Why Telecom Operators Are Adapting Open AI Models for Network Operations

Telecom operators are turning to open AI models for more than lower costs. Open weights and training methods let them adapt models to network operations, choose where to run them and inspect how they behave. That flexibility matters when AI must work with operational data and support tasks such as network configuration, customer care and incident triage.

NVIDIA’s latest State of AI in Telecommunications report found that 89% of respondents consider open source models and software important to their company’s AI strategy. The figure points to a shift in how operators are thinking about model choice: rather than relying on one system for every task, they can combine open and closed models according to workload needs.

Open models can provide access to capable systems at lower cost, leaving closed models available for tasks where operators judge they offer greater value. The Artificial Analysis Intelligence Index v4.3.2 tracks competition from leading open models across demanding reasoning, coding, scientific and agentic tasks. Results on broad benchmarks, however, don’t establish how well a model will handle a particular operator’s network or procedures. That requires testing against the work it will actually perform.

Open weights and training recipes also give operators a path to tune models on network, customer and industry data. This can help a model interpret telecom terminology and follow workflows familiar to the operator. Greater access to model components can also make it easier for teams to evaluate behavior and adapt it to business requirements. Operators can then choose to run models across public cloud, private infrastructure or edge environments, depending on the workload and deployment constraints.

Adapting open foundations to telecom work

NVIDIA’s Nemotron family includes open models for reasoning and agentic workflows, as well as models with speech capabilities for voice applications. NVIDIA describes the family as providing weights, training data and recipes, resources that can support further adaptation and evaluation.

SoftBank Corp. is using open models as a base for telecom-specific development. The company draws on foundation models, including Nemotron, to develop its SoftBank Large Telecom Model and advance AI for network operations, design and management.

“Open models allow SoftBank Corp. to build on the rapid progress of global foundation models while applying the network knowledge and operational expertise we have accumulated over many years,” said Rajeev Koodli, principal fellow of SoftBank Corp. and senior vice president of SB Telecom America. He said the company is using open foundations to develop and continuously advance its telecom model for those operational uses.

NVIDIA has also described a 30-billion-parameter Nemotron 3 Large Telco Model, fine-tuned by AdaptKey on open source telecom datasets. The model is intended to improve performance on telecom-specific tasks, including interpreting industry terminology and reasoning through workflows such as network configuration and customer incident triage. The source provides no numerical accuracy result, so the reported improvement should be read as a description of the model’s aim rather than a quantified benchmark claim.

To help operators adapt this model or other open systems to their own networks, NVIDIA published a fine-tuning recipe using the NVIDIA NeMo libraries. The process is designed to use operator-specific networks, customers and procedures. A reusable recipe can lower the work needed to start adaptation, but each operator still has to determine whether its data is suitable and whether the tuned model performs reliably on its own tasks.

Production needs more than a model

Open weights alone don’t make an AI workflow ready for production. Operators also need ways to prepare data, manage access and turn model outputs into actions that fit existing processes. Where records contain sensitive information, the data pipeline may need to anonymize them or use synthetic examples designed to preserve privacy before fine-tuning begins.

AT&T’s approach emphasizes choosing models by workload. “At AT&T, we believe the future of AI is not about choosing a single model; it’s about intelligently matching every workload to the right combination of performance, cost and control,” said Andy Markus, the company’s chief data and AI officer. He said open models are part of that strategy and described production scale, reliability and governance as requirements for putting it into practice.

NVIDIA’s proposed platform combines data pipelines, open models, agent orchestration, secure runtimes and simulation. It is built around NVIDIA AI Enterprise and the NVIDIA Agent Toolkit. The components address different stages of deployment: preparing data, coordinating model-driven agents, running them in a controlled environment and testing workflows through simulation. Operators will still need to assess how those pieces fit their own systems and operational requirements.

The larger point is that model selection is only one part of the production decision. Data preparation, workflow design and oversight help determine whether an AI system can move from a demonstration to a useful network task. NVIDIA points to a partner ecosystem working across the platform; the practical result for operators will depend on how these tools perform together in deployment.

Local services depend on local adaptation

Operators can also host models for customers who want to use them directly or build applications on top of them. Fine-tuning for local languages and industries can make AI services more relevant to the people and organizations using them. That opportunity depends on access to suitable data, the ability to maintain the model and a deployment setup that meets local data-handling needs.

Indosat Ooredoo Hutchison, one of Indonesia’s largest telecom operators, is developing Sahabat-AI, a family of open source models intended to reflect Indonesian language and culture. Chirag Sukhadia, the company’s chief data and AI officer, said the work is intended to adapt AI to local language, culture, data and practical needs, and to support local applications.

For telecom operators, open models offer room to customize, inspect and deploy AI across different environments. Their value will depend on whether that control translates into better results on specific operational tasks—and whether data pipelines, testing and oversight are strong enough to support those systems in day-to-day use. NVIDIA technologies for telecommunications are one set of tools operators can evaluate as they build that infrastructure.