Telecom operators want control as well as lower cost
NVIDIA has detailed why telecommunications companies are adopting open models for network operations, customer care and local AI services. Its 6 October article cites a company survey in which 89 per cent of respondents said open-source models and software were important to their AI strategy. Survey results reflect the selected participants and NVIDIA's framing, but the operational motivations are clear: customisation, deployment flexibility and control over model artefacts matter in critical infrastructure.
Operators manage specialised terminology, large networks and sensitive customer data. A general closed model may be useful for some tasks, while an open model can be adapted to local procedures and run on infrastructure the operator controls. NVIDIA argues for matching workloads to different models instead of choosing one provider for everything. That approach requires a disciplined evaluation layer so flexibility does not become an unmanageable collection of model variants.
Nemotron gets a telecom-specific adaptation
The announcement introduces the 30-billion-parameter Nemotron 3 Large Telco Model, fine-tuned by AdaptKey on open telecom datasets. NVIDIA says it is intended to understand industry terminology and reason through workflows such as network configuration and customer-incident triage. The company positions it as a baseline that operators can adapt rather than a finished model for every network.
Domain tuning can improve relevance, but public telecom data will not capture each operator's equipment, topology or internal procedure. Local adaptation therefore needs representative examples and careful separation between training and evaluation sets. Teams should measure factual accuracy, tool selection, escalation and recovery from incomplete information. A plausible network command produced from outdated assumptions can be more dangerous than an obvious refusal.
The recipe is part of the release
NVIDIA has released an end-to-end fine-tuning recipe through its NeMo open libraries. The recipe is designed to help operators adapt Nemotron and other open models using their own operational data. Publishing a repeatable pipeline can reduce engineering effort and make choices about data preparation, training and evaluation easier to inspect than a fully managed black box.
Operational data often contains customer identifiers, network vulnerabilities and commercially sensitive information. Preparation must include access control, minimisation and anonymisation, with synthetic data used where appropriate. Teams should track the origin and permitted use of every dataset. Model weights and checkpoints may memorise sensitive material, so storage, distribution and deletion controls need to cover the training artefacts as well as the original records.
Production requires more than an open model
NVIDIA links the model work to AI Enterprise software and its Agent Toolkit for data pipelines, orchestration, secure runtimes and simulation. AT&T, SoftBank and Indosat are presented as examples of operators pursuing model choice or locally adapted AI. Vendor and customer statements demonstrate interest, not guaranteed results. Each workflow still needs performance, safety and cost evidence in the operator's environment.
Network operations are a high-consequence domain. An agent should begin in observation or recommendation mode, with people approving configuration changes. Simulation and digital twins can test proposals before they reach live equipment. Logs must show the model, data, tools and approvals involved. Open weights increase inspectability and control, but they do not automatically make behaviour safe or explainable.
Local AI can support language and sovereignty goals
Procurement should include that lifecycle cost in any comparison with a managed model. Open weights reduce some dependencies while creating new operational ones. The best choice may vary by workload, with sensitive network tasks kept local and lower-risk services using an external provider under contract.
Open models also shift responsibility to the operator. A hosted provider may manage abuse monitoring, patching and capacity, whereas a telecom company running its own weights must design those functions. Fine-tuned models need a release process, model cards and rollback when a new checkpoint performs worse. Security teams should scan dependencies and restrict access to training infrastructure, because altered data or weights can compromise many downstream services. The attraction of ownership is real, but ownership includes the ongoing burden of proving which model is running, how it was produced and whether it still meets the safety and reliability requirements of each network workflow.
NVIDIA also argues that operators can host models adapted to local languages, regulations and government requirements. Indosat's Sahabat-AI is cited as an example of building for Indonesian language and culture. Telecom companies already operate distributed infrastructure and trusted customer relationships, which may place them well to offer regional AI services.
The business case depends on sustained demand and the cost of maintaining models across hardware generations. Operators should compare local hosting with managed services, including staffing, energy, update and security costs. NVIDIA's open telco model and recipe expand the available options and make experimentation easier. The responsible path is to preserve that flexibility while applying rigorous data governance, independent evaluation and staged authority before models influence a live network.