1. Executive Overview
The global telecommunications industry faces a unique technological paradox. While standard foundation artificial intelligence models demonstrate impressive general capabilities, they consistently struggle when deployed inside specialized enterprise environments. Modern telecom infrastructure relies on a vast, intricateweb of domain-specific standards, proprietary protocols, legacy hardware operational logs, and strict regulatory frameworks—nuances that generic off-the-shelf enterprise Large Language Models (LLMs) rarely capture.
To bridge this operational gap, AT&T has spearheaded the development of the Open Telco (OTel) model series: a family of open-source, domain-native AI architectures designed to inject deep telecommunications expertise directly into enterprise workflows. Following the widespread adoption of OTel 1.0—which surpassed 25 million downloads—AT&T undertook the engineering challenge of building OTel 2.0.
+-----------------------------------------------------------------------------------+
| AT&T OTel 2.0 WORKFLOW |
+-----------------------------------------------------------------------------------+
| Raw Telecom Data (GSMA Docs, Protocols, Standards) + Open-Source Synthetic Engine |
| │ |
| ▼ |
| Data Processing & Generation Engine (~1 Trillion Tokens) |
| ├─ Microsoft Phi-4 (~700B Tokens/Mo for Data Prep) |
| ├─ OSS-120B (High-Reasoning & Complex Validation) |
| └─ Gemma-4 (Task-Specific Fine-Tuning Workflows) |
| │ |
| ▼ |
| Microsoft Foundry Managed Compute Layer (~530 GPUs Total) |
| ├─ 430x AMD Instinct™ MI300X GPUs (High-Memory Workloads) |
| └─ Heterogeneous NVIDIA GPU Instances |
| │ |
| ▼ |
| Final Trained OTel 2.0 Model Corpus (~400 Billion Tokens) |
| │ |
| ▼ |
| Democratized Ecosystem Release (GSMA Open Telco Alliance) |
+-----------------------------------------------------------------------------------+Developing OTel 2.0 required solving a broader structural problem facing global enterprises: how to build, fine-tune, and deploy highly specialized, domain-specific AI systems at enterprise scale while maintaining strict operational efficiency, hardware flexibility, and financial restraint.
By leveraging Microsoft Foundry Managed Compute, AT&T abandoned rigid single-model paradigms in favor of a flexible multi-model approach using high-performing open-source architectures (including Phi-4, OSS-120B, and Gemma-4). Grounding this workload across a cluster of approximately 530 dedicated GPUs—including 430 AMD Instinct™ MI300X accelerators—AT&T successfully processed roughly 1 trillion tokens and trained OTel 2.0 on a refined corpus of 400 billion tokens.
This strategy yielded major technical and economic advantages: by using open-source models for synthetic data generation and data preparation served via managed compute rather than relying on proprietary frontier model APIs, AT&T saved tens of millions of dollars, proving that specialized industrial AI can be deployed at scale without cost overruns.
2. The Domain Specificity Deficit in Telecommunications
Why Generic Foundation Models Fall Short
General-purpose artificial intelligence models are trained primarily on broad web crawls. While this gives them strong conversational fluency and wide baseline knowledge, it creates distinct failure modes when applied to critical network operational environments:
- Standards and Regulatory Complexity: Telecom operations are governed by dense, highly technical standards bodies such as the 3rd Generation Partnership Project (3GPP), the GSM Association (GSMA), and the Open RAN (O-RAN) Alliance. Standard models often hallucinate standard numbers, confuse frequency allocation rules, or misinterpret core protocol interfaces.
- Context-Heavy Diagnostics: Troubleshooting mobile core networks, radio access networks (RAN), and backhaul routing requires context windows filled with complex telemetry metrics, system log events, and signal degradation parameters. Generic LLMs lack the domain logic required to correlate these disparate signals into root-cause analyses.
- Data Security and Sovereignty: Telecommunications data contains sensitive subscriber telemetry, structural metadata, and mission-critical network topologies. Sending raw telecommunications data through commercial third-party public API endpoints often conflicts with corporate compliance standards and national data sovereignty regulations.
The Evolution of OTel
AT&T launched the Open Telco initiative to establish an open standard for telco-native AI models. The success of OTel 1.0 demonstrated a strong demand across the sector, amassing over 25 million downloads. However, building OTel 2.0 required expanding beyond basic network concepts into advanced multi-step reasoning, synthetic data generation, and complex diagnostic pipelines.
Achieving this required processing massive datasets, creating an execution strategy that balanced computational scale, model selection, and hardware infrastructure.
3. Detailed Chronology: Building OTel 2.0
Phase 1: Formulating the Multi-Open-Model Strategy
Rather than relying on a single monolith model, AT&T implemented a modular, multi-open-model architecture. By deploying open models via Microsoft Foundry, AT&T retained total visibility and control over model training pipelines, data governance protocols, and unit economics.
The model stack was organized into distinct operational tiers:
- Synthetic Data Generation and Processing Engine (Phi-4): To train OTel 2.0 effectively, AT&T needed to transform massive troves of unstructured technical documentation—including raw specifications from the GSMA—into clean, structured instruction pairs. Microsoft’s Phi-4 served as the high-throughput driver for this workflow, processing around 700 billion tokens per month to perform document parsing, synthetic dialog generation, and dataset enrichment.
- Reasoning and Logical Validation (OSS-120B): For complex network domain reasoning, architectural evaluation, and protocol verification, AT&T utilized OSS-120B. This larger parameter model provided the deep contextual reasoning needed to generate synthetic troubleshooting paths and complex network logic.
- Development Workflow & Targeted Fine-Tuning (Gemma-4): Google’s Gemma-4 model family was deployed across development workflows to assist in refining domain task vectors, fine-tuning downstream model behavior, and evaluating interim OTel 2.0 checkpoints.
+----------------------------------------------+
| RAW GSMA & TELECOM DOCUMENTATION |
+----------------------------------------------+
│
▼
+----------------------------------------------+
| SYNTHETIC DATA PIPELINE |
| Phi-4: High-Throughput Token Generation |
| (~700 Billion Tokens / Month Executed) |
+----------------------------------------------+
│
▼
+----------------------------------------------+
| HIGH-REASONING VALIDATION |
| OSS-120B: Protocol & Logical Verification |
+----------------------------------------------+
│
▼
+----------------------------------------------+
| TRAINING DATASET CORPUS |
| ~1 Trillion Tokens Processed & Curated |
| ~400 Billion High-Quality Tokens Selected |
+----------------------------------------------+
│
▼
+----------------------------------------------+
| OTEL 2.0 TRAINING |
| Executed via Foundry Managed Compute |
| (430x AMD MI300X + Heterogeneous GPUs) |
+----------------------------------------------+Phase 2: Processing 1 Trillion Tokens
Building OTel 2.0 required ingesting, synthesizing, and training on an immense corpus of network data:
- Ingestion & Augmentation: AT&T gathered standard technical literature, operational telemetry formats, and GSMA specifications, assembling an initial raw corpus.
- Synthetic Expansion: Using Phi-4, AT&T expanded this foundational core with synthetic technical queries, diagnostic edge-case scripts, and network configuration scenarios. In total, the system processed approximately 1 trillion tokens.
- Filtering and Curation: The raw combined dataset was processed down to a refined, high-density training set of approximately 400 billion tokens, forming the primary training foundation for OTel 2.0.
Phase 3: Provisioning Managed Infrastructure
Traditionally, managing thousands of specialized compute nodes requires weeks or months of operational overhead, including setting up driver stacks, orchestrating compute clusters, managing thermal throttling, and balancing load across networks.
AT&T bypassed this infrastructure friction by deploying on Microsoft Foundry Managed Compute. The platform enabled AT&T to provision a dedicated heterogeneous GPU cluster of approximately 530 GPUs in a matter of days. A core component of this deployment was the integration of 430 AMD Instinct™ MI300X GPUs, working alongside additional NVIDIA GPU instances within a unified managed environment.
4. Compute Economics & Silicon Strategy
The Financial Advantage of Dedicated Compute vs. Frontier APIs
For enterprise-scale AI projects, unit economics are a primary constraint. Processing trillions of tokens using closed-source, commercial frontier APIs often introduces severe financial friction.
- Frontier API Economic Model: In a standard third-party commercial API paradigm, generating 1 trillion synthetic tokens at enterprise rates ($2.50 to $10.00+ per million tokens depending on input/output split and model class) can push raw data synthesis costs into tens of millions of dollars before primary model training even begins.
- Dedicated Managed Open-Model Strategy: By serving open architectures like Phi-4, Gemma-4, and OSS-120B on dedicated compute instances via Microsoft Foundry Managed Compute, AT&T decoupled token processing costs from API pricing models.
By leveraging dedicated hardware for high-throughput synthetic data generation, AT&T saved tens of millions of dollars compared to proprietary frontier APIs. These savings directly enabled AT&T to run broader experimentation cycles and expand the scale of OTel 2.0.
COST COMPARISON MODEL (DATA SYNTHESIS AT 1 TRILLION TOKENS)
Frontier API Model:
[░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░] Tens of Millions ($10M - $30M+)
Foundry Managed Compute + Open Models (AT&T):
[░░░░░░░░░░░░] Fraction of API Cost (Saved Tens of Millions)Silicon Heterogeneity: AMD Instinct™ MI300X Integration
AT&T’s deployment of 430 AMD Instinct™ MI300X GPUs represents a major real-world validation of multi-vendor GPU strategies in top-tier enterprise AI environments.
+-------------------------------------------------------------------------------+
| HETEROGENEOUS COMPUTE CLUSTER ALLOCATION |
+-------------------------------------------------------------------------------+
| Hardware Allocation: ~530 Total Dedicated Accelerators |
| |
| [AMD Instinct™ MI300X GPUs] ################################### 430 Units |
| (High HBM3e Memory Density) (81% of Total Cluster Capacity) |
| |
| [NVIDIA GPU Instances] ######## 100 Units |
| (Targeted Parallel Tasks) (19% of Total Cluster Capacity) |
+-------------------------------------------------------------------------------+The AMD Instinct MI300X architecture, featuring 192GB of HBM3 memory capacity and high memory bandwidth, is well-suited for high-throughput batch synthetic data generation and large-parameter token execution. By decoupling software deployments from a single vendor’s hardware ecosystem, AT&T established operational flexibility:
- Workload-Specific Hardware Allocation: Compute tasks were routed dynamically to the hardware platform best suited for the job, matching memory-bandwidth-bound inference workloads to high-capacity AMD nodes.
- Mitigating Supply Chain Bottlenecks: Leveraging multiple silicon architectures shielded the project from single-vendor hardware availability bottlenecks, keeping the development timeline on schedule.
- Rapid Onboarding: Microsoft Foundry Managed Compute abstracted driver orchestration and hardware setup, reducing compute deployment times from weeks down to days.
5. Supporting Context & Key Metrics
The quantitative parameters underlying the OTel model program highlight the operational scale required to execute specialized domain AI:
Data and Compute Specifications
| Parameter Metric | Empirical Value / Volume | Operational Role / Significance |
|---|---|---|
| OTel 1.0 Total Downloads | > 25 Million | Established open domain model standard for global telco network operators |
| Total GPUs Deployed | ~530 Accelerators | Dedicated compute infrastructure managed via Microsoft Foundry |
| AMD Silicon Density | 430 AMD Instinct™ MI300X GPUs | High-bandwidth memory instances used for data preparation and fine-tuning |
| Raw Tokens Processed | ~1 Trillion Tokens | Ingested GSMA standards, network specifications, and synthetic data |
| Trained Dataset Size | ~400 Billion Tokens | Filtered, domain-specific dataset used for core OTel 2.0 training |
| Phi-4 Data Generation Scale | ~700 Billion Tokens / Month | High-throughput synthetic data synthesis and documentation parsing |
Model Function Matrix
| Model Architecture | License / Model Type | Primary Task Assignment | Unique Operational Value |
|---|---|---|---|
| Phi-4 | Open Model | Synthetic data generation & data pre-processing | Processed ~700B tokens/mo; dramatically lowered pre-training API generation costs |
| OSS-120B | Open Model | High-reasoning evaluation & protocol logic | Handled dense multi-step logic validation and network diagnostic paths |
| Gemma-4 | Open Model | OTel 2.0 development workflows & task adaptation | Supported fine-tuning loops and modular domain performance testing |
| OTel 2.0 | Community Open Telco Standard | Telco-native foundation model | Open domain-specific foundation model for the global telecom ecosystem |
6. Official Statements & Expert Analysis
Industry leaders from Microsoft, Hugging Face, and AT&T emphasized how open models, combined with managed compute platforms, enable specialized AI engineering at scale:
"Every company in the world needs to build its own AI, and that is only possible with open models and open source. AT&T is championing this vision, building on open models like Phi-4 and Gemma, and giving OTel back to the community as a telecom AI foundation others can build upon. Microsoft Foundry makes this practical at scale, bringing the latest open models from the Hugging Face collection together with AMD and NVIDIA GPUs in one place, so teams can pick the right model and the right hardware, then deploy in hours instead of weeks."
— Jeff Boudier, Vice President of Product, Hugging Face
Boudier’s perspective points to an industry-wide trend: proprietary models are useful for general workflows, but building sovereign, enterprise-owned AI requires fine-tuning open architectures over dedicated compute. The ability to deploy open-source models paired with hardware flexibility—switching between AMD and NVIDIA architectures without rewriting core infrastructure code—is becoming a cornerstone of enterprise AI strategy.
"When you are processing hundreds of billions of tokens, infrastructure becomes part of the problem you solve. Foundry Managed Compute gave us access to GPU capacity at scale so our teams could focus on advancing OTel2.0 instead of managing infrastructure."
— Mark Austin, Vice President, Data Science and AI at AT&T
Austin highlights an operational truth in enterprise AI development: at extreme token volumes, raw computing infrastructure ceases to be an abstract backend utility; it becomes an active variable in the engineering equation. Systems like Microsoft Foundry Managed Compute remove low-level orchestration friction, allowing data science teams to focus on domain accuracy and data quality rather than hardware management.
7. Industry Implications & Future Outlook
Democratizing Domain-Specific AI
The release of OTel 2.0 back to the telecommunications ecosystem marks a major step forward for vertical enterprise AI. By contributing OTel 2.0 to the global community through frameworks like the GSMA Open Telco initiative, AT&T provides smaller telecommunications operators, virtual network operators (MVNOs), and network equipment vendors access to a telecom-native foundation model without requiring them to invest millions in raw data generation.
+-----------------------------------------------------------------------------------+
| GLOBAL TELECOM AI ECOSYSTEM IMPACT |
+-----------------------------------------------------------------------------------+
| AT&T OTel 2.0 Core Model Architecture |
| ├─ Ingests 3GPP / GSMA Standards & Local Operational Telemetry |
| └─ Fine-Tuned via Open-Source Pipeline on Heterogeneous Hardware |
| │ |
| ▼ |
| Distributed via GSMA & Open-Source Community Ecosystem |
| │ |
| ├───────────────► Network Equipment Vendors (Automated RAN Tuning) |
| ├───────────────► Global Network Operators (Autonomous System Diagnostics) |
| └───────────────► Enterprise Ecosystem (Custom Vertical Integrations) |
+-----------------------------------------------------------------------------------+Key Takeaways for Enterprise AI Operations
- Synthetic Data Realizes Massive Savings: Using specialized open models like Phi-4 to generate domain training data can cut pre-training costs by tens of millions of dollars compared to using commercial frontier API platforms.
- Heterogeneous Hardware Strategy Delivers Flexibility: Combining hardware options—such as AMD Instinct MI300X accelerators and NVIDIA GPUs—prevents vendor lock-in, optimizes memory-bound inference costs, and mitigates supply chain delays.
- Managed Abstraction Accelerates Development: Leveraging managed execution environments like Microsoft Foundry Managed Compute eliminates low-level cluster orchestration overhead, shortening infrastructure setup timelines from weeks down to days.
As artificial intelligence shifts from exploratory research to core enterprise infrastructure, the OTel 2.0 project provides a clear operational template. By pairing open-source models with dedicated, managed hardware ecosystems, enterprises can efficiently build high-performing, domain-specific AI models that deliver real commercial value at scale.
