From counting tokens to creating value: China's AI path

Editor's note: Ma Xiaobai is director and a research fellow at the Multinational Corporations Research Office, Enterprise Research Institute, Development Research Center of the State Council. Chen Gong is a lecturer and PhD at the Education and Training Center for Officials and Entrepreneurs, State-owned Assets Supervision and Administration Commission of the State Council. This article reflects the authors' opinion and not neccesarily those of CGTN.

Artificial intelligence (AI) is emerging as a strategic and foundational technology in China. From exploring pathways toward artificial general intelligence (AGI) to advancing the "AI Plus" initiative, China has continued to strengthen policy support, integrating AI into industrial and social sectors to drive digital and intelligent transformation.

The National Data Administration is also advancing work on a token-based value system and exploring new models, including token trading.

The scale of AI activity is already striking. China's daily token usage has surged from 100 billion at the beginning of 2024 to more than 140 trillion as of March this year. Tokens are becoming a new indicator of activity in the intelligent economy.

But there is a more important question behind that number: Does using more tokens necessarily mean creating more value? To answer that, we need to look at how different AI development paths are taking shape around the world.

The US has clear advantages in computing power, chips and capital. Leading companies have placed greater emphasis on "scaling laws," continuously pushing the performance ceiling of foundation models by investing more in parameters, data and computing power. This approach has helped push the boundaries of what AI models can do. But it also comes with a huge bill: Higher costs for training and inference, as well as growing energy consumption.

Chinese AI companies are also racing to develop models with cutting-edge capabilities. But alongside raw performance, they are putting greater emphasis on a different equation: capability, cost and usability.

Algorithmic innovation, sparse activation, engineering optimization and open-source ecosystems are helping lower the barriers to AI adoption. The rapid iteration of models such as DeepSeek, Kimi and GLM reflects this trend.

On August 26, Zhipu launched and open-sourced GLM-5.3-Flash. Artificial Analysis, a third-party AI evaluation organization, gave the model an Intelligence Index score of 57, well above the median of 28 for comparable open-weight models, indicating a favorable balance between capabilities and cost.

And this points to something bigger.

The competition among large AI models is no longer simply about who can make the most powerful model. It is increasingly about who can make that intelligence affordable, efficient to run and scalable enough to be deployed widely.

Of course, the AI development paths of China and the US cannot simply be reduced to a question of which one is better.

The more important question is whether these different technological approaches can eventually support sustainable business models.

Global investment in AI is heating up, but so are concerns about excessive spending and bubble risks. Whether there is a bubble cannot be judged by capital expenditure or model valuations alone. The real test is much more straightforward: Is AI actually making the economy more productive?

If massive investment only produces higher rankings, bigger models and impressive demos, the commercial prospects of AI will inevitably come under scrutiny. But if AI can raise labor productivity, reduce social costs, create new products and generate new demand, then what looks like massive spending today could become the infrastructure of tomorrow.

This is where China's particular strengths in AI development come into play. China has a comprehensive industrial system, a huge market and an unusually wide range of real-world application scenarios.

For consumers, intelligent assistants are rapidly making their way into smartphones, automobiles and home devices.

In cultural tourism, personalized itineraries, intelligent tour guides, multilingual services and visitor-flow forecasting are reshaping how people experience travel.

In finance, intelligent customer service, risk identification, compliance reviews, and research and investment assistance are becoming increasingly integrated into everyday business processes. And perhaps nowhere is the transformation more tangible than in industry.

AI is already being applied to intelligent quality inspection, equipment failure prediction, production scheduling and digital twins. These are not just demonstrations of what AI can do. They are tests of what AI is actually worth. That is why the key to turning tokens into real value is to put technological investment against actual output. Have companies reduced costs and energy consumption? Has product quality improved? Have R&D cycles become shorter? Have new sources of revenue and employment emerged?

Token usage can tell us how active the AI economy is. But productivity created per token may tell us much more about the quality of AI commercialization. This may be the more meaningful race ahead.

China's AI sector needs to move from competing over parameters to competing over applications, and from counting tokens to creating value. That means balancing fundamental innovation with industrial empowerment, combining cost reduction with open ecosystems, and matching rapid technological development with agile governance.

The real sign that AI has crossed the threshold into commercialization is not that machines can generate more content, write longer answers or score higher on benchmarks. The real question is whether AI can deliver higher productivity across more industries at a cost sustainable for businesses and society. Ultimately, the value of intelligence is not measured by how much a machine can produce, but by how much better we can shape the real world.