Conceptual diagram of AI Photo: VCG

US companies are reportedly increasingly adopting Chinese-built artificial intelligence (AI) models as they expand AI-powered tools and products. The trend is being driven primarily by cost-performance advantages rather than a simple shift in preference, highlighting the rising global competitiveness of Chines AI models, experts said.

Recent model releases from Chinese AI companies such as DeepSeek and Zhipu AI have been viewed by many users as highly competitive with leading frontier systems from US firms including Anthropic and OpenAI. At the same time, the token prices of some of the most advanced American AI models have been rising, creating unexpected cost pressure for businesses deploying AI at scale, according to US media outlet CNBC. 

Data cited by CNBC showed that the share of tokens used by US companies on Chinese AI models through OpenRouter — a platform that gives developers access to a wide range of models — has remained above 30 percent each week since February 8, reaching as high as 46 percent. That compares with an average share of just 11 percent over the past 12 months and only 4.5 percent in the first half of 2025.

Lindy, a US-based AI startup, switched all of its traffic from Anthropic's Claude to DeepSeek in June. The company's CEO said the move would save the company millions of dollars over the coming months, CNBC reported.

Industry analysts pointed out that reducing costs and improving efficiency is the core reason US companies are choosing Chinese AI models. Against the backdrop of accelerated digital transformation, Chinese open-source models, which combine performance and cost advantages, are well suited to companies' core needs for self-reliance, controllability, and higher quality at lower cost. 

In addition, tighter regulatory scrutiny of leading domestic AI models in the US is further encouraging overseas companies to explore Chinese alternatives. As a result, Chinese AI models are seeing expanding opportunities for global adoption, they added.

Chinese AI models have become dramatically less expensive to use due to a combination of technical and market factors, Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute, told the Global Times on Wednesday. He pointed to the continued evolution of Mixture of Experts (MoE) architecture, open-source expansion and low-price strategies, as well as the rapid rise of agent-based applications, which place greater emphasis on low-cost inference.

"The AI market is moving from who has the strongest model to who is more cost-effective and easier to deploy," Tian said. He cited Zhipu AI's GLM 5.2 as an example, saying it reportedly comes within 1 percent of Anthropic's Opus 4.8 on a major agent benchmark while costing only about one-fifth as much.

Xiang Ligang, a veteran telecom industry observer, said the final stage of AI deployment is inference, where token consumption becomes the key cost factor. For many practical applications, including autonomous driving, medical services and home robots, access to the largest and most expensive models is often unnecessary.

Xiang told the Global Times that domestic developers such as DeepSeek, Zhipu AI and Alibaba's Qwen have focused more on architectural optimization, token efficiency and smarter inference strategies. As a result, they can deliver competitive performance on many enterprise tasks while using less computing power per request, which directly lowers costs for both providers and users.

In addition, US export restrictions that limit China's access to the most advanced GPUs have pushed Chinese AI labs to focus heavily on algorithmic efficiency. Techniques such as mixture-of-experts architecture, multi-head latent attention and FP8 mixed-precision training have enabled Chinese teams to achieve strong performance with fewer computing resources, Tian said.