Companies increase investment in the "morpheme economy" industry chain layout

The data show that as of this March, our country’s average daily Token usage has already exceeded 1.4 trillion, a more than 1,000-fold increase over two years. The surge in Token usage not only confirms that AI application scenarios are continually deepening, but also creates new opportunities—relevant companies are accelerating their deployment along the “Token economy” industrial chain such as high-performance compute supply, Token operations services, and high-quality datasets.

So-called Tokens are the smallest information unit processed by large models—for example, in everyday work and life, every time you ask AI questions or use AI to generate content is calling Tokens. Especially this year, the breakout of “agents” such as “lobsters” has driven a significant increase in Token consumption.

“Since late January this year, some model companies have set performance records in which their revenues in 20 days surpassed their full-year revenue for 2025. Behind these numbers, a new business logic based on Token-based billing is accelerating its evolution.” Liu Rihong, Director of the National Data Administration, said at the 2026 Annual Conference of the China Development Forum recently. Around the calling, distribution, and settlement of Tokens, a new value system is accelerating into formation and has become an important path for monetizing the AI industry.

The surge in Token usage is behind the consumption of compute power, the running of algorithms, and the investment of electricity, which also puts higher demands on compute power. Recently, many companies have been speeding up innovative deployments for high-performance compute, and architectural innovations represented by supernodes are an important path to improving compute efficiency.

For example, Aoshushang Co., Ltd. (Tsinghua Unisplendour) recently rolled out its first wireless-cable container-type supernode, scaleX40. In typical inference scenarios, the inference throughput for the same number of cards improves by more than 4 times, significantly enhancing unit compute power’s Token output capability; ZTE Corporation has also launched supernode technology. By reconstructing the compute interconnect system, it integrates dozens to hundreds of multi-vendor GPU “compute into a unified compute unit, achieving system-level optimization of compute power.

In the view of industry insiders, the core innovation in the compute industry will focus on reducing effective Token costs. Deep coordination between compute and applications, full-stack optimization across software and hardware, and ecosystem coordination across the entire industrial chain will become key priorities. ZTE’s Strategic and Ecosystem Chief Expert Tu Jiashun said that the rollout and promotion of supernode technology will drive intelligent computing infrastructure to evolve toward higher efficiency, greener operations, and greater openness.

“Token surges demand higher requirements for compute density, memory-access bandwidth, communication efficiency, and more. This is pushing compute power to shift from pursuing single-card peak performance to system-level collaborative optimization of the ‘memory—bandwidth—interconnect.’” Lu Feng, Director of the Beijing Frontier Future Technology Industry Development Research Institute, told reporters. “The ‘Token economy’ will drive underlying compute infrastructure toward higher energy efficiency development.”

Given their measurable, tradable, and exchangeable characteristics, Tokens also serve as a “settlement unit” connecting technology supply and business demand. In recent times, the three major telecom operators have all said they will further explore “Token operations services.”

“In the intelligent era, to build a new form of intelligent economy, we must accelerate the transformation and upgrading from ‘traffic operations’ to ‘Token operations.’” Liu Guiling, General Manager of China Telecom, said. China Telecom’s Token operations already have preliminary practice. Taking an example of a certain company’s AI private deployment on the China Electric Information Terra platform, China Telecom custom-developed 73 agents, driving annual consumption of 1.2 trillion Tokens.

In addition, China Unicom will accelerate building a compute operations model of “agents + Tokens + AI cloud.” China Mobile said that by integrating high-quality AI models, it builds trusted inference services and links the complete service chain: “using Tokens for agents, with Tokens pulling compute power.” The Token market is quickly opening up.

The large increase in average daily Token usage indicates that dataset supply is also increasing substantially. The enablement of data elements for innovation-driven AI development has entered a stage of benign interaction. The reporter learned from the National Data Administration that by the end of 2025, more than 100,000 high-quality datasets had already been built nationwide, with a total size exceeding 890PB (petabytes; a unit for computer storage capacity), roughly equivalent to about 310 times the total amount of digital resources of China’s National Library.

“Data are processed and applied by AI after being broken down into Tokens, providing solid support for model iteration and application deployment.” Lu Feng said. As Tokens become a standardized pricing and circulation unit for AI capabilities, value across the industrial chain will also be redistributed to each optimization node across the “Token full lifecycle.” Among them, in the data services segment, high-quality Token supply will become a high-value resource. It also has the potential to drive the development of independent commercial products such as prompt engineering, Token compression, and vertical Token libraries.

It is reported that next, the National Data Administration will coordinate with all parties to deeply implement a new round of actions for building high-quality datasets. Guided by scenario needs, it will create AI-Ready (AI readiness) high-quality datasets that are technically feasible, practically convenient, and quality-assured, achieving both a quantitative and qualitative improvement in high-quality dataset supply.