NVIDIA argues open AI models can help telecom operators deploy AI with more control

Telecom operators are changing how they plan and deploy artificial intelligence, shifting attention away from a single “best” closed model and toward open models that operators can evaluate, adapt, and govern internally. In a new NVIDIA analysis of AI in telecommunications, the company says 89% of respondents report that open-source models and software matter to their AI strategy, citing benefits that extend beyond cost.

The core argument is that open models can be trusted and controlled in ways that matter for telecom. Operators handle sensitive customer information, run safety- and reliability-critical network functions, and often operate within strict data governance requirements. NVIDIA’s framing is that open models allow telcos to customize behavior and workflows across high-impact tasks, including autonomous network operations and customer service, rather than treating AI systems as black boxes.

NVIDIA points to its Nemotron family of open models, describing them as frontier-level reasoning models optimized for agentic workflows, along with speech capabilities for voice applications. The company says Nemotron models are available with open weights, training data, and recipes, aiming to make it easier for operators to start from an industry-ready baseline and adapt to their own operational context.

The report highlights examples of how operators can use open models as a foundation for telecom-specific AI. SoftBank Corp. is cited as using open models to build and continuously advance what NVIDIA calls its telecom-specific capabilities, including work applied to network operations, network design, and overall management. Rajeev Koodli, a principal fellow at SoftBank Corp. and senior vice president of SB Telecom America, is quoted saying open models let SoftBank build on global foundation model progress while applying network knowledge accumulated over many years, and that it uses open foundations extensively, including NVIDIA Nemotron models and others, to develop its SoftBank Large Telecom Model.

NVIDIA also describes a specific product step for tailoring open models to telecom tasks. The company announced a 30-billion-parameter Nemotron 3 Large Telco Model, which it says was fine-tuned by AdaptKey on open-source telecom datasets. NVIDIA characterizes the result as an open baseline meant to understand telecom terminology and to reason effectively through operational workflows such as network configuration and customer incident triage.

Beyond selecting models, the report emphasizes that deployment requires more than model choice. NVIDIA says operators need pipelines that prepare and protect data for fine-tuning, including anonymizing sensitive records and generating privacy-preserving synthetic datasets. The company also argues that operators need platforms capable of turning open models into governed, autonomous agentic workflows, with appropriate security controls and an operational way to validate performance before systems are put into production.

To support that end-to-end approach, NVIDIA says it provides an integrated platform powered by NVIDIA AI Enterprise software and the NVIDIA Agent Toolkit. The company says the stack spans data pipelines, open models, agent orchestration, secure runtimes, and simulation, positioning it as a practical route from telecom data to production-ready AI workflows. The report further argues that a partner ecosystem can fill gaps across the layers needed for production deployment.

The telecom context in the report is also about flexibility. NVIDIA quotes Andy Markus, AT&T’s chief data and AI officer, arguing that AI’s future is not about choosing a single model, but matching each workload to the right combination of performance, cost, and control. Markus is quoted saying open models are essential to that approach, and that partnering with NVIDIA helps bring the strategy into production with what AT&T describes as the scale, reliability, and governance required by its business.

Not everything is specified in the report. NVIDIA does not disclose detailed metrics for performance gains, deployment timelines, or how operators benchmark “control” and “governance” across different fine-tuning approaches and data environments. The document also does not spell out which specific open datasets, evaluation methods, or runtime constraints are used for each named operator beyond the general description of anonymization, synthetic data generation, and agentic workflow governance.