Executive Overview
As hyper-scale artificial intelligence reshapes corporate enterprise strategy, the global telecommunications sector faces a unique structural hurdle: commercial frontier models, despite their broad fluency, routinely fail to grasp the hyper-specialized operational realities of modern network engineering. Telecommunications networks operate on complex layers of proprietary telemetry, intricate 3GPP (3rd Generation Partnership Project) standards, localized fault-isolation protocols, and complex billing architectures. For industry giants like AT&T, generic large language models (LLMs) are simply inadequate for deep operational execution.
To solve this domain-knowledge deficit, AT&T initiated the Open Telco (OTel) model series—an ambitious open-weights AI foundation built specifically for telecommunications engineering, operations, and network orchestration. Following the overwhelming adoption of its predecessor, OTel 1.0, which generated over 25 million downloads globally, AT&T set out to architect OTel2.0.
+--------------------------------------------+
| AT&T OTel 2.0 |
| Domain-Specific Telco Foundation |
+--------------------------------------------+
^
|
+-----------------------------------+-----------------------------------+
| | |
+---------------------------+ +---------------------------+ +---------------------------+
| Phi-4 | | OSS 120B | | Gemma 4 |
| Data Prep & Synthetic | | Complex Reasoning & | | Development Workflows |
| (~700B Tokens/Month) | | Logic Pipelines | | & Optimization |
+---------------------------+ +---------------------------+ +---------------------------+
^ ^ ^
+-----------------------------------+-----------------------------------+
|
v
+--------------------------------------------+
| Microsoft Foundry Managed Compute |
| Unified Compute & Governance Layer |
+--------------------------------------------+
^
v
+--------------------------------------------+
| Heterogeneous Infrastructure Cluster |
| ~530 GPUs (430 AMD MI300X + NVIDIA Units) |
+--------------------------------------------+Building OTel2.0 required processing approximately 1 trillion tokens of domain data and executing fine-tuning across 400 billion tokens. Achieving this scale presented immense operational and financial challenges: managing infrastructure overhead, preventing runaway token-consumption costs, and orchestrating thousands of GPU hours across diverse hardware architectures.
By partnering with Microsoft and leveraging Microsoft Foundry Managed Compute, AT&T abandoned the high-friction model of self-managed GPU deployments. Utilizing a heterogeneous infrastructure fleet—including roughly 430 AMD Instinct™ MI300X GPUs alongside NVIDIA hardware—and a suite of open-weights models (including Phi-4, OSS-120B, and Gemma 4), AT&T constructed an agile, cost-effective AI development engine. The initiative yielded a deployable telecom foundation while saving tens of millions of dollars compared to proprietary API endpoints, cutting provisioning cycles from weeks to days.
Detailed Chronology: The Engineering and Development Narrative of OTel2.0
Phase 1: The Domain Gap and the Telecom Imperative
The transition from generic enterprise software to autonomous, self-healing network orchestration demands high-precision domain understanding. In early multi-modal and natural language testing, standard closed-source models frequently hallucinated network topology states, misapplied signaling standards, and lacked the context required to parse dense GSMA (GSM Association) technical documentation.
Recognizing that general-purpose models lacked this domain expertise, AT&T launched the Open Telco initiative. The initial release, OTel 1.0, proved that a smaller, heavily domain-tuned model could outperform massive closed-weight architectures on telecom-specific benchmarks. However, scaling up to OTel2.0 meant digesting vastly larger datasets: raw network engineering logs, technical specifications, operational runbooks, and massive volumes of legacy GSMA standards documentation.
Phase 2: Resolving the Compute Bottleneck via Managed Infrastructure
As the data preparation footprint expanded, AT&T encountered an infrastructure roadblock common to enterprise AI deployments: provisioning and operational overhead. Historically, internal data science teams spent significant time managing bare-metal GPU clusters, handling driver compatibility, configuring orchestration layers, and managing static workloads.
To break this bottleneck, AT&T integrated Microsoft Foundry Managed Compute. The platform provided AT&T with abstracted, fully managed access to dedicated GPU infrastructure. This eliminated compute provisioning lag, enabling data science teams to shift their focus entirely from infrastructure orchestration to data engineering and model training.
Traditional Private Deployments:
[ Provision Hardware ] --> [ Config Drivers/Cluster ] --> [ Deploy Workloads ] (Weeks)
Foundry Managed Compute:
[ Request Managed Capacity ] ---------------------------> [ Scale & Train Models ] (Hours/Days)Phase 3: Multi-Model Synthetic Pipeline Orchestration
To construct the 1-trillion-token corpus needed for OTel2.0, AT&T deployed a multi-model data preparation workflow. Raw technical documentation, while dense, was insufficient on its own to train complex reasoning logic. AT&T used open-weights models to generate massive volumes of high-quality synthetic telecom data, augmenting real-world GSMA specifications.
- Synthetic Generation & Scrubbing: Microsoft’s small yet remarkably capable Phi-4 model was designated as the primary data generation workhorse. Phi-4 processed roughly 700 billion tokens per month, systematically translating raw telecommunications standards into structured, synthetic Q&A pairs, instruction-following tasks, and domain-specific code snippets.
- High-Reasoning Validation: Complex logic pipelines and cross-domain edge cases were routed through OSS-120B, a high-parameter open model capable of rigorous technical reasoning and structural data validation.
- Targeted Fine-Tuning Workflows: Models like Google’s Gemma 4 were integrated into iterative development loops to refine output quality and streamline training downstream.
Phase 4: Heterogeneous GPU Execution
With the training corpus assembled, AT&T deployed a compute fabric comprising approximately 530 high-performance GPUs via Microsoft Foundry Managed Compute. A core component of this footprint was 430 AMD Instinct™ MI300X GPUs, selected for their high memory bandwidth and capacity, which are ideal for large-scale token processing. By orchestrating both AMD and NVIDIA hardware within a unified platform, AT&T demonstrated that enterprise-scale AI fine-tuning can thrive on a heterogeneous, vendor-agnostic infrastructure layer.
Supporting Context & Key Metrics
The development of OTel2.0 offers clear economic and technical lessons for enterprise AI deployment. As model sizes scale, relying solely on commercial, pay-per-token API endpoints becomes financially unsustainable for massive token pipelines.
Economic Comparison: Proprietary API vs. Open-Weights Managed Compute
When generating and processing 1 trillion tokens, closed-source API pricing models (often ranging from $2.50 to $10.00 per million tokens processed through high-tier frontier models) result in astronomical costs. By hosting open models like Phi-4 on dedicated, managed GPU compute, AT&T shifted its economic model from variable, high-margin API consumption to fixed, fully utilized hardware cycles.
This strategic shift saved AT&T tens of millions of dollars in raw compute costs during the synthetic generation and data preparation phases alone.
+-----------------------------------------------------------------------------------+
| AT&T OTel2.0 SCALE & METRICS SUMMARY |
+------------------------------------+----------------------------------------------+
| Metric Category | Operational Value |
+------------------------------------+----------------------------------------------+
| OTel 1.0 Legacy Adoption | > 25 Million Global Downloads |
| Total GPUs Provisioned | ~ 530 Units (Unified Fabric) |
| Primary Accelerator Architecture | 430 AMD Instinct™ MI300X GPUs |
| Total Processed Dataset Size | ~ 1 Trillion Tokens (Raw + Synthetic) |
| Total Training Corpus Size | ~ 400 Billion Tokens (Refined Telco Context) |
| Monthly Phi-4 Token Throughput | ~ 700 Billion Tokens / Month |
| Open Models Deployed | Phi-4, OSS 120B, Gemma 4 |
| Infrastructure Provisioning Speed | Reduced from Weeks to Days/Hours |
+------------------------------------+----------------------------------------------+Multi-Model Workload Allocation
The strategy relied on matching specific open-source architectures to the distinct functional requirements of the pipeline:
- Phi-4: Primary data preparation engine. Handled bulk dataset transformations, document structuring, and synthetic token generation at massive scale (~700B tokens/month).
- OSS-120B: Heavy-reasoning engine. Used to resolve complex network topology puzzles, validate standard compliance, and generate complex operational logic.
- Gemma 4: Optimization engine. Supported specialized iterative fine-tuning and development workflows prior to final target model weights compilation.
Official Statements & Executive Insights
The success of the OTel2.0 project underscores a growing consensus among technology leaders: open-source model weights combined with flexible, managed compute platforms represent the most viable path toward enterprise AI maturity.
Jeff Boudier, Vice President of Product at Hugging Face, emphasized the industry-wide significance of AT&T’s commitment to open telecom standards:
"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."
From an operational standpoint, managing raw compute platforms at trillion-token volumes transforms hardware from a back-end utility into a core strategic variable. Mark Austin, Vice President of Data Science and AI at AT&T, highlighted how Microsoft Foundry Managed Compute removed this operational friction:
"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."
Future Outlook & Strategic Implications
The deployment of OTel2.0 marks a major milestone in telecommunications infrastructure and enterprise AI architecture. By making the OTel2.0 foundation available to the broader ecosystem, AT&T and its industry partners—including the GSMA—are establishing a standard blueprint for telco-grade artificial intelligence.
[ Domain Data + GSMA Standards ]
|
v
[ Open Models (Phi-4, OSS 120B) ]
|
v
[ Microsoft Foundry Managed Compute ] <---> [ AMD MI300X / NVIDIA Cluster ]
|
v
[ OTel 2.0 Open Foundation ]
|
v
[ Global Telecom Open Ecosystem ]Implications for the Global Telecom Ecosystem
- Accelerated Network Automation: Open, highly capable telco models allow regional operators to deploy intelligent automated agents for network diagnostic routines, fault isolation, and dynamic capacity routing without incurring prohibitive model license fees.
- Standardization Around GSMA Guidelines: With OTel2.0 pre-trained on standardized GSMA framework documentation, global carrier integrations, cross-border roaming logic, and inter-provider billing reconciliations can be standardized across automated channels.
- Decoupling from Single-Vendor Silicon: AT&T’s successful deployment across 430 AMD Instinct MI300X GPUs alongside existing NVIDIA hardware demonstrates that enterprises can avoid single-vendor GPU lock-in.
The Enterprise AI Paradigm Shift
Beyond telecommunications, OTel2.0 serves as a practical model for other heavily regulated, highly specialized industries—such as healthcare, financial services, and energy infrastructure. The project proves that processing trillions of domain-specific tokens does not require proprietary, closed-system monopolies.
By combining open-source model architectures (such as Phi-4 and Gemma), unified management platforms (like Microsoft Foundry), and diverse silicon environments (AMD and NVIDIA), modern enterprises can construct secure, domain-specific AI platforms that scale efficiently while maintaining cost control.
