NVIDIA launches open AI stack for telecom networks

The rollout pairs a 30-billion parameter Nemotron model with containment tools so carriers can run local network agents without sending data to closed clouds.

NVIDIA formally introduced a software stack and open models on October 6, 2026, aimed at helping telecommunications carriers build autonomous network agents. In a technical post by Kanika Atri on the company’s official blog, the chipmaker stated that major telecom operators are steering clear of closed proprietary APIs to protect operational sovereignty and retain control over network traffic data. The release highlighted adoption by carriers including AT&T, SoftBank, and Indonesia’s Indosat Ooredoo Hutchison.

First: why open models matter in network infrastructure

Telecommunications providers manage regulated customer information, follow strict privacy mandates, and cannot tolerate unexpected network outages. When an operator sends call records or routing topologies to external proprietary cloud endpoints, it loses the ability to inspect how those inputs are handled and risks disruption from unannounced model changes.

With open-weight models, code runs inside the carrier’s private facilities. Engineering teams can examine weights, tune systems on internal operational records, and automate monitoring workflows without sending sensitive communications data outside their corporate borders.

The 30B Nemotron model and carrier deployments

The centerpiece of the release is the Nemotron 3 Large Telco Model (LTM), a 30-billion parameter model posted to Hugging Face by technical partner AdaptKey. The model was adjusted using telecom-specific datasets to identify hardware alarms and parse device configuration scripts, paired with open workflows for further domain tuning.

Several international carriers confirmed active projects:

• AT&T: Andy Markus, Chief Data and AI Officer, stated in a joint release that the carrier evaluates specific models for each task to balance operating costs, operational oversight, and latency with vendor support.
• SoftBank: Rajeev Koodli confirmed that the Japanese operator uses Nemotron base weights to build an in-house model for cellular network operations.
• Indosat: in Indonesia, the carrier employs the open framework for its Sahabat-AI project to meet national regulations and handle regional dialects.

To back its product positioning, NVIDIA cited metrics from its “State of AI in Telecommunications 2026” survey. In that vendor-sponsored report, 89% of surveyed telecom executives stated that open-source software and models are vital to their plans, while 66% reported active AI adoption in their operations, up from 49% in the previous survey.

Containment tools and current operational limits

To keep autonomous agents from triggering erroneous commands that could disable cellular links, NVIDIA packaged isolation features. The NVIDIA OpenShell environment and NemoClaw templates establish digital boundaries, specifying which commands an agent can test in simulation before engineers allow production execution. The release also includes NeMo Anonymizer and NeMo Safe Synthesizer to mask subscriber details and generate synthetic test logs.

Despite commercial interest in self-healing infrastructure, fully autonomous operation remains out of reach. Technical coverage from trade outlet RCR Wireless notes that live telecommunications networks still require human operators in the loop to review high-impact routing changes and hardware reconfigurations.

Additionally, independent studies verifying measurable operational cost reductions from this stack have not been published, and contract terms between NVIDIA and participating carriers remain undisclosed. The current rollout highlights a practical infrastructure preference: major carriers would rather run dedicated models on their own hardware than hand network operations over to third-party cloud providers.

Sources

• NVIDIA Official Blog — Announcement on open models for telecom operators
• NVIDIA State of AI in Telecommunications 2026 — Annual telecom industry survey
• RCR Wireless News — Technical analysis on telecom autonomy boundaries
• Hugging Face / AdaptKey — Nemotron 30B technical model repository
• @NVIDIAAI on X — Official announcement of the telecom stack