When cost and practical application takes priority: China surpasses US in AI adoption

In February 2026, the latest data released by OpenRouter, the world’s largest API aggregation platform for AI models, showed that Chinese AI models have achieved a historic breakthrough in global call volume, surging 127% within three weeks, surpassing the long-leading US for the first time. Chinese AI models also occupy four of the top five spots globally, signalling a clear turning point towards a “strong East, weak West” dynamic in the global AI application landscape.

My articles have tracked this shift — from catching up to overtaking — over the past three years. This overtaking is the result of three years of meticulous development in China’s AI industry, with sustained and focused efforts on cost efficiency and practical applications, while also highlighting a profound shift in the centre of global AI competition.

Notably, the majority of OpenRouter’s users are overseas developers, with 47.17% from the US and only 6.01% from China. This indicates that the rise of Chinese AI models is driven by global developers’ “foot voting” rather than merely domestic market support.

Three stages of development are clearly illustrated when looking back over the past three years of global AI call volume, as the gap between China and the US evolved from a vast divide to parity, and now to overtaking.

The first stage was “the US monopoly period” of 2023, when the global AI applications market was almost entirely dominated by US companies. Data from OpenRouter that year showed that the top five models worldwide by call volume were all from the US: GPT-4 by OpenAI, Claude 2 by Anthropic, Gemini 1.0 by Google, LlaMA 2 by Meta, and Copilot by Microsoft. Leveraging their first-mover advantage, these models accounted for over 90% of global call volume.

Meanwhile, back then, Chinese AI models such as Baidu’s ERNIE Bot and Alibaba’s Qwen were still in the stage of technical refinement and domestic market development. They primarily served local enterprises and users, accounted for less than 10% of global call volume, and were largely concentrated within Chinese-speaking communities and parts of Southeast Asia, with virtually no influence in the global market. 

At the time, the US held firmly in its hand the global AI industry’s discourse power, dominating both technical standards and business models. Chinese models were seen as “followers”, struggling to break through the US’s technological and market barriers.

The second stage was “China’s catch-up period” of 2024, when Chinese AI models began accelerating their global expansion and gradually narrowed the gap with the US. That year, Chinese companies such as MiniMax, Moonshot AI, Zhipu AI and DeepSeek launched their models on major global API aggregation platforms such as OpenRouter, formally initiating their overseas expansion.

Compared with US models, the core advantage of Chinese models lies in their exceptional cost-effectiveness, with pricing at just one-fifth to one-tenth of comparable US models. For instance, MiniMax M2.5 is priced at US$1.10 per million tokens, whereas Claude Sonnet by Anthropic costs as much as US$15 per million tokens. This price advantage has quickly attracted a large number of cost-sensitive overseas developers and small- and medium-sized enterprises.

Meanwhile, Chinese companies adopted a “rapid iteration” approach, optimising model performance on a monthly or even weekly basis, and gradually catching up with US competitors in key areas such as ultra-long context processing, multimodality and code generation.

By the end of 2024, the global call volume share of Chinese AI models had risen to 30-40% — still lagging behind the US but significantly narrowing the gap. Some Chinese models began entering the global top ten by call volume, breaking the long-standing dominance of US models.

The third stage is “China’s overtaking period” since February 2026, during which Chinese AI models achieved a historic breakthrough in formally surpassing the US in global call volume.

Data from OpenRouter showed that in early February, the call volume of Chinese models reached 4.12 trillion tokens, compared with 2.94 trillion tokens for the US, putting China about 40% ahead. By mid-to-late February, the call volume of Chinese models had surged to 5.16 trillion tokens, while the US saw a slight decline to 2.7 trillion tokens, widening China’s lead to 91%.

In just three weeks, Chinese models’ call volume skyrocketed 127%, reflecting global developers’ recognition of their performance and cost-effectiveness. Today, Chinese models are not only more affordable but have also overtaken their American counterparts in core capabilities.

For example, Kimi K2.5 supports ultra-long context processing to the millions in tokens, making it the preferred choice for handling professional documents in research, legal and financial sectors; GLM-5 and DeepSeek V3.2 now match GPT-4 Turbo in code generation efficiency and accuracy; and MiniMax M2.5, with its multimodal capabilities and emotional interaction features, has gained widespread favour among consumer-side developers and content creators.

With this, Chinese AI models have completely shed the “follower” label, and achieved a leap from catching up to leading the field.

As I have repeatedly stressed, China’s AI industry does not prioritise fundraising, instead focusing long term on two core aspects: cost and practical applications. After years of being underestimated by Western media, China’s AI sector has now fully demonstrated its capabilities, rivalling those of the US.

In terms of cost, Chinese AI models achieve high efficiency at low prices, thanks to system-level advantages such as innovative architectures, a thriving open-source ecosystem and scenario-driven optimisation, rather than mere price competition. 

At the computational level, Huawei Cloud’s CloudMatrix 384 supernode reduces latency to 0.8 microsecond through dynamic routing, boosts inference efficiency by 40-50%, and, with a liquid-cooling + green energy solution, lowers the power usage effectiveness to 1.09, significantly cutting operating costs. At the model level, Alibaba’s Qwen 3.5 uses a sparse activation architecture, greatly reducing computational demand and cutting inference costs to just US$0.0003 per token. 

At the ecosystem level, China follows an open-source, inclusive approach. Domestic open-source models such as DeepSeek-R1 spread research and development costs from hundreds of millions of dollars down to tens of millions, attracting global developers to contribute to iterative optimisation, further reducing overall costs. This system-level cost control gives Chinese models a formidable price advantage in the global market.

In terms of practical applications, unlike the “research-heavy, slow-iteration” approach of major US enterprises, Chinese AI companies adopt a strategy of rapid response and agile iteration, enabling them to quickly identify developers’ pain points worldwide and meet real-world user needs at speed.

For example, in response to overseas developers’ demand for ultra-long context processing and multimodal interaction, Chinese firms completed multiple rounds of iteration within just a year: Kimi rapidly scaled from supporting 100,000-level token context to million-level capacity, while MiniMax achieved seamless integration across text, image and speech modalities.

In response to enterprise users’ need for customisation, Chinese providers can rapidly deliver tailored solutions, whereas US models often take months to complete such adaptations. This user-centric iterative approach enables Chinese models to quickly align with the diverse demands of the global market, earning broader recognition from developers.

The competition in the global AI industry has long shifted from mere technical showmanship and conceptual breakthroughs to a comprehensive contest of ecosystem and applications. China’s AI call volume has now surpassed the US for the first time, not only reflecting a “belated” but genuine display of Chinese technological strength, but also signalling a new era of “East-West co-governance” in AI development. 

Unlike US AI, Chinese AI follows an inclusive path of “cost-effectiveness + scenario adaptation”, which better meets the needs of most countries and developers worldwide and better promotes global AI adoption.

Related: The next AI superpower? China’s quiet lead in the race for scale | How China’s all-in-one platforms are reshaping the AI battlefield