An AI factory exhibit at the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance held in Shanghai on July 29. Photo: IC PHOTO

At the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance, Chinese President Xi Jinping noted that as a new engine of world economic growth and an accelerator for the shift of growth drivers, AI is moving from the digital world into the physical world. The observation highlights AI’s significance for the global economy while also capturing the stage the technology has now reached and the emerging consensus surrounding its development. Over the past year, the global user base for AI products has expanded rapidly, bringing the technology within reach of a much broader public. At the same time, AI applications have begun to move beyond generating digital content to taking part in real-world production, deepening their integration with the real economy. The focus of competition has shifted accordingly—from computing power and language-model performance in the digital realm toward vast stores of data and real-world application scenarios in the physical realm.

Moving toward universal accessibility

AI products built around large language models (LLMs) are advancing and expanding at an unprecedented pace. By early 2026, ChatGPT had reached 900 million weekly active users, representing annual growth of more than 100%. QuestMobile data also showed that, by the first quarter of 2026, the native apps of Doubao, Qwen, and DeepSeek in China had reached 340 million, 170 million, and 130 million monthly active users, respectively, with both user numbers and engagement continuing to rise. AI products are growing faster than any previous generation of internet products.

Behind this rapid growth is a steep decline in the cost of using AI. According to Stanford University’s “The Artificial Intelligence Index 2025,” the cost of using models that deliver comparable inference performance fell over two years to just 1/280 of its earlier level, while the performance gap between open-source and closed-source models narrowed sharply. The drop in price has put advanced models within the financial reach of individual developers and small- and medium-sized enterprises, further enlarging the user base. Declining costs and increasing adoption have reinforced each other, helping turn AI into an everyday service accessible to the general public.

The spread of LLMs is also changing how economic activity is measured and transactions are conducted, giving rise to the increasingly prominent “token economy.” A token is the smallest unit into which text is divided for model processing, and the number of tokens produced and consumed can be used to measure the scale of intelligent services. Pricing and trading mechanisms built around tokens are also beginning to take shape. Tokens, however, are simply a technical byproduct of combining the Transformer architecture with tokenization mechanisms and are specific to LLMs. The pricing models for large models are themselves evolving. Outcome-based, task-based, and per-agent-seat pricing have already emerged, while newer directions such as world models and embodied intelligence do not inherently rely on discrete tokens as units of measurement or billing. The significance of the token economy therefore lies not in tokens themselves, but in what they reveal about an emerging economic form: For the first time, intelligence has become a commodity that can be measured, priced, and traded, much as electricity enters the market in kilowatt-hours. The unit of measurement may change, but the commodification of intelligence appears increasingly irreversible.

A clear gap nevertheless remains between the breadth of AI adoption and the depth of its application. According to a global survey released by McKinsey in 2025, most surveyed companies had already adopted AI in at least one business function, but only a small minority had achieved large-scale deployment. In practice, many companies have rolled out numerous AI applications without seeing a corresponding improvement in business performance. So far, adoption has been concentrated mainly in individual use and workplace assistance. The deeper integration of AI into enterprise production and operations—and its conversion into measurable productivity gains—remains at an early stage.

Integrating with the real economy

Moving from widespread use to deep application will depend on AI’s ability to make the transition from an information tool to a production tool. Earlier forms of AI were used mainly for information retrieval, content recommendation, and image recognition, serving the acquisition and circulation of information. With the emergence of generative technologies, models can directly produce text, code, and design solutions, giving them for the first time the capacity to take on productive tasks. Yet this capability has so far been realized chiefly in the realm of digital content, while the core operations of many enterprises take place on factory floors, construction sites, and in fields—not in text. To participate meaningfully in production, AI must move beyond the digital world and into the physical one.

The trajectory of technological development points in the same direction. The supply of publicly available online text that can be used for training is approaching exhaustion, making it increasingly difficult to keep improving model capabilities through existing approaches alone. Turing Award laureate Richard Sutton has argued on this basis that the path of training machines primarily on existing human knowledge is approaching its limits. In the future, AI will need to learn through sustained interaction with real environments, entering what he calls the “era of experience.” Sutton has also cautioned that AI as it actually exists today remains weak and unreliable, far removed from the powerful image many members of the public have formed of it. Closing that gap, he argues, likewise requires exposing AI to the challenges and feedback of real-world environments.

The technical groundwork for engaging with the physical world is also advancing rapidly, particularly in two areas: world models and embodied intelligence. The former are intended to help machines understand the physical world, while the latter enable machines to act within it. At the core of a world model is prediction: Based on the current state of an environment and the action about to be taken, it infers what is likely to happen next. Prediction and judgment, however, can contribute to production only when they are translated into action. That requires giving intelligence a body capable of perception and action—precisely the direction being explored by embodied intelligence. In some industrial settings, the reliability of robotic operations is already approaching the level required for commercial deployment. As a result, AI is beginning to enter production in practice and take on real-world operational tasks.

Highlighting data and scenarios

When model capabilities can be obtained through open-source platforms or low-cost application programming interfaces, data and real-world scenarios become the scarcer resources. Interactive data from the physical world cannot simply be scraped from the internet; it must be accumulated gradually through actual production, at a cost far higher than that of acquiring textual data. This has given rise to a widely held view in the robotics industry that “models determine the starting point, while data determines the endpoint.” Autonomous driving offers a case in point. Companies entered the field with broadly similar algorithms, but what ultimately separated them was the accumulation of road-testing mileage and real-world driving data.

In the competition over data and application scenarios, China’s advantages are increasingly coming to the fore. China is the only country that possesses all industrial categories listed in the United Nations industrial classification system. Its manufacturing value added accounts for roughly 30% of the global total, and almost every type of production scenario can be found domestically. From port logistics and mining operations to power-grid inspections, intelligent upgrading is already underway across multiple industries, giving AI developers an exceptionally wide range of testing grounds for practical deployment. According to the International Federation of Robotics, roughly half of all newly installed industrial robots worldwide each year are deployed in China. China’s leading stock of industrial robots has made it the market where physical-interaction data is accumulating most rapidly. Its enormous domestic market also allows new applications to acquire users quickly and spread research and development costs across a much broader base. The U.S.-China Economic and Security Review Commission has argued that China and the United States have chosen two different development paradigms. The United States has concentrated resources on expanding computing capacity and training frontier models, while China has used open-source approaches and cost advantages to promote widespread model deployment, relying on data generated in the course of deployment for continuous iteration. As competition shifts toward data and application scenarios, the advantages of China’s approach are becoming increasingly apparent.

To realize these advantages fully, however, China still needs to address shortcomings in the institutional framework governing data. China has been at the forefront of institutional experimentation in this area, having moved relatively early to recognize data as a factor of production alongside land, labor, capital, and technology. Since 2024, enterprises have also been permitted to record data resources as assets on their balance sheets. In practice, however, data generated in production settings remains scattered across different industries and enterprises. Rules governing data rights, circulation, and security management are still incomplete, leaving large volumes of data undeveloped. The opening of public data and the interconnection of industry-specific datasets are likewise only beginning. Greater access to application scenarios must accompany greater access to data: New technologies need places where they can be tested and conditions that allow them to improve through iteration. This requires governments and enterprises to work together to establish institutional arrangements that allow data to circulate under clearer rules and application scenarios to be opened in a more orderly manner.

Zhong Zhou is an associate research fellow from the Institute of Quantitative and Technological Economics at the Chinese Academy of Social Sciences.