From "Data Traffic" to "Token": How AI Is Becoming a Utility Like Water and Electricity

Over the past decade, the fundamental unit of measurement for the digital economy was "data traffic."

Now, in the AI era, a new unit is entering the everyday vocabulary of businesses and individual users: the Token. In Chinese, it has been officially translated as "词元" (cí yuán).

Recently, China's three major telecom operators have started selling Token packages. China Telecom has launched a "Token + Connectivity + Security" service for enterprises, starting at 39.9 RMB per month. Shanghai Telecom has gone a step further—offering 250,000 Token credits for just 1 RMB, payable via phone bill, granting access to over 30 mainstream large language models. China Mobile and China Unicom are also piloting similar services in multiple regions. AI computing power is transforming from a professional cloud resource into a basic service that everyday users can understand and purchase.

Behind this is not merely an innovative pricing plan, but a larger shift: computing power is being Tokenized, and enterprise AI services are entering a phase of usage-based metering.

If "from data traffic to Token" describes the evolution of telecom operators' business models, then "AI as a utility" is an even more vivid metaphor. AI capabilities are beginning to exhibit the fundamental characteristics of basic services like water, electricity, and data connectivity—standardized metering, pay-as-you-go purchasing, package subscriptions, unified billing, elastic usage, and low-barrier access.

Underpinning this analogy is the accelerated national strategy for integrated computing power. With the deepening of the "East Data, West Computing" project and the construction of a national integrated computing network, computing power is moving from scattered resource pools to a cross-regional, uniformly dispatched infrastructure network. By introducing Token packages, operators are essentially providing standardized "meters" and "access valves" to the market, built upon the gradually established computing power "pipeline network." For enterprises, the significance of this change is not merely that "AI has become a little cheaper," but that AI will increasingly resemble a routine operational resource, integrated into budgeting, cost management, access control, compliance, and performance management systems.

I. From "Buying Cloud Resources" to "Buying Tokens": The Changing Metering of AI Services

In the past, enterprises typically accessed AI capabilities in a few ways: purchasing cloud resources (GPU servers, computing clusters, cloud hosts, or inference services); buying model APIs (paying by calls, input/output Tokens, or concurrency); or procuring specific applications (smart customer service, office assistants, marketing content tools, knowledge base Q&A, or industry-specific LLM solutions).

While these options are relatively clear for large enterprises and technical teams, they pose significant barriers for small and medium-sized businesses. Companies need to understand models, computing power, deployment methods, API calls, data security, cost forecasting, and a host of other issues.

The advent of Token packages essentially repackages these complex capabilities. A business no longer needs to first understand GPU instance hours, inference concurrency, or context length. Instead, it starts with a more intuitive unit: How many Tokens do I have this month? How much text generation, code assistance, customer service conversation, or marketing content does that support?

This closely mirrors the telecommunications industry's shift from "per-minute billing" to "data packages." Users don't need to understand network scheduling, but they know how much data they've bought, how much they've used, and how overage charges work. Tokenization is essentially transforming AI capabilities into service quotas that are easier to procure, allocate, measure, and manage.

Of course, making AI a "utility" requires not just power plants and water treatment facilities, but also distribution networks and standardized metering. The national integrated computing power initiative is precisely driving such a cross-regional, cross-entity dispatching system—allowing computing resources in western China to seamlessly support model calls in the east, and enabling computing pools from different cloud vendors to be metered and traded under unified standards. The emergence of the Token means this computing power network finally has a "unified measure" for end-users.

II. Are Tokens Universal? Can Telecom Packages Directly Connect to APIs like Kimi or Doubao?

This is a common point of confusion. Since Shanghai Telecom's public information mentions the Kimi K2.5 model as an example, and China Telecom's packages mention access to their own and third-party computing resources, enterprises naturally ask: If I buy a Token package from a telecom operator, can I directly call models like Kimi, Doubao, Tongyi, or DeepSeek? Are these Tokens interchangeable?

The answer requires a layered perspective.

First, technically, "aggregated calling" is achievable, but only if the operator's platform has already integrated the corresponding models. If an operator builds a unified model aggregation gateway, after purchasing Token credits, an enterprise can obtain an API key from the operator (e.g., China Telecom or Tianyi Cloud) and use the unified API interface to select different models from the operator's pool. The operator's platform then dispatches the call to its integrated models and consumes credits according to the platform's rules.

Public reports indicate that Shanghai Telecom users, after purchasing credits, can call over 30 mainstream LLMs via a standard API interface. China Telecom's relevant packages mention integrating the operator's own星辰 (Xingchen) LLM and other domestic models like GLM5.

Therefore, if Kimi, Doubao, or other models are included in the operator's model list, enterprises could potentially call these models through the operator's unified entrance. However, the key point is: this is platform-level aggregated calling, not inherent interoperability between the official API accounts of different models.

If an enterprise buys API credits directly from the Kimi Open Platform, that's a settlement relationship between the enterprise and the Kimi platform. Similarly, calling Doubao directly via Volcanic Engine is a settlement with Volcanic Engine. Thus, a telecom operator's Token package is more accurately described as an AI service quota within that operator's platform. It can aggregate multiple models through the platform, but it does not mean Tokens are natively interchangeable across all major LLM official APIs.

Second, Tokens are not entirely equivalent units across different models. Although everyone uses the term "Token," tokenization methods (especially for Chinese), input/output pricing, context length, caching mechanisms, multimodal pricing, and tool use fees can all differ. The same Chinese passage might be segmented into a different number of Tokens by different models.

Therefore, enterprises shouldn't just compare "how many Tokens per Yuan." They need to ask more specific questions: Which specific models and versions are supported? Do they meet business needs? How are inputs and outputs charged? Are multimodal and tool use features extra? Is the interface stable? Where does the data go? Does it meet enterprise security and compliance requirements? These are crucial questions when procuring Token packages.

III. The True Value of Operator Tokens: Lowering Access Complexity, Not Replacing Model Vendors

From an enterprise perspective, the value of an operator's Token package may not lie in replacing the official APIs of model vendors like Kimi, Doubao, Tongyi, or DeepSeek. Its greater value likely lies in providing a relatively low-barrier aggregated entry point.

Directly accessing multiple model platforms typically requires enterprises to separately register accounts, apply for API keys, sign service agreements, manage top-ups, read interface documentation, handle different data formats, and manage various bills and access controls. For teams with limited technical resources, this is substantial work. If an operator's Token package can offer a unified entry point, unified API, unified bill, and unified service support, it significantly lowers the barrier for enterprises to trial AI capabilities.

Within the framework of the national integrated computing power strategy, operators are playing a role in integrating the computing network and standardizing access. They provide not just connectivity, but through a unified Token metering and settlement system, they act as dispatchers and integrators of cross-regional computing resources. For enterprises, this means buying Tokens in the future might not just be buying service from a single cloud vendor or a single computing center, but buying access to a nationally scalable, elastically dispatched computing network.

Thus, the operator Token package can be positioned as: not a single model API, but an aggregation portal for AI services; not just selling Tokens, but selling a combined service of "Token + Connectivity + Security + Computing Dispatch + Application Delivery."

IV. Impact on Enterprises: AI Transitions from Project Investment to Routine Operating Cost

After Tokenization, the logic of enterprise AI usage will change. Previously, many companies viewed AI as a specific project:立项 (project approval), procurement, testing, launch, acceptance. Budgets typically resided in IT or digital departments, with business units primarily acting as requesters.

But as Token packages become common, AI capabilities will resemble everyday business tools. Sales teams generating proposals, marketing teams creating content, customer service handling inquiries, finance preparing reports, legal reviewing contracts, HR generating training materials—all these could consume Tokens.

This leads to three direct impacts:

• First, AI costs become easier to budget for. Companies can allocate Token quotas by department, project, or scenario, shifting AI usage from one-time procurement to monthly budget management.

• Second, AI usage becomes easier to manage granularly. Companies need to know not only how much AI service was bought, but also who is using it, for which scenarios, what effects were achieved, and whether any data or content risks emerged.

• Third, AI investments become easier to include in ROI assessments. With Tokens as a clear consumption unit, companies can further measure whether consuming a certain number of Tokens reduces labor hours, improves customer response speed, increases sales leads, or lowers content production costs. This pushes companies from the stage of "having AI" to the stage of "using AI cost-effectively."

V. Opportunities for Enterprises: Not Just "Integrating Tokens," But "Managing Tokens"

Opportunities around computing power Tokenization may fall into several categories:

First, Token distribution and integration. Enterprises with appropriate qualifications, customer channels, and service capabilities can build on operators, cloud vendors, or model platforms to offer Token packages and AI application combinations for specific industries.

Second, industry application development. The Token itself is just a unit; customers pay for application scenarios. Whoever can transform Tokens into practical tools—tax assistants, compliance assistants, customer service assistants, investment research assistants, sales assistants—gets closer to customer budgets.

Third, AI cost optimization. As usage grows, Token costs become a real issue. Model routing, caching, prompt compression, context management, task layering, and low-cost model substitution will become valuable capabilities.

Fourth, AI governance and compliance services. The more enterprises rely on AI, the more they need internal rules covering data input boundaries, account permissions, sensitive information handling, output auditing, log retention, and vendor management.

Fifth, operator ecosystem partnership. If operators make Tokens the new gateway for AI services, developing applications, agents, industry-specific solutions, and security services around their package system will be a key path for smaller service providers to enter the market. Especially with the national integrated computing power push, service providers helping enterprises with cross-regional computing dispatch and hybrid cloud deployment will find new growth space.

For most enterprises, the opportunity is likely not becoming a foundational Token supplier, but building application, service, and management capabilities around Tokens.

VI. Compliance Considerations for Token-Related Businesses

As the Token package and computing power service market develops, some enterprises are exploring business models related to Tokens, including integrating services or managing quotas for clients based on operator or cloud vendor capabilities. When engaging in such activities, enterprises must first clarify the nature of their business to map compliance requirements accordingly.

The Token discussed here is essentially an AI model call quota, computing power service metering unit, or service rights package. Compliance assessment depends not on the name "Token," but on the actual service provided.

• First, if an enterprise provides underlying cloud resources or computing platforms, it needs to assess whether value-added telecommunications service permits are required, especially related to cloud computing, IDC, or internet resource collaboration services.

• Second, if an enterprise develops SaaS applications for business clients based on existing compliant cloud services and model capabilities, it typically needs to consider ICP filing or ICP licenses, generative AI service filings, algorithm filings, and general compliance obligations for data security, personal information protection, and content safety.

• Third, if an enterprise provides AI services to the public via websites, apps, or mini-programs, at a minimum ICP filing is generally required. If it constitutes commercial internet information services, an ICP license may also be needed.

• Fourth, if an enterprise provides generative AI services to the Chinese public, including generating text, images, audio, or video, it needs to address generative AI service filings, security assessments, model disclosure, and content safety requirements.

• Fifth, if services include algorithmic recommendation, generative synthesis, or deep synthesis functions for the public, algorithmic filing or deep synthesis compliance requirements should also be assessed.

The core compliance criteria are three questions: First, does the enterprise directly provide underlying cloud or computing resources? Second, does it provide generative AI services to the public? Third, does it process user data, enterprise data, or personal information?

VII. Conclusion

From data traffic to Tokens, from isolated cloud resources to a national integrated computing network, the AI infrastructure is transitioning from the "construction phase" to the "operation phase."

In the past, AI was mostly a topic for technical teams and innovation departments. Now, with Tokens being packaged, billed, and serviced, and with the national integrated computing network enabling cross-regional, cross-platform unified dispatch, AI is entering the everyday language of enterprise operations and management.

"AI as a utility" is a vivid phrase, but the real determinant of enterprise value remains the ability to embed such usage-based AI services into business processes and convert them into measurable productivity.

For operators, cloud vendors, model vendors, and industry service providers, Tokens are not the destination, but a new gateway. True competition will return to fundamental questions: Who can provide stable supply? Who can lower costs? Who can dispatch national computing resources? Who understands the scenarios? Who can ensure security? Who can help customers turn AI into operational value?

The next phase of AI is not just about more powerful models, but about more usable, manageable, and sustainable enterprise intelligence services.

-END-