Telecom Operators Turn to Open AI Models for More Control

Telecom operators are changing how they build AI systems. Instead of putting every workload on proprietary models, a growing number of operators are turning to open AI models that they can customise, deploy and govern around their own networks.

The shift is showing up in some of the industry’s biggest AI projects. SoftBank is developing a large telecom model using NVIDIA’s Nemotron framework, while Indonesia’s Indosat Ooredoo Hutchison is using open AI technology to build models adapted to local languages and cultural requirements.

The attraction is fairly practical: more control over data, more freedom to customise models and potentially lower costs at the scale telecom companies operate. NVIDIA’s latest telecommunications research suggests the movement is becoming difficult for operators to ignore.

Open AI Is Becoming Part of Telecom Strategy

Telecom companies handle enormous amounts of network, customer and operational data. That makes AI particularly useful, but it also creates a problem. Operators need models that understand highly specialised telecom environments without unnecessarily exposing sensitive information to external systems.

According to NVIDIA’s State of AI in Telecommunications report, released on October 6, 2026, 89% of surveyed telecom companies said open-source models and software are important to their AI strategies. The figure points to a broader change in how operators think about AI infrastructure.

Rather than treating an AI model as a finished product, operators increasingly want something they can tune around their own networks, data and business requirements.

Telecom Operators Want Models They Can Customise

Open models give telecom companies more room to adapt AI to specialised workloads. A general-purpose model may be capable of handling ordinary questions, but network operations require a very different understanding of terminology, configurations, incidents and infrastructure.

That is where fine-tuning and domain-specific training become important. Operators can adapt open models using proprietary network and customer data while maintaining greater control over where that information is processed.

The approach also gives operators more choice over deployment. Depending on the workload, a model can be adapted for private infrastructure or other controlled environments rather than forcing every AI task through a single external provider.

SoftBank Is Building a Large Telecom Model With NVIDIA

SoftBank Corp. is one of the companies pushing this model-specific approach. Its Large Telecom Model uses NVIDIA’s Nemotron framework and is designed around the requirements of telecom network operations and management.

The goal is to make AI useful inside the machinery of the network itself. That includes areas where operators need to understand network conditions, manage infrastructure and respond to operational requirements.

For SoftBank, an open-model foundation provides the flexibility to develop technology around its own telecommunications environment instead of relying entirely on a general-purpose model built for a much wider range of industries.

Indosat Uses Open AI to Bring Local Language Into Telecom

Indonesia offers another interesting example. Indosat Ooredoo Hutchison’s Sahabat-AI initiative uses open AI frameworks to develop models suited to Indonesian languages and cultural contexts.

That localisation matters because AI systems do not operate in a vacuum. Language, terminology and cultural context can affect how useful an AI system becomes when it moves from a demonstration into everyday customer and business interactions.

Sahabat-AI shows how open models can also become part of a country’s broader AI strategy. Instead of simply importing a model developed elsewhere, local organisations can adapt AI technology to their own linguistic and operational needs.

AT&T Is Using Open Models for Telecom-Specific AI

AT&T is taking a similar approach, with its AI strategy placing emphasis on flexibility and governance. The company’s OTel 2.0 model was built using around 400 billion telecom-specific tokens, creating a model foundation tailored to telecommunications workloads.

The company has argued that open models are important because they give operators greater control over how AI workflows are developed and governed.

That becomes increasingly relevant as operators move AI into sensitive areas such as network management, customer incident handling and autonomous operations. The more responsibility an AI system takes on, the more important control over the underlying model becomes.

NVIDIA Is Building an Open AI Stack for Telecom

NVIDIA is becoming a major part of this shift through its Nemotron family of open models and the supporting NVIDIA NeMo software ecosystem.

The company’s tools allow telecom operators to customise models for specific jobs rather than treating AI as a single general-purpose application. Potential uses include network configuration, customer incident triage and increasingly autonomous network operations.

NVIDIA has also introduced the Nemotron 3 Large Telco Model, which is fine-tuned on telecommunications data with the aim of improving performance on industry-specific workloads.

This creates an interesting layer in the AI infrastructure market. The competition is no longer only about which company has the strongest general-purpose model. It is also about who can provide the models, datasets and tools that allow entire industries to build specialised AI systems.

Open Models Could Give Telecom Companies More Control Over Data

Privacy and governance sit near the centre of the open-model argument. Telecom operators process customer information and operational data that can be particularly sensitive, while their networks form part of critical communications infrastructure.

An open or open-weight model does not automatically make an AI system private or secure. Operators still need proper access controls, deployment architecture, monitoring and governance.

But having greater control over the model and where it runs can give operators more options when dealing with sensitive workloads. That flexibility is one reason open AI is becoming attractive alongside proprietary models rather than simply replacing them.

Cost Is Another Reason Operators Are Looking Beyond Proprietary Models

The economics become more significant when AI is running across a telecom company at huge volumes. Customer interactions, network monitoring and internal operations can generate enormous numbers of AI requests.

Open models can give operators more control over how those workloads are served and optimised. They can choose models according to the complexity of a task instead of automatically sending everything to an expensive frontier system.

The broader corporate market is moving in the same direction. Recent reporting has highlighted companies increasingly using open-weight models because of their lower operating costs, customisation options and ability to keep sensitive data within controlled infrastructure.

Open AI Could Become Important to Autonomous Networks

The most consequential use cases may sit inside network operations. Telecom companies are already working toward networks that can automatically detect problems, adjust resources and respond to changing demand.

AI-native network research is also moving toward agentic systems capable of intent-driven automation and closed-loop resource optimisation across network and computing infrastructure.

That puts a premium on models that understand telecom infrastructure rather than simply generating convincing text. A network agent needs to interpret technical information and operate within strict boundaries.

Open, domain-specific models could give operators more control over that process.

Telecom AI Is Moving Toward Industry-Specific Models

The bigger story is not simply that telecom companies are choosing open AI. It is that they are increasingly building AI specifically for telecommunications.

New research is producing specialised telecom models trained on areas such as 3GPP standards, O-RAN specifications, network telemetry and telecom software repositories. One recent open telecom research project, OTel, combines datasets, benchmarks and open-weight models designed for telecom-specific AI development.

Other research is exploring specialised models that can understand telecom operations and even assist with software development and production code workflows.

The direction is becoming clear: telecom AI is developing its own technical layer rather than remaining dependent on generic models.

The Future May Be a Mix of Open and Proprietary AI

Open models are unlikely to eliminate proprietary AI from telecom. Operators have different requirements for different workloads, and some applications may still benefit from the capabilities of closed frontier models.

The more likely outcome is a hybrid AI environment. An operator might use a proprietary model for one complex workload, an open model for another, and a smaller specialised model for a highly controlled network task.

That gives telecom companies something they have historically valued in infrastructure: choice.

For an industry built around massive networks and long-term technology investments, that flexibility could prove just as important as raw model performance.

Telecom Operators Are Building More of Their AI Infrastructure

The move toward open AI models signals a broader change in telecom technology strategy. Operators are no longer simply asking which AI provider they should buy from. They are asking which parts of the AI stack they should control themselves.

That includes models, data, infrastructure, governance and increasingly the agents that interact with network systems.

SoftBank, Indosat and AT&T show three different versions of the same direction: build models around the industry, adapt them to local requirements and maintain enough control to use AI inside critical operations.

For telecom, open AI is becoming less of an alternative technology choice and more of an infrastructure strategy.

Sources

Related Posts