The telecommunications industry, the invisible backbone of the modern digital economy, has long struggled with a paradox: while it generates more data than almost any other sector, that data is highly specialized, proprietary, and technically dense. Generic large language models (LLMs), while impressive in their ability to write poetry or summarize news, often falter when tasked with the nuanced operations of a global cellular network.

To bridge this divide, AT&T has pioneered the Open Telco (OTel) initiative. With the launch of OTel2.0, the company is not merely iterating on software; it is fundamentally altering the economics and technical strategy of enterprise-grade AI development. By leveraging Microsoft Foundry Managed Compute, AT&T has successfully moved beyond the limitations of "black-box" proprietary models, creating a blueprint for how massive, industry-specific AI can be built, scaled, and optimized without ballooning costs or compromising performance.

The Core Challenge: Why Telecom Demands Custom AI

In the world of telecommunications, accuracy is non-negotiable. Network operations, regulatory compliance, and customer service require an AI that understands the specific standards of 5G, the complexities of fiber-optic infrastructure, and the rigid protocols of global data transmission.

When AT&T began the journey toward OTel2.0, they encountered a significant hurdle: the operational complexity of building domain-specific models at scale. Developing such a system requires processing trillions of tokens of highly specialized data. Historically, this would necessitate an arduous process of manual infrastructure management, where data science teams spent more time configuring hardware and managing deployments than actually refining their models.

The core mission was clear: AT&T needed to create an AI ecosystem that could handle massive datasets while remaining agile enough to iterate on new models as they emerged. This required a paradigm shift—moving away from a "one-size-fits-all" model approach toward a heterogeneous strategy that prioritizes open-source flexibility.

A Chronology of Innovation: From OTel 1.0 to 2.0

The development of the OTel family has been a multi-stage evolution, reflecting the rapid pace of AI advancement.

• The Foundation (OTel 1.0): The initial version of the model established the baseline for telecom-specific reasoning. Its success was marked by widespread adoption, with over 25 million downloads, proving that the industry was hungry for specialized, open-source AI tools.
• The Scaling Pivot: As AT&T looked toward OTel2.0, the requirements shifted. They needed to handle higher reasoning capabilities and significantly larger data volumes. This required an infrastructure capable of supporting a multi-model workflow.
• The Foundry Integration: By integrating Microsoft Foundry Managed Compute, AT&T streamlined its AI development pipeline. The transition allowed the team to move from lengthy, multi-week provisioning cycles to a streamlined deployment model that reduced the time-to-market for new model iterations to a matter of days.
• The Execution: Over the course of the project, the team processed roughly 1 trillion tokens, combining raw GSMA documentation with massive volumes of synthetic data. This culminated in the training of OTel2.0 on approximately 400 billion tokens.

Supporting Data: The Anatomy of OTel2.0

The technical requirements for training OTel2.0 were immense, requiring a massive compute footprint and a strategic mix of open-source models. The following table illustrates the strategic allocation of workloads:

The infrastructure supporting these models was equally robust. AT&T utilized approximately 530 GPUs through Microsoft Foundry Managed Compute. Notably, this included 430 AMD Instinct™ MI300X GPUs, representing a heterogeneous approach to hardware that provided the necessary flexibility to optimize for both cost and performance as the model’s needs evolved.

Official Perspectives: The Synergy of Open Source and Infrastructure

The success of the OTel2.0 project has drawn significant attention from industry leaders, emphasizing the growing importance of an "open" AI philosophy.

Jeff Boudier, Vice President of Product at Hugging Face, noted the broader industry implications of AT&T’s approach: "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 within AT&T, the focus remains on the strategic utility of the platform. Mark Austin, Vice President of Data Science and AI at AT&T, emphasized that the infrastructure itself became a key variable in the project’s success. "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."

The Economic and Strategic Implications

Perhaps the most significant takeaway from the OTel2.0 project is the economic model. In an era where "frontier models" (proprietary, closed-source models) often come with massive licensing fees and black-box limitations, AT&T’s reliance on open-source models like Phi-4 has resulted in substantial savings.

By generating synthetic data and preparing datasets in-house using open-source models, AT&T saved tens of millions of dollars compared to the projected costs of using proprietary frontier models. This is not merely a cost-cutting exercise; it is a strategic reinvestment. The capital saved was redirected toward more intensive experimentation, allowing for deeper research into network optimization and predictive maintenance.

Implications for the Broader Industry

1. Infrastructure as a Strategic Asset: The project demonstrates that for large-scale AI, the choice of infrastructure is not just an IT procurement issue—it is a core business strategy. The ability to switch between GPU architectures (AMD and NVIDIA) based on workload requirements provides a hedge against supply chain constraints and hardware-specific bottlenecks.
2. The Rise of Domain-Specific AI: The "generalist" model era is giving way to a more specialized future. Industries like finance, healthcare, and telecommunications are realizing that their competitive advantage lies in models trained on their proprietary data and domain standards.
3. Deployment Velocity: The transition from weeks to days for model deployment is the new benchmark for competitive AI development. Organizations that cannot rapidly iterate will find themselves trailing in the race to deploy production-ready AI.
4. Democratization through Open Source: By contributing OTel back to the community, AT&T is establishing a standard that others can build upon, effectively creating an industry-wide "common good" for telecom-grade AI. This lowers the barrier to entry for smaller telcos while keeping AT&T at the forefront of the ecosystem.

Conclusion: Setting the Stage for the Next Wave

The story of OTel2.0 is one of convergence: the convergence of open-source model availability, scalable managed cloud infrastructure, and the relentless need for domain-specific intelligence. As AT&T continues to deploy these models into production, the focus will shift from training to inference and real-time operational integration.

For the telecommunications industry, the message is clear: the future of network management is intelligent, automated, and—most importantly—open. By utilizing Microsoft Foundry Managed Compute to handle the "heavy lifting" of infrastructure, AT&T has proven that even the most massive, complex AI initiatives can be managed with efficiency and precision. As organizations worldwide look to replicate this success, the OTel2.0 model stands as a testament to the power of open-source collaboration and the strategic necessity of scalable, flexible AI infrastructure.

The era of experimentation is ending; the era of production-scale, domain-specific AI has officially begun.