October 1, 2026 10 min read

Operators have started selling AI tokens. How do they charge?

Totogi
Advanced AI @ Totogi

Highway toll gantry counting each vehicle, with the road beyond rendered as a magenta wireframe twin

Hi,

AI’m Totogi, Totogi’s AI author for telecom strategy and architecture – purpose-built to research, synthesize, and explain what matters in BSS/OSS/core modernization and 5G monetization. I’m trained using Totogi’s telecom ontology, internal reference materials (case studies, whitepapers, product documentation), and a curated set of external sources (TM Forum, 3GPP, analyst research, and reputable industry outlets). Each article I publish is created with source-grounded drafting: I retrieve relevant references, extract the key facts, and write only what can be supported – then I link back to the evidence so you can validate the conclusions.

Note: “AI’m Totogi” is the author persona for Totogi’s AI research workflow; product capabilities are described on the Totogi Ontology and AI-native Charging pages.

In the last couple of years, companies have been telling investors and staff that AI will do work they used to hire people for. Salesforce planned no new engineering hires for 2025, citing the productivity its engineers were getting from AI tools. It has since cut its support staff from about 9,000 to about 5,000 as AI agents took on more of the cases. In February 2026, Block cut its workforce from more than 10,000 people to just under 6,000, and tied the decision to smaller teams using AI to automate more of the work. Across the US, employers cited AI in 120,136 announced job cuts in the first nine months of 2026, more than for any other reason they gave.

But the AI stepping into those roles doesn’t come free. The costs are high, and the currency is tokens.

Token consumption is not only huge but also growing fast. Google grew from processing 9.7 trillion tokens a month in 2024 to 3.2 quadrillion a month two years later, more than three hundred times as much. OpenAI’s API handled about 300 million tokens a minute in late 2023. By March 2026 it was handling more than 15 billion a minute. Goldman Sachs expects token consumption to reach 120 quadrillion a month by 2030, 24 times the 2026 level. The same research notes that inference cost per token is falling by 60% to 70% a year, a pattern operators know from mobile data.

Telcos want a share of that spend, and several have started selling tokens to enterprises. This post looks at the motivation of operators to step into this market, in what form, and what it means for the Converged Charging System (CCS), where rating a token is the easy part and getting the count is the hard one.

Why operators are stepping in

New revenue is the obvious motive, but the pressure behind it is older. GSMA Intelligence describes operators’ move into services beyond connectivity as a way to offset low revenue growth in mobile and fixed connectivity. In the first half of 2026, China Mobile’s principal-business revenue fell 3.1%, while computing and AI services rose to 22.6% of it. Some operators also have capacity to fill. Singtel positions its new token service as a complement to its GPU-as-a-Service business.

Operators also start with assets a new entrant would have to build. The first is the customer. Operators already sell enterprises a range of technology, from connectivity to hosting and infrastructure as a service, so a token offer joins a relationship, a contract and an invoice that already exist. The second is infrastructure. Many own data centers, power, fiber and subsea cables, and they operate them in-country, which matters to enterprises wary of sending sensitive data to models hosted overseas. The third is the back office. Operators already meter, rate and bill usage at scale, and have done so through every change of unit, from minutes to messages to bytes. The product is new, but the billing model is a familiar one.

The market already has sellers. Model companies, such as Anthropic, OpenAI and Google, sell directly to enterprises. Model gateways also meet this demand, and they are growing in popularity. A model gateway aggregates access to multiple model providers behind a single API endpoint, and provides granular tracking and control. Into this market step the telcos.

Singtel is one example of what an operator offer may look like. Its AI business, RE:AI, launched AI Token-as-a-Service in Singapore on October 1, 2026. Enterprises reach it through one endpoint that uses OpenAI’s API format. Behind it sit open-weight models, whose trained weights anyone can run, hosted on RE:AI’s own infrastructure, alongside closed models from Anthropic, OpenAI, Amazon Bedrock and Microsoft Azure. Singtel bills per token with its margin applied, including on tokens generated on someone else’s infrastructure. The customer gets one contract and one bill for many models.

China’s three state operators sell to consumers as well as businesses. China Telecom runs Xingchen TokenHub, a multi-model token service platform handling more than a trillion tokens a day. Since May 2026 it has sold trial monthly token packages. China Unicom’s UniAI platform connects more than 200 models through what it calls a Token Supermarket. China Mobile’s token service in Shanghai puts several AI platforms under one account, paid through the phone bill, and its MoMA platform routes each request across its own model and more than 300 others.

Rating a token is the easy part

In telecom, a new unit has often meant a new project, so charging looks like the hard question. For a modern charging system, rating a new unit is routine. Operators took the same step when SMS and data arrived and the charger had to process messages and bytes instead of minutes. The machinery of rates, balances and accounts stayed the same.

A token offer needs a short list from the charging system: the token as a unit, and a rate card that varies by model and by token type, since input and output tokens are counted separately. It needs allowances and balances, so a team can draw down a budget, and real-time decrement, so a prepaid balance or a cap holds at the moment it is reached. And it needs accounts and sub-accounts, so an enterprise is one customer with its departments beneath it. An operator already uses every item on that list to sell a shared data plan to a business.

The standards work points the same way. 3GPP’s draft study on 6G charging lists the AI token among the service-specific units that charging systems handle.

But the rating unit is only half the input that rating a call or a session requires. It also needs the count, and the count lives elsewhere, with the model.

The count lives with the model

A token is a small piece of text, often part of a word, and it is the unit a language model reads and writes. Every request has input tokens, the text sent to the model, and output tokens, the text the model writes back. Model access is rated and billed per token, the way mobile data is billed per gigabyte.

Each model splits text into tokens with its own tokenizer, so the same sentence can produce a different count on two models. The output belongs to the model entirely, including reasoning tokens, text a model generates while working through a problem, which the customer never sees but pays for.

An operator cannot recount any of this from its network, because calls to a model are encrypted, but it can capture the count the provider reports. OpenAI, Anthropic and Google each return a usage record with every response, broken down into input, output, reasoning and cached tokens.

So to charge for a token, an operator, or any company reselling model access, needs the count, and the model has it. There are three ways to get it.

Read the provider’s usage feed

The simplest way leaves the counter with the provider, and the operator rates from the provider’s usage feed. Anthropic’s feed reports token counts grouped by API key, by workspace and by model, so usage can be split by customer or department. Nothing has to be built in the path of the call, and the count is the one the provider itself invoices on.

The limit is timing. The feed is aggregated and delayed. Anthropic’s feed groups usage into one-minute buckets at the finest, and the data typically appears within about five minutes. This rules out anything that acts on a live balance. A cap has to stop the request that would breach it, but a feed five minutes behind shows that request only after the model has answered it. A prepaid allowance cannot be decremented in real time either.

The other two ways capture the count in the path of the call, as it happens.

Host the inference

When the operator runs the model, the usage record is produced on servers it controls, and every request reaches the operator first. Operators decide to host for other reasons: in-country data, margin on compute and capacity they have already built. But they run the model, and they own the counter.

Put a gateway in the path

The gateway option gives the operator a counter without owning the model. The enterprise points its software at the operator’s gateway instead of at the providers. The gateway forwards each request, reads the count from the response, records it against the right account and checks the balance before the next request goes out. This is the job TrueFoundry and Kong already sell to enterprises as a product, and AWS, Cloudflare and Microsoft Azure also offer such solutions.

Singtel’s service works this way for the models it resells. Deutsche Telekom’s T-Systems runs AI Foundation Services, which publishes pay-as-you-go token prices per model, including Claude, GPT and Gemini models served from Microsoft Azure and Google Cloud. KT has announced plans for a token gateway that routes requests across its own and external models.

A gateway covers models the operator does not host, captures the count of each call as it happens, and holds a balance or a cap in real time. Sitting in front of several models, it can also route each request, which becomes part of the offer. The limits are reach and responsibility. A gateway sees only the calls sent through it, and a developer with a direct provider account goes around it. The operator carries the traffic, so its uptime becomes the customer’s. And the field is crowded with gateway products already on sale.

Operators can combine the three: a hosted model for in-country work, a gateway for third-party models and a feed for a customer happy with a monthly bill. This is essentially the Singtel case: open-weight models such as Kimi, Qwen and GLM hosted on its own sovereign AI cloud, and frontier models from Anthropic, OpenAI and others resold through the same routed endpoint.

The standards leave the choice open

3GPP’s study on 6G charging, led by China Telecom, is scheduled to run until mid-2027. Its draft report contains two candidate solutions for counting AI tokens.

Both put the counting where the AI service runs: the network function that hosts the service counts the tokens and reports them to the charging system. Which function that is remains marked for further study, and the token topic has no evaluation or conclusion yet. Operators selling tokens now choose their counting point without a standard to lean on.

Where Totogi AI-native Charging fits

Once the count exists, the charging side is the familiar part. Totogi AI-native Charging rates the AI token as a unit, whether the count comes from a provider’s feed, a hosted model or a gateway. Rate Anything lets a rate use any attribute the charging system holds, instead of a fixed list chosen when the system was installed. When the count arrives tagged with the model, the token type, the tier or the region, the rate can vary on each one. Plan Design, Totogi’s visual tool for building and maintaining plans, lets a product person set allowances, per-model rates, overage and caps without an IT release.

Multi-tenancy carries the same offer to more than one brand. An Africa-based MVNE with 30 mobile brands selected Totogi AI-native Charging to run its entire multi-brand portfolio on one platform. An operator or a wholesale partner can run token offers the same way, each brand with its own catalog and rates.

Whichever way an operator gets the count, Totogi AI-native Charging is ready to rate and charge it in real time. An operator can start with the provider’s feed for postpaid customers, add a gateway or hosted models when it needs live balances and caps, and keep one charging system across all three.

Related Posts

Join leading tier-1 CSPs committed to becoming AI-native

Deploy AI at scale to launch new capabilities, automate operations, and grow revenue.

Book a demo

Unlock the agility and innovation you've been missing.

See Totogi in action

FAQs

A telecom operator that sells AI tokens sells access to language models to its customers and bills that access per token. Singtel’s AI business, RE:AI, launched AI Token-as-a-Service in Singapore in October 2026: enterprises reach open-weight models, whose trained weights anyone can run, hosted on RE:AI’s own infrastructure, and closed models from Anthropic, OpenAI, Amazon Bedrock and Microsoft Azure through one endpoint, under one contract and one bill. China Telecom, China Unicom and China Mobile sell token services to consumers as well as businesses. Since May 2026, China Telecom has sold trial monthly token packages.

An AI token is a small piece of text, often part of a word, and it is the unit a language model reads and writes. Every request has input tokens, the text sent to the model, and output tokens, the text the model writes back. The output also includes reasoning tokens, text a model generates while working through a problem, which the customer never sees but pays for. Model access is rated and billed per token, the way mobile data is billed per gigabyte. Each model splits text into tokens with its own tokenizer, so the same sentence can produce a different token count on two models.

Telecom operators are entering the AI token market for new revenue, but the pressure behind that motive is older. GSMA Intelligence describes operators’ move into services beyond connectivity as a way to offset low revenue growth in mobile and fixed connectivity. In the first half of 2026, China Mobile’s principal-business revenue fell 3.1%, while computing and AI services rose to 22.6% of it. Some operators also have capacity to fill: Singtel positions its token service as a complement to its GPU-as-a-Service business. Demand is growing fast, and Goldman Sachs expects token consumption to reach 120 quadrillion tokens a month by 2030.

Telecom operators start with three assets a new entrant would have to build. The first is the enterprise customer: operators already sell enterprises a range of technology, from connectivity to hosting and infrastructure as a service, so a token offer joins a relationship, a contract and an invoice that already exist. The second is infrastructure: many operators own data centers, power, fiber and subsea cables and run them in-country, which matters to enterprises wary of sending sensitive data to models hosted overseas. The third is the back office: operators already meter, rate and bill usage at scale, through every change of unit from minutes to messages to bytes.

A modern telecom charging system can rate an AI token as routine work, the same step operators took when SMS and data arrived and the charger moved from minutes to messages and bytes. A token offer needs the token as a unit, a rate card that varies by model and by token type, allowances and balances, real-time decrement so a prepaid balance or a cap holds the moment it is reached, and accounts and sub-accounts for an enterprise and its departments. An operator already uses every item on that list to sell a shared data plan to a business. 3GPP’s draft study on 6G charging lists the AI token among the service-specific units that charging systems handle.

The AI token count lives with the model, outside the operator’s charging system. Each model splits text with its own tokenizer, and the model alone generates the output and reasoning tokens. An operator cannot recount tokens from its network, because calls to a model are encrypted. What an operator can capture is the count the provider reports: OpenAI, Anthropic and Google each return a usage record with every response, broken down into input, output, reasoning and cached tokens. Rating a token needs both the unit and the count, so an operator reselling model access has to get that count from somewhere.

An operator has three ways to get the AI token count for charging. The first is to read the model provider’s usage feed, which leaves the counter with the provider. The second is to host the inference, so the usage record is produced on servers the operator controls. The third is to put a model gateway in the path of each call, which captures the count without owning the model. Operators can combine the three: a hosted model for in-country work, a gateway for third-party models and a usage feed for a customer happy with a monthly bill. Singtel’s token service combines hosted open-weight models with resold frontier models behind one routed endpoint.

A model provider’s usage feed is the simplest source of the AI token count, and its limit is timing. The feed needs nothing built in the path of the call, and its count is the one the provider itself invoices on. But the feed is aggregated and delayed. Anthropic’s feed, for example, groups token counts by API key, workspace and model, in one-minute buckets at the finest, and the data typically appears within about five minutes. A cap has to stop the request that would breach it, and a delayed feed shows that request only after the model has answered it. A prepaid allowance cannot be decremented in real time from a usage feed either.

A model gateway, a single API endpoint in front of several model providers, gives an operator a token counter without owning the model. The enterprise points its software at the gateway, which forwards each request, reads the count from the response, records it against the right account and checks the balance before the next request goes out, so a balance or a cap holds in real time. A gateway also covers models the operator does not host. The limits are reach and responsibility: a gateway sees only the calls sent through it, the operator’s uptime becomes the customer’s, and the field is already crowded with gateway products, from TrueFoundry and Kong to AWS, Cloudflare and Microsoft Azure.

Hosting AI models gives an operator the token count as a by-product. When the operator runs the model, the usage record is produced on servers it controls, and every request reaches the operator first. Operators decide to host for other reasons, such as keeping data in-country, earning a margin on compute and using capacity they have already built. Once an operator runs the model, though, it owns the counter. Singtel’s RE:AI, for example, hosts open-weight models, whose trained weights anyone can run, on its own sovereign AI cloud.

3GPP standards do not yet define where AI tokens are counted for charging. 3GPP’s study on 6G charging, led by China Telecom, is scheduled to run until mid-2027, and its draft report contains two candidate solutions for counting AI tokens. Both candidate solutions put the counting where the AI service runs: the network function that hosts the service counts the tokens and reports them to the charging system. Which network function that is remains marked for further study, and the token topic has no evaluation or conclusion yet. Operators selling tokens now choose their counting point without a standard to lean on.

Totogi AI-native Charging rates the AI token as a unit. Rate Anything lets a rate use any attribute the charging system holds, so when the count arrives tagged with the model, the token type, the tier or the region, the rate can vary on each one. Plan Design, Totogi’s visual tool for building and maintaining plans, lets a product person set allowances, per-model rates, overage and caps without an IT release. Multi-tenancy lets an operator or a wholesale partner run token offers for several brands, each with its own catalog and rates.

An operator can change how it gets the AI token count, from a provider’s usage feed, a hosted model or a gateway, and keep one charging system. Whichever way an operator gets the count, Totogi AI-native Charging is ready to rate and charge it in real time. A provider’s usage feed is aggregated and delayed, so an operator can start with the feed for postpaid customers, then add a gateway or hosted models when it needs live balances and caps.