China's three major telecom operators accelerate token business as AI economy drives demand surge

As artificial intelligence permeates various industries, the commercialization of the token economy has drawn increasing attention.

The newly released China Token Factory Development White Paper (2026) (hereinafter referred to as the White Paper) indicates that token factories have moved beyond the concept incubation phase and entered a trial stage where project construction, capability procurement, and platform operations are advancing in parallel. However, from an industry perspective, the proportion of projects that are truly operational remains low. The core challenge is also evolving: previously, the focus was on "whether computing power is available"; now, the question is "whether stable, continuous production of cost-competitive effective tokens can be achieved."

This shift is driven by the rapid growth in AI application scale. According to data from the National Data Administration, China's daily token calls stood at approximately 100 billion in early 2024, surged to 100 trillion by the end of 2025, and exceeded 140 trillion as of March 2026. Liu Liehong, Director of the National Data Administration, previously stated that tokens are measurable, priceable, and tradable, serving as a "settlement unit" connecting technology supply and commercial demand. Meanwhile, data from the Ministry of Industry and Information Technology shows that as of the end of June 2026, China's intelligent computing power reached 2,185 EFLOPS, with an overall rack utilization rate of 71.4%.

From "Selling Resources" to "Selling Services"

According to the White Paper, a "token factory" is defined as an industrialized production system that "uses electricity as the energy input, GPU hours as the production process, tokens as the product, and unit effective task cost as the core metric." Notably, the emphasis is not on simply increasing token volume, but on "effective tokens."

Under this definition, only output tokens that simultaneously meet service-level requirements such as latency, availability, and accuracy—and are ultimately seen and used by end users—are considered effective tokens and eligible for commercial measurement.

Industry observers believe that the emergence of the "token factory" concept reflects a shift in how AI infrastructure is evaluated. In the past, after a data center or intelligent computing center was built, investors focused primarily on server, GPU, and facility scale. But in the inference stage, what truly determines commercial value is whether these hardware assets can be utilized continuously, stably, and efficiently. This means AI infrastructure is gradually transitioning from selling "resources" to selling "services."

This trend aligns with the development stage of the computing power industry. As of June 2026, the national computing facility rack utilization rate was 71.4%, with intelligent computing power reaching 2,185 EFLOPS. Faced with growing computing supply, simply expanding hardware scale no longer automatically translates into economic value. Improving computing utilization, reducing unit inference costs, and converting computing power into stable AI services have become key priorities for the next phase of industry development.

At the first Token Cloud Service Conference, the China Academy of Information and Communications Technology (CAICT) proposed that tokens are gradually becoming an important service unit connecting computing supply, large model capabilities, and industry applications. It called for further improvements in service architecture, resource scheduling, metering and billing, quality assurance, and security and trustworthiness standards. The national standard Artificial Intelligence Token Metering Framework and Testing Methods has also been initiated, with participation from organizations including the China Electronics Standardization Institute.

Wang Qinglin, Manager of the Ruidanheng Industrial Research Institute, believes that the impact of token factories may go beyond adding a new type of computing project. It could drive the industry to use multiple effective metrics to measure the return on investment (ROI) of AI infrastructure. For local governments, computing investors, and cloud service providers, this also means that the previous approach of "build first, find demand later" needs adjustment. Whether a project can generate sustained orders will increasingly determine investment returns.

Operators Race to Capture the Entry Point

While token factories remain in the trial phase, telecom operators—who are also actively exploring token operations—have entered a period of rapid product and platform iteration.

According to a recent research report by Omdia analyst Xia Maosen, when the three major operators first launched token packages in May, prices generally ranged from 3 to 5 yuan per million tokens. By July, core package prices had dropped to around 1 yuan per million tokens, approaching the levels of leading domestic cloud service providers. This round of price adjustments indicates that operators are quickly abandoning the strategy of profiting solely from token price margins, and are instead seeking comprehensive value from networks, computing power, models, channels, and industry customer resources.

From a strategic perspective, the three operators have adopted different focuses.

China Mobile emphasizes a platform-based approach. In June, it established a group-level Token Office responsible for integrating the entire process of "creating, delivering, and applying tokens." At the same time, China Mobile began combining tokens with traditional services such as 5G packages, offering multi-model capabilities to individual and enterprise users through an AI distribution platform. Xia Maosen believes that the significance of this move lies in unifying previously fragmented token pilot projects at the group level, reducing coordination costs across different business departments.

China Telecom places greater emphasis on productization. From its "From Bit to Token" value proposition to its Tianyi Cloud Token products and the Xingchen Token Hub, China Telecom is packaging tokens as standardized products for individuals, families, and enterprises. It has launched a personal package of 9.9 yuan/month for 10 million tokens and an enterprise package of 299.9 yuan/month for 150 million tokens. Its Token Hub offers over 100 model options. This approach closely resembles the package-based, product-oriented operations of traditional telecom services.

China Unicom focuses more on industries and scenarios. After a July organizational restructuring, China Unicom incorporated "strengthening token operations" into the responsibilities of its product innovation and marketing department. At its recent partner conference, China Unicom further proposed enhancing the full-chain capability of token "creation, storage, transport, cleaning, and application," and fully opened its "Agent + Token + AI Cloud" operating system.

China Unicom Chairman Dong Xin publicly stated that the company aims to integrate computing supply, model calls, and scenario returns into a single "accounting book," shifting from "transporting traffic" to "supplying tokens," from "transmitting bits" to "empowering intelligence," and from "connectivity" to "productivity." China Unicom also launched Unicom Cloud 8.0, a trusted token infrastructure, and Unicom Yuanjing MaaS, aiming to upgrade tokens from a simple billing unit to a foundational service for industry clients.

More notably, China Unicom is addressing the long-standing "metering black box" in token commercialization. The company introduced the "Yuanzhu Metering and Billing System" to standardize token measurement across different models, where token standards and usage traceability have been inconsistent. Unicom Cloud 8.0 further proposes "six-coordination" of computing power, network, chip, data, model, and security, targeting key scenarios such as government services, healthcare, education, finance, and state-owned enterprises.

The strategies of the three operators indicate that tokens are no longer just a new package for operators. They are redefining the relationship between operators, the AI industry, and end users.

Omdia's assessment aligns with this logic: token services themselves are rapidly commoditizing. In the long term, operators' true opportunity to build differentiated advantages lies not in simply "selling tokens," but in leveraging their existing capabilities—national networks, computing infrastructure, enterprise customers, billing systems, and security—to become the AI infrastructure and delivery layer between models and end users. In other words, tokens may be just the entry point; the real competition is for the AI service orchestration and industry customer service rights behind the tokens.

Three Hurdles to Mature Commercialization

The boom in AI development has provided a clear demand foundation for the token economy. More importantly, token consumption is expanding from C-end chat and Q&A to more complex industry scenarios such as enterprise office work, software development, and industrial production.

The Ministry of Industry and Information Technology has explicitly called for deepening "AI + Manufacturing," promoting the integration of large models into R&D, production, marketing, and operations management. This suggests that as enterprise AI applications gradually enter production systems, token demand is expected to shift from sporadic calls to sustained, stable industrial loads.

However, the path from rapid growth to mature commercialization presents clear challenges.

First, the "metering standard" issue. Although tokens inherently have metering attributes, the computational cost and ultimate value of a token vary across different models, tasks, and context lengths. Wei Kai, Director of the CAICT AI Research Institute, previously noted that the industry should not focus solely on token unit prices. Standards for measuring "high-quality tokens" are needed. This aligns closely with the White Paper's emphasis on "effective tokens": future price competition will likely shift from "cost per million tokens" to "how many tokens and how much computing power are needed to complete a specific task, and what the final outcome is."

Second, the "cost and benefit" issue. Token factories are capital-intensive, with significant upfront hardware and infrastructure investments but relatively low marginal costs. Consequently, if utilization is insufficient, the unit cost of effective tokens can rise rapidly. The White Paper specifically notes that utilization rates have a greater impact on unit costs and project returns than hardware purchase prices. The same hardware can yield significantly different profitability when handling different workloads. In other words, future computing projects cannot simply compete on "who builds the biggest," but on "who can keep the computing power fully utilized over the long term."

Third, the "value realization" issue. For enterprise customers, purchasing tokens is not the end goal. The real business outcome is reducing R&D costs, improving customer service efficiency, optimizing production processes, or enabling agents to complete tasks. Therefore, token operations may eventually evolve from "pay-per-call" to "pay-per-task, pay-per-outcome, or pay-per-service-level."

Huang Wei, Deputy Director of CAICT, publicly noted that tokens have given AI services a metering foundation similar to electricity or data traffic for the first time. However, moving from "meterable" to "tradable" and "scalable" still requires improvements in standards, service quality, and business rules.

In the near future, the token industry is likely to remain in a phase of "rapid growth" coexisting with "model experimentation." Token factories will address large-scale production; operators and cloud service providers will handle supply, scheduling, and delivery; and model and agent companies will continue to extend upstream and downstream along the application value chain. As all segments of the industry gradually form unified metering standards, tokens may truly evolve from a number on a large model API bill into a foundational commercial unit connecting computing power, data, models, and the physical economy.

(Source: China Business Network)