Token gestures: are AI tokens the telecom industry's next money maker?

As AI adoption accelerates, some believe that AI tokens could become the next billable unit, with operators acting as distributors of AI services in much the same way they once sold connectivity. Adam Hughes asks the question: should operators be selling tokens or commercialising the outcomes they enable?

Could AI tokens become the next gigabyte?

In its quest to bill the world, the telecom industry is once again searching for its next resource to monetise – first it was minutes, followed by messages, and then they came for the data.

Now, Tokens-as-a-Service is in its sights.

A growing number of industry voices argue that AI tokens could become the next billable unit, much as gigabytes became the commercial foundation for mobile internet. Suman Kanuganti, CEO and founder of Personal AI, thinks that telcos’ future lies in selling bundles of AI tokens directly to customers.

The idea is gaining traction. China's three major telecom carriers – China Mobile, China Telecom and China Unicom – have already started selling AI token plans across multiple model providers directly to enterprises and consumers. This comes as AI usage across China has risen to over 140 trillion tokens every day.

Tokens are the basic units that large language models use to process language.

AI models don’t process language in the same way we do. Put simply, text is divided into smaller units called tokens, which are converted into numerical representations that the model can process.

Because more tokens generally require more computation, memory and GPU processing, they have become a practical way for AI providers to estimate inference costs. A token can be a word, part of a word, punctuation, an individual character or a chunk of code. For example, a sentence such as:

The customer cancelled the subscription.

can be broken up into something like:

[The] [customer] [cancel] [led] [the] [subscription] [.]

… producing seven tokens.

Tokens are consumed as either input tokens or output tokens. Say you ask an AI chatbot to “Explain quantum physics in 500 words”. Reading this input might consume only 10 tokens, but generating an output could consume hundreds of tokens, because they are generated sequentially, with each token helping the model predict the next.

Tokens roughly correlate with compute consumption, memory requirements and inference costs.

In other words, tokens provide a practical way to relate language processing to computational demand and cost. Could they also provide a unit of billing?

Traditional telecom revenues have matured. Voice has become effectively unlimited; SMS was largely displaced by OTT messaging platforms like WhatsApp and iMessage. Data remains commercially important, but competitive pressure and ever-larger allowances have steadily eroded its ability to command premium pricing.

Now, imagine a mobile subscription that includes an AI assistant capable of managing emails, booking travel, conducting research and handling customer enquiries. Instead of a handful of chatbot interactions each day, autonomous AI agents may eventually perform thousands of actions on a customer's behalf, creating a continuous stream of measurable consumption.

The opportunity here is clear: AI workloads consume tokens, and token usage can be measured and charged for.

Suppose an operator in the future offers access to an AI assistant for £15 per month. The operator purchases wholesale AI tokens from multiple providers. Usage records arrive via APIs, and the BSS/OSS stack handles charging and billing behind the scenes. The customer never sees the underlying tokens.

If autonomous agents consume thousands or millions of tokens each day, this usage could become sufficiently regular and measurable for operators to package into subscription services. Whether customers would accept such a model is less certain, but it’s technically feasible.

The argument goes that, over time, AI will come to resemble a utility, and utilities are priced in units that are easy to measure – water in litres, electricity in kWh, and mobile data in bytes.

Just as streaming a movie uses more data than sending a text, writing a single email uses fewer tokens than analysing a 500-page technical document, and charging the same for both these requests makes no sense.

Most customers don't know – and don’t want to know – how many tokens their last request consumed, which LLM they used or what a model context window is. They just expect their assistant to deliver the desired outcome.

One of the clearest arguments for token monetisation comes from Mavenir and Red Hat, which together recently launched an Integrated AI Platform designed specifically for telecom operators, with usage-based charging and billing capabilities. AI tokens simply become another category of usage event. The platform therefore appears to strengthen the case for token-based metering, rather than necessarily for selling token bundles directly to customers.

A related model is emerging around sovereign AI infrastructure, with Nvidia proposing that operators become “sovereign AI factories” providing AI services to governments, enterprises and developers. In this model, tokens are a wholesale accounting mechanism rather than a retail product.

Many enterprises don’t want to purchase and manage GPUs, just as they don’t want to purchase storage arrays or database servers. They want AI applications that solve business problems. Tokens simply provide the accounting framework required to measure consumption and allocate costs behind the scenes.

The opportunity for telcos lies in applying their existing strengths – in charging, billing, settlement and service management – to commercialise AI services delivered through sovereign or operator-owned AI platforms.

However, the case for selling tokens directly to customers is less straightforward.

AI tokens are valuable because the compute to generate them is scarce. The upfront cost of training frontier models has run into billions of dollars, and inference requires expensive GPUs, which are increasingly in short supply. But over time, as models become more efficient and open-source alternatives improve, the cost of inference is likely to continue falling.

Usage-based metrics in telecoms tend to commoditise as scarcity disappears. Minutes used to be expensive because capacity was tight, but eventually, network efficiency improved, and voice became effectively unlimited. SMS followed a similar path, with messaging platforms such as WhatsApp chipping away at the economic value of SMS. Operators believed data would replace declining voice and messaging revenues, but consumers now increasingly expect large or even unlimited allowances.

The more a resource becomes abundant, the harder it becomes to monetise. Selling token bundles could eventually sound as strange as an operator advertising a daily allowance of 100 WhatsApp messages.

If a telco offered two plans – one bundled with 50 million tokens, the other with 100 million tokens, most customers would ask: what can I actually do with that? Similar concerns were once raised about mobile data: few consumers could translate kilobytes or megabytes into a precise number of emails, web pages or streamed videos. Operators nevertheless made data allowances understandable through bundles, usage alerts and practical guidance.

Tokens present an additional challenge, however, because their cost and usefulness can vary considerably by model, provider and task. Furthermore, continued improvements in AI efficiency could quickly undermine the value of fixed token allowances, with newer models able to deliver better results at a much lower cost.

Customers may eventually expect unlimited AI in the same way they expect unlimited calls. Arguably, some developers already do; Uber managed to use up its annual token budget in just four months earlier this year.

Free tiers complicate the matter further. The overwhelming majority of consumers today experience AI through free tiers; only about 3% of AI users pay, according to 2025: The State of Consumer AI. The other 97% probably aren’t thinking about tokens or context windows.

The industry has seen this before. For years, businesses talked about the "API economy". The assumption was that APIs would become products in their own right, but in reality customers were primarily interested in the services they enabled. The API was just the method of delivery – tokens may now follow a similar path.

Rather than allowing hyperscalers to capture all AI value, operators could build AI platforms, host AI workloads and monetise AI consumption themselves. This could involve investing in infrastructure to run open-source models such as DeepSeek. However, operators would be entering a competitive market in which differentiation may be limited and margins under pressure.

Operators attempting to build a retail token business could face exactly the same problem they experienced with minutes and data: the underlying unit becomes commoditised.

China's emerging token economy is often presented as evidence that token-based telecom services are inevitable, but the situation there is significantly different.

DeepSeek and other Chinese models have dramatically reduced inference costs and expanded access to AI compute. At the same time, Chinese telcos are being positioned as AI infrastructure providers rather than simply connectivity providers. This forms part of a broader national AI infrastructure strategy supported by conditions that simply don’t exist in Western markets, including state-backed operators and a rapidly expanding domestic model ecosystem.

There’s also a strong cultural dimension; while the narrative around AI in the West veers between “gentle singularity” and “white-collar bloodbath”, China's messaging has focused on AI as a productivity tool, an industrial capability and an issue of national competitiveness.

The strongest opportunity for telcos probably lies in commercialising AI services.

If operators begin offering or aggregating those services, they may receive usage records describing model activity, token consumption, latency and service levels. These records would need to be mediated and validated, with the resulting usage rated, charged, billed and settled.

AI services also introduce a different and potentially more dynamic set of charging variables than traditional telecom services. Costs can vary according to model type, reasoning requirements, latency and autonomous agent activity. This creates new challenges for charging systems, product catalogues and partner management platforms.

However, consumers are unlikely to buy millions of tokens in the way they buy hundreds of GBs of data. Instead, customers might subscribe to an “AI Assistant” or an “AI Research Agent” tuned to a desired outcome, while the BSS tracks token consumption internally. Customer-facing charges could be based on tasks completed or enquiries resolved, with the BSS platform managing the underlying economics.

Partner management may be the most significant opportunity. Operators could handle services from OpenAI, Anthropic, Google or future regional providers, creating a wholesale ecosystem similar to roaming, content partnerships or cloud marketplaces. Usage would need to be tracked, costs allocated and settlements performed across multiple suppliers.

These are areas where telecom operators already possess mature operational capabilities, and where BSS vendors already specialise.

If operators believe they have discovered a new equivalent of mobile data, they may be disappointed; the history of telco monetisation suggests that resource-based pricing eventually collapses into utility economics. Instead, tokens are more likely to disappear into the background, remaining critical to operations while becoming less important as a customer-facing product.

The telecom industry's history suggests that every successful metering unit eventually becomes invisible. Customers no longer think about minutes or individual messages. Increasingly, many do not think about gigabytes – AI tokens may ultimately follow the same trajectory.