lead in frontier AI but are losing ground in distribution and cost.
Chinese models are cheaper and more widely used, forcing companies to choose between price and performance.
The death zone is where models neither lead in capability nor cost are left behind.
Author: Evan Null
The race split in two
The AI race has split into two distinct contests: capability and distribution. While US labs still hold the frontier, Chinese models are winning in scale and cost. This shift is reshaping how companies choose AI tools, with price and volume becoming key factors. The US is leading in advanced AI but losing in the broader market. Chinese models are not just cheaper—they are more widely adopted and integrated into developer workflows.
Welcome to the death zone
The death zone is where models that are neither the best nor the cheapest get crushed. Companies are choosing between premium frontier models and low-cost open models, leaving the middle behind. This bifurcation is clear in OpenRouter data, where Anthropic holds a small share but captures most spending. The middle is not sustainable, and many enterprise AI strategies are stuck there. The cost of being in the middle is high, and the pressure to choose a side is growing.
China built this on purpose
China’s success in AI is not accidental. Export controls forced labs to innovate with efficiency, leading to cheaper and more powerful models. State support further reduced costs, allowing companies like Xiaomi to slash API prices. This strategy has paid off, with Chinese models dominating both usage and downloads. The US is now playing catch-up, trying to match the efficiency and scale that China has built over years of constraint.
The builder’s playbook for 2026
For AI builders, the key is to embrace hybrid routing, treat efficiency as a weapon, differentiate above the model layer, and avoid the middle. Companies that route work to the right models—frontier for critical tasks, open for cost-sensitive ones—can cut costs without sacrificing quality. Efficiency is no longer a secondary concern but a core strategy. Differentiation through data, application layers, and domain-specific tuning is now essential. Staying in the middle is not an option in 2026.
America needs an open weight answer now
The US is not responding effectively to the rise of Chinese open models. While security concerns justify some restrictions, a blanket ban is not a strategy. Open weights are how ecosystems grow, and the US has no equivalent to China’s open model dominance. Meta’s retreat left a gap that China filled quickly. To compete, the US needs to release frontier-class open models regularly and support them with procurement incentives. The future of AI is not just about being the best, but about being the most accessible and affordable.
