The figures were contrasted with another widely circulated claim that 58 per cent of AI tokens in the United States are used for coding and enterprise automation.
The comparison quickly fuelled a narrative that China was directing its computing resources towards entertainment and consumer content, while the United States was concentrating on software development and technologies shaping the future.
However, a closer examination of the figures raises questions about their reliability, methodology and suitability for measuring technological progress.
According to the assessment, the claim that short dramas and video generation consume 55 per cent of China’s AI tokens originated from an industry estimate circulated online and included in the investor-relations records of a listed company with significant investments in AI-generated short dramas.
It was not based on a comprehensive statistical survey or an authoritative national measurement framework.
ByteDance’s AI chatbot Doubao has also reportedly questioned the reliability of the computing-power figures. The company indicated that approximately 10 per cent of its total computing resources were used for training and inference involving video-generation models.
Meanwhile, no authoritative source has been identified to substantiate the claim that 58 per cent of US AI token consumption goes towards coding and enterprise automation.
Why token consumption cannot measure AI leadership
A central argument in the debate concerns the meaning of tokens, which are units used by AI systems to measure and bill model usage.
Token consumption reflects the volume of processing undertaken by an AI model and provides an indication of computing workloads. However, it is not a direct measurement of productivity, economic output or technological advancement.
Different AI applications also consume computing resources at significantly different rates.
Video generation, for example, requires substantial computational resources to produce visual content, whereas coding tasks can often be completed with considerably lower token consumption.
Consequently, comparing the two sectors solely on the basis of token usage can produce misleading conclusions about their economic importance and technological sophistication.
The broader argument is that the value created by AI matters considerably more than the number of tokens consumed.
China’s industrial ecosystem provides a large testing ground
China’s extensive industrial base and enormous consumer market provide AI developers with a wide range of commercial applications and opportunities to test their technologies under real-world conditions.
Applications such as short-video production, e-commerce, livestreaming, digital marketing and automated content creation may not always be associated with frontier technological research. Nevertheless, they expose AI systems to substantial commercial demands and large numbers of users.
For instance, AI video-generation models used extensively in short-drama production must meet practical requirements involving visual consistency, production schedules, creative quality and cost efficiency.
Repeated commercial use creates opportunities for developers to identify weaknesses, improve performance and refine their models.
The argument is that technologies tested and improved through thousands of real-world applications can generate meaningful industrial value, even when their uses appear less technically sophisticated than specialised software-development applications.
Mass-market adoption, therefore, should not automatically be equated with low-value technological activity.
Lower costs are expanding AI adoption
Large-scale commercial deployment is also helping reduce the cost of accessing AI technologies.
As more users adopt AI services, inference costs can be distributed across a larger customer base, making advanced tools more affordable for smaller companies, independent developers and individual creators.
According to the assessment, some Chinese video-generation models now cost less than one yuan per second to operate, while AI-generated short-drama production costs have fallen from several thousand yuan per minute to a few hundred yuan.
The reduction in costs is expanding opportunities for smaller businesses and individual creators to experiment with AI-powered production.
At the same time, widespread deployment is exposing developers to engineering challenges that can influence future research priorities.
China’s progress in open-source models, lightweight deployment and low-cost inference has been closely connected to the need to provide AI services efficiently at scale.
This creates a cycle in which commercial adoption supports technological improvement, while better technology encourages further adoption.
China’s research capabilities remain significant
The argument that China prioritises commercial applications at the expense of fundamental research is also challenged by its research and innovation record.
A 2024 report by the World Intellectual Property Organization (WIPO) identified China as a leading contributor to artificial intelligence research publications, citations and patent output. It also highlighted China’s position as the world’s largest source of generative AI patent publications.
Chinese AI models have also recorded substantial usage on OpenRouter, an international platform that aggregates access to different AI models through APIs.
According to the figures cited in the assessment, Chinese models led weekly usage on the platform for 26 consecutive weeks, accounting for approximately two-thirds of total usage.
The share of tokens used for coding tasks on the platform reportedly increased from around 10 per cent in February to approximately 50 per cent in September.
These figures suggest that Chinese AI models are increasingly being used for technical and software-development tasks alongside entertainment, content generation and commercial applications.
China’s approach, therefore, is not limited to deploying existing technologies in consumer-facing industries. It also involves developing foundational models and improving technological capabilities while connecting research with large-scale commercial demand.
A different approach to AI development
The debate over token consumption highlights a broader question about how countries should measure progress in artificial intelligence.
Comparing national AI capabilities through figures drawn from different sources and inconsistent methodologies risks creating an inaccurate picture of technological competition.
The distinction between developing advanced AI systems and deploying them across industries is also increasingly difficult to maintain, as real-world applications can provide valuable feedback for improving foundational technologies.
China’s large industrial ecosystem, consumer market and expanding commercial adoption offer opportunities to connect research, engineering and business applications.
Rather than treating mass-market AI adoption and frontier research as competing priorities, the country’s development strategy seeks to strengthen both.
Ultimately, the number of tokens consumed by a particular industry cannot independently determine which country is leading the global AI race. A more meaningful assessment requires examining research quality, model performance, innovation, commercial adoption, productivity improvements and the economic value generated by AI.
The wider message is that large-scale application should not be dismissed as a sign of technological weakness. When supported by research and continuous innovation, it can become an important foundation for advancing artificial intelligence and expanding its benefits across the global economy.
