Tokens on the Network: The Telecom Opportunity in Personal AI

For most consumers, the role of a telecom network is fairly straightforward. It carries calls, messages, and data. The technology behind those services has changed dramatically over the years, but the basic relationship has remained familiar.

Artificial intelligence could complicate that picture in an interesting way.

As AI agents become more capable, they will need somewhere to run, a reliable way to identify the people they represent, and a trusted environment in which to communicate with other agents and services. Telecom operators already manage much of that underlying infrastructure. They know who their customers are, operate networks built around reliability, and maintain billing relationships with millions of people and businesses.

That raises a larger question: could operators do more than carry AI traffic? Could they become the trusted service providers behind everyday AI experiences?

Vishal Mathur, Chief Commercial Officer at the Telecom Infra Project, explored that question with Suman Kanuganti, founder and CEO of Personal AI, in a recent TIP podcast.

Kanuganti’s argument begins with the way telecom services have evolved. Voice gave people a way to speak across distance. Messaging made communication asynchronous and programmable. Data created the conditions for the mobile internet, APIs, applications, and digital services.

AI introduces another kind of exchange. Rather than moving only voice, text, or application data, AI systems communicate through tokens that represent language, context, and intent.

Kanuganti believes those tokens could become a new network primitive.

“The data gave us apps, and tokens will give us agents,” he explained during the conversation.

That distinction matters because agents could change how people interact with digital services. Today, booking a hotel, ordering a ride, transferring money, or making a purchase usually requires the customer to open a specific app or website and follow its process. An AI agent could instead understand the customer’s intent and coordinate directly with another company’s service agent.

The customer would not necessarily need to know which model was being used or where the request was processed. They would simply ask for something to be done.

A personal agent tied to a trusted identity

Personal AI is developing private agents with memory that can learn about an individual over time. Within a telecom environment, Kanuganti envisions those agents being associated with a subscriber’s existing identity, including their phone number.

That could make the experience feel less like downloading another AI application and more like activating a new network service.

One of the clearest examples is missed-call handling. Instead of sending a caller to a static voicemail box, an AI agent could answer on the subscriber’s behalf, understand the relationship between the two people, collect relevant information, and respond within permissions established by the subscriber.

Over time, the agent could become more useful because it would retain context. It might understand how its owner prefers to communicate, which requests require immediate attention, and which contacts should receive different kinds of responses.

The same basic architecture could eventually apply to other connected assets, including vehicles, robots, devices, and machines. Each could have an evolving AI identity capable of interacting with people and other systems.

This is where Kanuganti sees a meaningful advantage for telecom operators. Technology platforms can provide advanced AI models, but carriers already manage authenticated identities, regulated customer relationships, privacy commitments, billing, and service-level expectations.

Those capabilities become especially important when an agent is allowed to act on someone’s behalf.

A conversational chatbot can tolerate some inconsistency. A personal agent answering calls, handling private information, or completing transactions needs a much higher level of trust.

Why distributed inference matters

The personal-agent model also creates a major infrastructure challenge.

An agent that learns continuously will accumulate memory, preferences, and relationship context. If every interaction requires sending that expanding context back to a centralized cloud model, the service can become more expensive and less predictable. A response might arrive almost instantly in one instance and take several seconds in another.

That inconsistency may be acceptable for some consumer AI products, but it is harder to reconcile with the reliability customers expect from a network service.

Distributed inference offers another path. Smaller, more targeted AI workloads can run closer to the person or device using them. An AI grid could distribute those workloads across available infrastructure, rather than relying entirely on a single centralized environment. For the customer, the result could be quicker and more consistent responses. For the operator, it could improve the economics of delivering personalized AI at scale.

Kanuganti pointed to work Personal AI presented with Comcast and NVIDIA at NVIDIA GTC. The demonstration examined how small language models could operate across an AI grid while maintaining predictable performance as a user’s memory expanded.

According to Kanuganti, the approach was 40 times more token-efficient than comparable hosted solutions for the use cases tested. It also showed that response latency could remain relatively stable even as the agent accumulated more information.

Those results point toward an important distinction between general cloud AI and network-native AI. A carrier service may need to support millions of small, persistent workloads, each connected to a specific person or device and each expected to respond reliably. That requires different design assumptions than a conventional application sending occasional requests to a large model.

A service customers may never think about in tokens

Even if tokens become central to the underlying service, customers may never see them.

Most people do not track every megabyte of data their phone consumes, particularly when they are on an unlimited plan. AI could follow a similar pattern. An operator might include a certain level of AI service within a mobile or household subscription, or offer enhanced agents as part of a higher-value package.

The customer would experience the outcome rather than the unit of consumption.

That could give carriers several ways to create value. They might offer personal agents directly, provide distributed inference infrastructure to enterprises, support AI services developed by partners, or bundle a combination of models and capabilities behind one network relationship.

Kanuganti also sees an opportunity for operators to support agent-to-agent communication. A customer’s personal agent could communicate with the agent of a hotel, bank, retailer, or transportation provider. Identity and authorization could be handled through the network, while the customer avoids jumping between separate applications and accounts.

The idea is still developing, and significant work remains around privacy, interoperability, billing, and control. Customers will need clear ways to decide what their agents can access, remember, disclose, and do. Operators, technology providers, and standards organizations will also need common approaches that prevent the market from breaking into another collection of closed ecosystems.

That is where industry collaboration becomes particularly important. TIP can help operators, infrastructure providers, silicon companies, AI developers, and other ecosystem partners test these models together. Reference architectures and real-world demonstrations will be necessary to determine which ideas can scale beyond a controlled proof of concept.

The commercial opportunity is significant, but Kanuganti argues that the implications go further. Telecom operators could influence whether personal AI becomes broadly accessible, whether it remains tied to a handful of major platforms, and whether consumers have meaningful privacy and control.

Carriers have spent years looking for ways to create more value above connectivity. Personal AI and distributed inference may offer one of the more credible opportunities to do so, provided operators begin treating AI as a service they can help shape rather than traffic they simply transport.

Talk, text, and data built the modern telecom industry. Tokens may give operators a new role within it.

Watch the full conversation on YouTube: https://youtu.be/ZI0JE0_FlSA

Listen to the podcast on Spotify: https://open.spotify.com/show/7mO0914KcVqzXI9GpTz8W2?si=IsKvxN1KSsuxbJYhhM6StA