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

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, which was fine-tuned by AdaptKey on open-source telecom datasets to deliver accuracy gains for industry-specific tasks. According to Hugging Face model cards, the AdaptKey-Nemotron-30b variant utilizes a LoRA fine-tuning approach trained on over 1.3 million telecom domain examples. This extensive dataset encompasses 3GPP standards, IETF protocols, network traces, anomaly detection, and network function configuration.

SoftBank Corp. is actively utilizing these open foundations to develop its SoftBank Large Telecom Model. Rajeev Koodli, principal fellow of SoftBank Corp. and senior vice president of SB Telecom America, explained the strategic advantage: "Open models allow SoftBank Corp. to build on the rapid progress of global foundation models while applying the network knowledge and operational expertise we have accumulated over many years."

To support this customization, NVIDIA released a full recipe detailing the end-to-end fine-tuning pipeline. This allows operators to adapt open models to their specific networks using NVIDIA NeMo open libraries. Furthermore, NVIDIA introduced Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model with only 3 billion active parameters per token, designed to reduce inference compute for high-volume agent workloads. For even more demanding tasks, the Nemotron 3 Ultra model offers a hybrid Latent Mixture-of-Experts architecture with 550 billion total parameters and 55 billion active parameters, demonstrating that open models are now achieving frontier-level reasoning performance across complex scientific and agentic workloads.

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

The transition to open models is driven by the need for operational flexibility and cost management. Andy Markus, chief data and AI officer of AT&T, stated, "At AT&T, we believe the future of AI is not about choosing a single model; it’s about intelligently matching every workload to the right combination of performance, cost and control."

This model-choice strategy is evident in the broader industry push for standardized open models. According to the GSMA, only “16% of Gen AI deployments target networks” and network operations, despite networks representing the largest cost center in the telecommunications industry. To address this gap, the GSMA launched the Open Telco AI initiative, providing purpose-built models like the OTel-LLM-8.3B-Classification, which achieves 99% accuracy on the TeleLogs benchmark for 5G network root cause analysis.

Collaboration across the hardware and software stack is accelerating this adoption. At the AMD Advancing AI conference, AT&T Chief Technology Officer Jeremy Legg and AMD’s Senior Vice President Dan McNamara introduced OTel 2.0. Red Hat reported that “OTel 2.0 has already surpassed 5 million downloads” since its release, following nearly 30 million downloads of the initial OTel 1.0 models. This collaboration, which includes Dell and Microsoft, combines open-source methodologies with specialized infrastructure to create a blueprint for training domain-specific models without vendor lock-in. Concurrently, the O-RAN alliance is investigating cross-domain AI and generative use cases within Open RAN architectures, signaling a unified industry standard for network intelligence.

What Makes Telecom Network Data Incompatible With Standard LLMs

The pivot toward open, domain-specific models is a direct response to the limitations of general-purpose artificial intelligence in complex network environments. IBM notes that telecom network observability data originates from highly proprietary sources, encompassing alarms, performance metrics, probes, and ticketing systems that capture incidents, defects, and configuration changes.

This data is deeply embedded in a domain-specific language featuring technical concepts from 5G, IP-MPLS, and other network protocols. Standard foundational large language models are typically not trained on this specialized telemetry. Consequently, applying general models to telecom operations often results in hallucinations or inaccurate triage, which is unacceptable in carrier-grade environments where operational efficiency and safety are paramount. The rapid pace of technological change in telecommunications further complicates matters, as static models quickly become outdated without continuous, targeted fine-tuning.

Open-weight models solve this by allowing operators to fine-tune the base intelligence using their own proprietary data. Because the weights and training recipes are accessible, telcos can implement strict data governance, anonymizing sensitive records and generating privacy-preserving synthetic datasets before the model ever processes the information. This ensures that critical network topology and customer data remain secure while still benefiting from advanced reasoning capabilities.

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

Integrating these models requires more than just accessible weights; it demands a fundamental restructuring of operational workflows. Deutsche Telekom, which serves more than 300 million customers and employs over 200,000 people, has set a strategic objective to become an AI-native telecommunications provider. The company views this not as a supplementary software rollout, but as a core transformation of decision-making and service delivery.

Jonathan Abrahamson, Chief Product & Digital Officer at Deutsche Telekom, summarized this operational shift:

"Becoming AI-native is not about adding AI to the way we work today. It is about redesigning the work itself."

To support this redesign, the Linux Foundation Networking (LFN)-applied observability for artificial intelligence operations (AIOps) working group, with Red Hat serving as technical lead, is developing an open best practice guide. This group is building a simulated proof-of-concept for end-to-end observability across service provider network stacks, ensuring that AI agents can reliably monitor and interact with complex infrastructure.

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

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

For enterprise network engineers and IT practitioners, the proliferation of open telecom AI models fundamentally changes how network troubleshooting and configuration are executed. Historically, diagnosing a 5G core network fault or configuring an Open RAN architecture required deep manual analysis of proprietary logs or reliance on black-box vendor tools.

With open models like AdaptKey-Nemotron-30b and the OTel family, engineers can deploy specialized reasoning engines directly within their own secure environments. The hardware-agnostic nature of the OTel open-source training repository means that teams are not forced to adopt a single vendor’s silicon to train or run inference on these models, granting them access to GPU capacity and open toolchains without prohibitive infrastructure costs.

Furthermore, the introduction of routing tools like NVIDIA’s NeMo Switchyard allows practitioners to build multi-agent systems that dynamically allocate compute. An engineering workflow can route high-volume, routine log parsing to a lightweight model like Nemotron 3.5 Lightning, while reserving a larger, more complex model for intricate, multi-step root cause analysis. This granular control ensures that network operations teams can scale autonomous workflows transparently, turning open-weight AI from an experimental concept into a daily operational utility that directly addresses the industry’s largest cost centers.