From "selling data traffic" to "selling Tokens," the year 2026 has witnessed a collective shift for China's three major telecommunications operators. With the explosive growth in demand for AI computing power, the nation's daily Token call volume has surged from the hundreds of billions at the start of 2024 to 180 trillion. Tokens are becoming the new metric for operators to redefine their own value.
In March 2026, China Telecom Corporation Limited (ASX: 601728) was the first to propose Token-based operations, aiming to reshape its business system. At the "2026 World Artificial Intelligence Conference (WAIC)," the "Token Economy" became a core topic for operator transformation. From production and scheduling to networking, what kind of full-chain innovative solutions will China Telecom deliver? Standing at this turning point for business transformation, could the Token economy become the next traffic economy?
Entering a New Phase of Token Operations
"The essence of Token operations is to provide artificial intelligence services to customers," stated Ke Ruiwen, Chairman of China Telecom, during the "Stars Coexist, Intelligence Benefits Partners" AI Ecosystem Forum at WAIC 2026. He emphasized that the "Five-in-One" intelligent cloud system vigorously promoted by China Telecom represents the tangible implementation of Token operations.
Taking "Zhiyun Shanghai" as an example, since its launch in May, user numbers have grown rapidly, effectively driving swift growth in Token consumption. In the era of the Token economy, the evaluation criteria for computing power network operational efficiency are undergoing a change.
In the view of Hu Zhiqiang, General Manager of Tianyi Cloud, the past evaluation of computing power networks focused solely on utilization rates, whereas now the focus is on "Token production capacity." Effective Token production capacity is inseparable from a robust technological foundation. To enhance this core metric, China Telecom has achieved breakthroughs in three key areas: underlying software, inference engines, and cluster scheduling.
"Our Token service scale has continued to experience high-speed growth over these past months," Hu Zhiqiang noted. To address customer pain points in Token consumption—such as difficulty in model selection, cost control challenges, and the complexity of model access and billing—China Telecom has upgraded the operational capabilities of its one-stop Token operation service platform.
Feng Wen, Chief Architect at MiniMax, shared that MiniMax's flagship models have been fully adapted to the Xiran domestic computing power and distributed inference network. The Xingchen TokenHUB can achieve unified access, intelligent scheduling, security auditing, and cost control, enabling government and enterprise clients to efficiently, stably, and cost-effectively call MiniMax's flagship models, thereby accelerating the large-scale deployment of large models in critical scenarios such as government affairs, industry, finance, and R&D.
At the China Telecom exhibition booth, one visitor remarked after an exchange, "Previously, accessing computing power required dealing with several different cloud providers separately. Now, one platform handles it all—one Token account, one set of APIs. How much you've spent and used is clear at a glance."
The Next Traffic Economy?
Looking back at the traffic economy era, its start relied on infrastructure development and transparent pricing. In 2026, the Token economy seems to be replaying this path: computing power infrastructure is taking initial shape, service platforms are unifying measurement, and pricing exploration is on the agenda.
But history does not simply repeat itself. The traffic economy once became mired in price wars that increased volume without increasing revenue. If the Token economy wishes to avoid repeating this mistake, it must answer a core question: How can Tokens transition from being "consumables" to becoming "value carriers"?
Yu Xiaohui, President of the China Academy of Information and Communications Technology, offered a key term for breaking this impasse during the forum: "Token Intelligence Density." He pointed out that while large models have made significant progress in complex reasoning and long-range tasks over the past year, against the backdrop of constrained chip supply, improving "Token Intelligence Density" is the key to achieving high cost-effectiveness.
In Yu Xiaohui's view, OPC (One-Person Companies) and intelligence-native approaches will be the future direction of the Token economy. Although intelligence-native is still in its early stages, and the mature forms of intelligent agents across various fields are far from arriving, this precisely represents a window period to avoid homogeneous competition.
He Zhongjiang, Chairman of China Telecom AI Technology Company, mentioned that from Q&A dialogues to task execution, statistics show that intelligent agent users consume approximately 10 million Tokens per day, which is about 1000 times that of traditional conversational software. As an implementation example, the internal test version of China Telecom's Xingchen Super Intelligent Agent, TeleAgent, already boasts over 300,000 users, with daily Token consumption exceeding 200 billion.
However, high consumption does not equate to high value; users care more about task completion capabilities in office scenarios. He Zhongjiang revealed that TeleAgent's excellence rate is 20% higher than that of competitors, while its Token consumption is only 50% to 80% of competing products.
As enterprises connect to an increasing number of large models, new concerns are emerging: How are costs accounted for? How are permissions managed? How is sensitive data protected? In the AI security demonstration area, staff switched between different large models, and the on-screen display of model call chains, Token usage, calling costs, and permission policies updated accordingly.
Hu Zhiqiang stated that providing intelligent model routing, unified Token accounting, and establishing a secure and trustworthy service system are the main approaches to solving the aforementioned problems. This includes the localization of inference frameworks and chips, supporting Token quota allocation, recycling, and hierarchical permission management.
Comments