November 22, 2026
Telecom Operators Shift to Open-Weight AI Models to Cut Network Costs and Retain Control

Telecom Operators Shift to Open-Weight AI Models to Cut Network Costs and Retain Control

NVIDIA Nemotron 3 helps telecom operators cut network costs with open-weight AI models and fine-tuned infrastructure

Telecommunications giants including AT&T, SoftBank, and Deutsche Telekom are building their artificial intelligence strategies on open-weight foundation models rather than relying exclusively on proprietary systems. This shift, accelerated by specialized releases like the NVIDIA Nemotron 3 Large Telco Model and the GSMA’s Open Telco AI initiative, allows carriers to fine-tune infrastructure for highly specific operational tasks while maintaining strict data governance and reducing inference costs.

How the Nemotron 3 Large Telco Model and AdaptKey Fine-Tune Network Operations

NVIDIA recently announced the 30-billion-parameter Nemotron 3 Large Telco Model to improve accuracy for specific industry tasks. AdaptKey fine-tuned this model on open telecom data to make it much more accurate for daily operations. Hugging Face model cards show the AdaptKey-Nemotron-30b version uses a specialized LoRA fine-tuning method. It was trained on over 1.3 million different telecom examples from various sources. This large dataset includes 3GPP standards, IETF protocols, network traces, anomaly detection, and network function configuration.

SoftBank Corp. is using these open models to build its own SoftBank Large Telecom Model. Rajeev Koodli is a principal fellow at SoftBank Corp. and senior vice president of SB Telecom America. He explained the main strategic benefit of this important choice for the company. He said, "Open models let SoftBank Corp. build on fast global progress." "We can also apply the network knowledge and operational expertise we gathered over many years."

NVIDIA released a full guide to help companies customize these models for their own needs. This lets operators adjust open models for their own networks using NVIDIA NeMo open libraries. NVIDIA also introduced Nemotron 3.5 Lightning to handle heavy computing tasks more efficiently. This is a 30-billion-parameter mixture-of-experts model with only 3 billion active parameters per token. It significantly lowers computing power for heavy agent tasks across the entire network. For harder jobs, the Nemotron 3 Ultra model uses a hybrid Latent Mixture-of-Experts setup. It has 550 billion total parameters and 55 billion active ones for complex tasks. This proves open models now reach top-level reasoning for complex scientific and agent tasks.

Why AT&T and the GSMA Open Telco AI Initiative Prioritize Open Weights

Companies are switching to open models to gain flexibility and better manage their overall costs. Andy Markus is the chief data and AI officer at the telecommunications company AT&T. He explained their view on this important shift in the broader industry. He said, "At AT&T, we believe the future of AI is not about choosing a single model." "It is about smartly matching every workload to the right mix of performance, cost and control."

This choice of models shows a wider industry push for standard open models across telecom. The GSMA notes that only “16% of Gen AI deployments target networks” and network operations. This is surprising since networks are the biggest cost center in the telecom industry. To fix this, the GSMA started the Open Telco AI initiative for better results. It offers custom models like the OTel-LLM-8.3B-Classification for specific and complex network needs. This model hits 99% accuracy on the TeleLogs benchmark for 5G network root cause analysis.

Teamwork across hardware and software is speeding up this change in the telecom industry. At the AMD Advancing AI conference, AT&T CTO Jeremy Legg and AMD SVP Dan McNamara introduced OTel 2.0. Red Hat reported that “OTel 2.0 has already surpassed 5 million downloads” since its recent release. This follows nearly 30 million downloads of the first OTel 1.0 models last year. This teamwork includes Dell and Microsoft to create a much better system. It mixes open-source methods with special hardware to train specific models without vendor lock-in. Also, the O-RAN alliance is studying cross-domain AI and generative uses in Open RAN setups. This points to a single industry standard for network intelligence in the future.

What Makes Telecom Network Data Incompatible With Standard LLMs

Switching to open, specific models directly answers the limits of general AI in complex networks. IBM notes that telecom network data comes from highly private sources within the companies. These sources include alarms, performance metrics, probes, and ticketing systems for daily operations. These systems capture incidents, defects, and configuration changes across the entire telecom network.

This data uses a special language with technical ideas from 5G, IP-MPLS, and other network protocols. Standard large language models are usually not trained on this specific type of data. So, using general models for telecom work often causes errors or bad sorting of issues. This is unacceptable in carrier networks where efficiency and safety are absolutely vital. Also, fast tech changes in telecom make things much harder for the companies. Static models quickly become outdated without constant, targeted training on new network information.

Open-weight models fix this by letting companies train the base AI with their own private data. Because the weights and training steps are open, telcos can enforce strict data rules. They can hide sensitive records and make safe fake datasets before the model sees the information. This keeps critical network layouts and customer data safe from any unauthorized access. It still lets them use advanced reasoning skills for their daily network operations.

How Deutsche Telekom and Red Hat Are Redesigning Workflows for Carrier-Grade AI

Using these models needs more than open weights, so it requires changing how the company works. Deutsche Telekom serves over 300 million customers and employs more than 200,000 people, and they plan to become an AI-native telecom provider. This is not just a simple software update, but a major change to how they make decisions and deliver services.

Jonathan Abrahamson, Chief Product & Digital Officer at Deutsche Telekom, explained this major change in their daily operations:

"Becoming AI-native is not just about adding AI to our current daily tasks, but it is about completely redesigning the work itself."

To help with this change, a working group is writing an open best practice guide. This is the Linux Foundation Networking (LFN) applied observability for artificial intelligence operations (AIOps) group, and Red Hat is the technical lead. They are building a simulated proof-of-concept for end-to-end observability across service provider network stacks, which ensures AI agents can reliably monitor and interact with complex infrastructure.

The GSMA is also providing practical tools to move these models from basic theory to actual use. The initiative offers the Satellite terminal UI for running benchmarks and submitting to leaderboards. It also includes an Agentic AI Testbed for evaluating AI agents against real-world scenarios. Finally, it provides Nika, a simulated network arena to test how well AI diagnoses and resolves injected faults.

What the Shift to Open Telco AI Means for Enterprise Network Engineers

For enterprise network engineers and IT workers, the growth of open telecom AI models changes how they troubleshoot and configure networks. In the past, finding a 5G core network fault or setting up an Open RAN architecture required deep manual analysis of proprietary logs. It also meant relying on black-box vendor tools to get the job done.

With open models like AdaptKey-Nemotron-30b and the OTel family, engineers can deploy specialized reasoning engines directly in their secure environments. The OTel open-source training repository works with any hardware, so teams do not have to use a single vendor’s chips to train or run these models. This gives them access to GPU power and open tools without huge infrastructure costs.

Also, routing tools like NVIDIA’s NeMo Switchyard let workers build systems that dynamically share computing power. A workflow can send high-volume log parsing to a small model like Nemotron 3.5 Lightning, but save a larger model for complex, multi-step root cause analysis. This control helps network teams scale their automated tasks clearly and easily. It turns open-weight AI from an experiment into a daily tool that lowers the industry’s biggest costs.