After telecom operators start selling AI as data packages, large model companies collectively enter a new battleground.

Source | Silicon-based Quadrant

When users no longer fret over whether to upgrade their monthly data plans, they might start worrying about how many Tokens to buy each month.

Tokens are about to be packaged and sold as standardized services by telecom operators, just like data, broadband, and SMS.

Recently, China's three major telecom operators have successively launched Token package products: monthly subscription-based Token plans for individual users, tiered computing power packages for developers and enterprise clients, and the announcement that dozens to hundreds of large models have been integrated into their platforms, allowing for "monthly purchases, multi-model access, and bill payments."

China Telecom has introduced personal and enterprise Token packages, with monthly fees starting as low as 9.9 yuan for 10 million Tokens. Local operators such as Shanghai Mobile and Shanghai Telecom have launched quota-based or universal Token billing models, with Shanghai Mobile offering 400,000 Tokens for 1 yuan.

As operators begin selling Token services, the cost for users to switch large models will significantly decrease. For large model companies, "user stickiness" will weaken, and only by "competing more fiercely" can they retain their market share.

In the future, large model companies like Doubao, Qianwen, and DeepSeek will not only compete on price and "Token quality per unit of energy consumption" but also on "higher-value AI application solution capabilities."

01 What is Token Service?

To understand Token services, one must first understand what Tokens are.

Computers cannot directly recognize text; they only recognize binary code (0s and 1s). Therefore, every word, character, voice, and punctuation we input is converted into binary code through specific encoding mechanisms.

In the context of large models, digital encoding is also the first step, and the number of digits in the encoded representation varies slightly for each character.

Tokens are the smallest computational units for large models to process information. User input, contextual memory, and model output are all calculated in Tokens. The more complex the model call, the longer the context, and the deeper the Agent execution chain, the higher the Token consumption.

Typically, in English, one Token roughly corresponds to four letters. In Chinese, due to the higher information density of characters, one Chinese character, one punctuation mark, or one phrase often corresponds to 1-2 Tokens.

Since large models think and output Token by Token, the industry sells and settles the invocation costs and usage quotas of large models in the form of "Per Million Tokens" or "quota points."

Currently, large model companies charge for Tokens on a tiered basis. Ordinary users can use models like Doubao and Qianwen in standard mode for free, while enterprise-level heavy users can purchase different tiers of API monthly subscriptions or metered services.

Since last year, operators have opened "computing power supermarkets" for large models. Model vendors are "tenant merchants," and operators collect "platform fees + computing power fees + channel fees." Users are not buying "operator models" but rather accessing any large model on the telecom platform using the operator's computing power, billed by Tokens.

In July 2025, China Mobile launched the model service platform MoMA (Mobile Model Access); in April, China Telecom released the Xingchen TokenHub operation service platform, and in May, China Unicom launched the "Unicom Star Network" Token service platform. These platforms integrate multiple mainstream large models from Baidu, Alibaba, ByteDance, DeepSeek, etc., with unified APIs, authentication, and billing.

Operators' platforms adapt to multiple large models internally, allowing users to switch smoothly by simply changing the model name (Model ID).

02 Why Are Operators Selling Tokens?

The explosion of Token services is no accident.

First, the billing model has changed. In the traditional cloud computing era, users were accustomed to paying for "server rental time" or "fixed bandwidth" (i.e., IaaS-layer computing power), buying bandwidth speed and time. However, with the development of large models, the capabilities provided by different models and the cost disparities for different tasks vary greatly. For example, stronger models have higher Token costs; longer contexts consume more Tokens; higher inference complexity leads to higher actual costs. Billing by Tokens aligns "the degree of intelligence consumed by users" with "the computing power costs incurred by vendors."

Second, it lowers technical barriers and "trial-and-error costs." The research and deployment of large models often require tens or even hundreds of millions of dollars in investment. For most small and medium-sized enterprises (SMEs) and individual developers, building their own models is unrealistic. Token services package and fragment "Artificial General Intelligence (AGI)" capabilities, allowing developers to call APIs on demand and pay Token fees without worrying about the underlying tens of thousands of GPUs consuming electricity.

Finally, there is an urgent demand driven by the explosion of application-layer scenarios. Entering 2026, AI Agents, AI-assisted programming, and multimodal content generation have exploded in application-layer scenarios. These applications require frequent "throughput" interactions with underlying large models during daily operation. An automated AI code-writing tool might consume millions of Tokens overnight. This high-frequency, massive interaction necessitates the market to provide more standardized, stable, and price-competitive Token package services.

Over the past two decades, operators' business models have undergone three core changes in measurement units.

The first stage was the voice era, where operators sold minutes. The second stage was the mobile internet era, where they sold data in GB. Now, in the AI era, operators are beginning to sell Tokens.

Tokens are undergoing an evolutionary process similar to that of data. Initially, they were just technical indicators; later, they became billing units; and eventually, they evolved into standardized commodities.

Operators' entry signifies that Tokens have begun to transcend the technical realm and enter the consumer system.

In the coming years, the way users purchase AI capabilities may fundamentally change: individual users buy "AI monthly packages," enterprises procure "Token resource pools," family broadband includes AI quotas, and government and enterprise dedicated lines integrate Agent services. Tokens will become as fundamental a resource as electricity, water, and data.

However, this does not mean that operators will replace large model companies.

03 How to Buy Tokens Appropriately?

Should Token services be purchased directly from native large model companies or through operator platforms? Currently, both business models have their pros and cons.

The first model is the native large model vendor approach, which charges per million Tokens. Vendors like OpenAI, Anthropic, DeepSeek, and Qianwen generally adopt this system, where users pay separately for input and output Tokens. Some, like Qianwen, may use a pre-purchase at the beginning of the month and settle at the end of the month.

The second model is the operator's monthly subscription for Token quotas. For example, Shanghai Telecom offers a minimum of 9.9 yuan for 10 million Tokens, with additional charges for excess usage, and plans to integrate Token benefits into the family's "Beautiful Home" digital space, supporting one-click payment via phone bills.

This "all-inclusive" or "phone bill integration" model allows Chinese users to purchase large model computing power just like buying data packages.

The overseas market primarily uses API tiered pricing by native large model companies, while the domestic market has pushed Token services into a "package" era similar to mobile phone bills.

Currently, both billing models have their advantages, as the Token package user base mainly falls into three categories.

The first category consists of independent developers and tech enthusiasts (Geeks). They utilize API interfaces provided by various vendors to build their own personalized AI applications, such as productivity tools, automatic translation plugins, and personal knowledge bases.

The second category is SMEs, startups, and B-end independent software vendors (ISVs), which are the core customer base for Token services. Whether purchasing Tokens for employees' programming needs, developing AI Agents for specific industries, or embedding AI assistance into existing enterprise ERP and CRM systems, SMEs need to subscribe to "team Token packages" from cloud vendors or operators.

The third category is "heavy AI-dependent" professionals and ordinary households who frequently use AI at home for copywriting, coding, or assisting children with homework.

For SMEs and startups, from a technical economics perspective, the pure Token-based billing model of native large models is more scientific.

The operator's package model, however, has two advantages. On the one hand, independent developers are not tied to a single large model and can independent choice (autonomously select) multiple large models through the platform. On the other hand, Token services may reach mass consumption faster. Most people understand what 100GB of data means but cannot perceive what 10 million Tokens represent.

Operators' monthly subscriptions essentially lower the cognitive barrier. Users do not need to understand Tokens; they can start by understanding their needs with a basic 9.9 yuan/10 million Tokens plan.

As operators begin selling Token services, "Doubao and peers" are about to start competing fiercely on three levels.

From "competing on parameters" to "competing on energy efficiency": Large model companies will no longer be able to blindly pursue larger parameters and higher energy consumption. Instead, they will focus on capabilities like model distillation, quantization, and inference optimization, which can output higher-quality Tokens with less energy.

Price competition will intensify further. After operators aggregate hundreds of models, user switching costs decrease. If Model A raises prices, users can replace it with Model B through the platform. When the differences in model capabilities are insufficient, price becomes a core competitive factor.

The profit center for large model companies will shift. Simply selling APIs offers limited profits. Future profitability may focus on Agents, industry applications, and enterprise solutions. Models themselves will gradually become infrastructure, while the application layer becomes the value center.

Perhaps a "two-sided market" is forming: operators control the entry points, while model vendors control the capabilities.