Open telco AI: Training a model for an industry

Domain-specific artificial intelligence (AI) is becoming a practical priority for the telecommunications industry. At AMD Advancing AI, AT&T Chief Technology Officer Jeremy Legg joined AMD’s Dan McNamara, Senior Vice President and General Manager of Compute and Enterprise AI, to introduce OTel 2.0 and show how telecom-focused models are moving closer to production network use. The early response underscores that momentum: OTel 2.0 has already surpassed 5 million downloads since release, following nearly 30 million downloads of OTel 1.0 models.

Bringing together Red Hat, AT&T, AMD, Dell, Microsoft, and GSMA, this collaboration aims to advance carrier-grade AI. By combining open source methodologies, including the OTel hardware-agnostic open source training repository and the Red Hat open source repository for synthetic data generation (SDG Hub), with specialized infrastructure, the collaboration creates a practical blueprint for how an entire industry can train, secure, and scale domain-specific models without vendor lock-in.

Beyond model training, the Linux Foundation Networking (LFN)-applied observability for artificial intelligence operations (AIOps) working group, where Red Hat serves as technical lead, is building an open best practice guide and simulated proof-of-concept (PoC) for end-to-end observability across service provider network stacks. OTel 2.0 provides the models for the inference engine in that environment, interpreting telemetry, diagnosing faults against 3GPP standards, and generating actionable explanations in a closed-loop automation flow. The observability pipeline will surface the right data and OTel 2.0 helps make sense of it. Together, they close the gap between collecting telemetry and acting on it.

Synthetic data generation: Creating training data from telecommunication standards

Building AI that understands service provider environments requires training data grounded in industry standards. Generic AI models struggle with service provider-specific tasks because they lack exposure to the specialized language and concepts encoded in 3GPP specifications, IETF request for comments (RFCs), and complex network architectures.

AT&T used Red Hat's open source SDG Hub to transform raw telecommunications documentation into high-quality training data for OTel 2.0. GSMA provided the initial dataset of approximately 15 billion tokens by collecting and processing documents from seven standards development organizations, including 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN, and TM Forum, and converted dense technical specifications into material suitable for model training.

The SDG Hub's knowledge tuning flows combine 4 knowledge extraction strategies to create comprehensive training data:

• Document-based knowledge tuning processes complete sections directly to generate question and answer (Q&A) pairs from the full context.

• Extractive summary knowledge tuning pulls out the most important passages from long documents to focus on technically significant content.

• Detailed summary knowledge tuning creates thematic summaries that capture broader conceptual frameworks.

• Key facts knowledge tuning breaks documentation into atomic elements to build fine-grained knowledge for precise definitions.

Running all 4 flows on the same source material and combining the outputs builds a richer, more diverse training dataset. 

For example, a 3GPP specification describing the RRC connection establishment procedure might use extractive summary tuning to pull out the key procedural steps from a 50-page document, then generate Q&A pairs that teach models both what the procedure is and how it relates to broader 5G network operations.

From the current telecom corpus, this approach generates thousands of training examples. As the initiative scales across a broader telecom knowledge base, it can produce hundreds of thousands of instruction-tuned examples, all while maintaining technical precision.

Training the model

AT&T trained OTel 2.0 using supervised fine-tuning (SFT) with full weight training on AMD graphical processing unit (GPU) compute infrastructure. Supervised fine-tuning takes a pre-trained base model and trains it further on the specialized Q&A pairs generated from the corpus. This approach updates every parameter in the model, allowing it to internalize complex terminology and procedural knowledge.

Looking ahead, open source techniques like Orthogonal Subspace Fine-Tuning (OSFT) offer promising capabilities for future training cycles. This method is available through Red Hat’s open source Training Hub, and addresses a challenge in continual learning known as catastrophic forgetting. When models learn new tasks sequentially, they often lose performance on earlier tasks. This technique constrains parameter updates, preserving prior knowledge while learning new capabilities as service provider network technologies evolve. 

Bringing this to scale

The Open Telco AI initiative demonstrates how open collaboration accelerates innovation, but the true value comes from making these capabilities accessible to all industries. 

SDG Hub's knowledge tuning flows aren't service provider-specific, they work on any domain's technical documentation. A financial services company can use the same flows to create training data from regulatory documents. A healthcare organization can apply them to medical research papers, and a manufacturing company can process equipment specifications and maintenance manuals. The approach is universal: take your domain's authoritative documents, run them through knowledge tuning flows, and generate training data that teaches models your industry's language and concepts.

Organizations gain the tools to create specialized AI for their specific domains while maintaining full control over data residency and operational resilience. The open source foundation means the community can continuously improve the underlying capabilities. Whether deploying on-premises or in hybrid cloud environments, organizations can use these building blocks to accelerate their AI initiatives without vendor lock-in.

Making models production-ready

Deploying AI in production environments requires rigorous security analysis and operational hardening, which is where Red Hat's open source AI safety tools come in.

The same SDG Hub that generates training data also generates adversarial test cases to evaluate model safety  and help quantify AI risk exposure. Combined with Garak, an open source LLM vulnerability scanner, extended with domain-aware red teaming probes, and Guardrails for mitigation during production, the approach systematically tests models against prompt injection attacks, hallucination of network parameters, and attempts to bypass safety constraints.

As part of the AI red teaming initiative, Red Hat actively contributes to these open source projects to integrate security throughout the AI lifecycle from data generation to continuous monitoring in production. Rather than bolting on security after deployment, this approach bakes safety testing into the development workflow, making security-focused carrier-grade AI accessible without requiring deep security expertise in each component.

This integrated approach to security, validation, and reliability helps close the gap from demo to  production-grade AI that service providers can trust in their critical environments.

Why open collaboration drives value

The initiative succeeds because each partner brings critical capabilities. GSMA contributed approximately 15 billion raw tokens through Open Telco AI. AT&T used Red Hat's open source SDG Hub to process more than 1 trillion tokens on Microsoft Managed Compute, using approximately 530 GPUs for data preparation, primarily AMD MI300X GPUs. That work generated roughly 440 billion training tokens for OTel 2.0. AT&T then trained the model on on-premises AMD MI355X GPUs provided through Dell Technologies infrastructure and servers. Day 0 inference is available on Microsoft Foundry, Featherless AI, and Red Hat, with weekly model weight updates planned as training continues.

Open technology demonstrates that the best enterprise technology is built collaboratively, and with production deployment in mind from the start. The telecommunications industry runs on open standards. AI for service providers should too.