NVIDIA Ships a 30-Billion-Parameter Open Telco Model as 89% of Operators Call Open AI Core Strategy

NVIDIA put a number on something telecom operators have been saying quietly for a year. On Monday the company released the Nemotron 3 Large Telco Model (LTM) — a 30-billion-parameter open-weight model fine-tuned by AdaptKey on open-source telecom datasets — alongside a blog post arguing that operators are building their AI strategies on open models, and its own survey data explaining why. Per NVIDIA's latest State of AI in Telecommunications report, 89% of respondents said open-source models and software are important to their company's AI strategy.

The model itself is pitched as a domain-tuned starting point rather than a finished product. NVIDIA says it understands telecom terminology out of the box and is aimed at the kind of work operators actually run: network configuration and customer incident triage. Alongside the weights, NVIDIA published the full end-to-end fine-tuning recipe through its NeMo open libraries, so an operator can adapt the LTM — or any other open model — to its own networks, customers and procedures. All capability and performance descriptions are NVIDIA's characterizations and have not been independently verified.

The strategic case NVIDIA lays out is fivefold: lower-cost access to frontier-level intelligence, with closed models reserved for the workloads that justify them; telco-specific customization built on open weights and published recipes; greater visibility into model artifacts and behavior for governance and regulatory alignment; deployment flexibility across public clouds, private infrastructure and edge; and the ability to package locally adapted AI services for enterprise and government customers. The company also points to the Artificial Analysis Intelligence Index v4.3.2 as third-party evidence that leading open models are closing the gap on reasoning, coding, scientific and agentic work.

Three operators supplied quotes, and each describes a different posture rather than a single use case. Rajeev Koodli, principal fellow at SoftBank Corp. and senior vice president of SB Telecom America, said the company uses open foundations extensively, including Nemotron models, in building its SoftBank Large Telecom Model for network operations, design and management. AT&T chief data and AI officer Andy Markus framed the issue as matching each workload to the right combination of performance, cost and control rather than standardizing on a single model. Chirag Sukhadia, his counterpart at Indonesia's Indosat Ooredoo Hutchison, pointed to Sahabat-AI, an open-model family adapted to Indonesian language, culture and data.

NVIDIA is also explicit about the catch: models alone do not get autonomous network operations into production. By its own account, operators additionally need data pipelines that anonymize sensitive records and generate privacy-preserving synthetic datasets for fine-tuning, plus agent orchestration, secure runtimes and simulation to turn open models into governed agentic workflows. That gap is exactly what NVIDIA says its commercial stack — NVIDIA AI Enterprise software and the Agent Toolkit — fills, which makes the open-model argument also an argument for the paid platform wrapped around it.

Still, the direction is clear. Telecom is becoming the proving ground for vertical open models: a domain where data residency, regulation and mission-critical reliability make "we can inspect it, fine-tune it and keep it inside our network" worth more than a few benchmark points. If the weights-plus-recipe pattern sticks, the obvious next candidates are the other industries with the same constraints — utilities, banking and healthcare — each of which has the data volume and the compliance burden to justify its own domain-tuned open model.