AI-related business concept illustration Photo: VCG

Editor's Note:

Currently, China's economy is steadily advancing along the path of high-quality development, even as domestic and international circumstances become increasingly complex. Some Western media, due to misunderstanding or bias, have repeatedly questioned or even distorted China's economic development. Accordingly, the Global Times launches the "Q&A on China's Economy" column to publish opinion pieces to present facts and clarify perceptions.

Recently, a set of figures on the share of token consumption by large artificial intelligence (AI) models has drawn considerable public attention in China. According to the figures, short dramas and video generation account for 55 percent of domestic token consumption, e-commerce and marketing 24 percent, and software development only 15 percent. Some have contrasted this with the claim that "58 percent of US tokens go to writing code and enterprise automation," giving rise to the catchy narrative that "China uses computing power to chase entertaining short dramas, while the US uses it to build the future."

Set against the current international discourse surrounding China-US competition in AI, the comparison is certainly attention-grabbing. Yet it does not withstand even the most basic fact-checking or scrutiny. The claim that "AI short dramas and video generation account for 55 percent" originated from an industry assessment, which circulates online and is contained in the investor-relations records of a listed company that is itself heavily invested in AI-generated short dramas. It is not a comprehensive calculation based on any authoritative statistical framework. Doubao has also recently clarified that the computing-power figures are "somewhat unreliable" and that, at ByteDance itself, only about 10 percent of the company's total compute is used for training and inference of video-generation models. As for the claim that "58 percent of the US token go to writing code," no authoritative source for the figure can be found at all.

Tokens are units used to measure and bill AI model usage, reflecting workload and computing consumption. They are neither a measure of productivity nor an equivalent of GDP, and therefore cannot be used directly to gauge the value of a country's AI industry. Meanwhile, video generation is a computationally intensive use case, while code generation is relatively low in token intensity but highly efficient. Comparing the two simply on the basis of token consumption is therefore methodologically unsound.

What matters is not how many tokens are consumed, but what value they create. China has the world's most complete and largest industrial system, giving AI an exceptionally rich range of application scenarios and the world's largest real-world testing ground. Short dramas, e-commerce, livestreaming and content generation may not fit the conventional image of cutting-edge, "hard-core" technology. Yet it is precisely these everyday applications that place AI technology on the front line, where it is tested by real users and repeatedly refined through viable commercial cycles. A video-generation model used thousands or tens of thousands of times in short-drama production must meet real requirements for visual consistency, production deadlines and cost efficiency. This model has far more opportunities to be updated, iterated and improved - and can generate far greater industrial value-than a technology used only by a small number of engineers or one that exists largely in PowerPoint presentations.

"Mass-market" does not mean "low-value," and mass adoption should not be set against frontier research. The scale of application itself helps fund research and points it toward the problems that matter. Large-scale use spreads inference costs across more users, helping transform model services from expensive laboratory tools into affordable public goods. In China, some video-generation models now cost less than one yuan per second, while AI short-drama production costs have fallen from several thousand yuan per minute to only a few hundred. Lower costs bring more small businesses and individual creators into the market, expanding the range of applications. In turn, the engineering challenges exposed by large-scale use help shape research priorities. China's strengths in open-source models, lightweight deployment and low-cost inference have, to a large extent, grown out of the demands of serving AI at scale.

China's AI development is by no means focused only on applications at the expense of technological research.  A 2024 report by the World Intellectual Property Organization (WIPO) already showed that China leads the world in the AI research publication volume, citations counts and patent output, and is world's largest source of generative AI patent publications. Chinese AI models have led weekly usage on OpenRouter, the world's largest AI model API aggregation platform, for 26 consecutive weeks, accounting for roughly two-thirds of total usage. The share of tokens used for coding tasks on the platform rose from about 10 percent in February to around 50 percent in September this year. China has never been absent from the development of foundational capabilities or frontier models. More importantly, it has not kept these advances locked inside laboratories. Instead, it has pursued development on two fronts, allowing research and development to interact with and reinforce the world's largest pool of real-world demand.

There is little reason to fuel anxiety about "China falling behind the US in AI" with questionable data. Turning statistics based on inconsistent methods and unclear sources into a simplistic contrast between "building the future" and "chasing short dramas" risks misleading the public and politicizing normal technological diffusion and industrial specialization. Large-scale application is, in fact, one of the strongest foundations of China's AI development. By closing the loop between research and real-world deployment, China is pursuing its own path while also creating greater room for the global diffusion of AI technologies and their industrial applications.