AT&T charts a new course with OTel 2.0, its new open-source AI foundation model

AT&T charts a new course with OTel 2.0, its new open-source AI foundation model

The telecommunications sector has never been immune to a hype cycle. As the dust settles around Generative AI, a new phase is emerging for telcos as they look to move past initial experimentation and navigate their way through the fog of AI. Telcos at the heart of this shift are grappling with a key question – do they deploy AI over open-source models or run over frontier models from the likes of Anthropic and OpenAI. This question is set against the backdrop of a much broader policy debate on open vs closed, but telcos also have to grapple with a cold, harsh economic reality that the decisions they take on how to run AI can have prohibitive financial implications.   

With some of the most complex datasets across the business world, telecom operators are increasingly looking towards open-source, highly optimized and domain-specific models to serve their highly specialized and demanding requirements as opposed to general-purpose frontier models. Indeed, the GSMA’s Open Telco AI leaderboard shows the top three performers to be “domain-adapted models that are highly accurate and can be significantly smaller in size.”  

Beyond the existential desire to retain control of the stack and underlying data, there is increasing recognition that general-purpose frontier models do not have the domain-specific knowledge that is crucial for telecoms domains. While this is theoretically possible to do, telcos are also wary of the prohibitive cost of relying on general-purpose frontier models for aiding core network functions.  Instead, telcos are increasingly creating or deploying domain-specific models. 

An example of this shift comes from the recent announcement by AT&T, of the launch of OTel 2.0, a newly trained telco-specific foundation model. While AT&T is certainly not the first operator to launch its own foundation model, they are taking an interesting approach built on a strategy that utilizes a multi-model approach built on hybrid infrastructure from multiple hyperscaler and GPU architectures and is optimized for lowest cost token consumption for AI inferencing. 

Why has AT&T taken chosen this path? 

The question merits a closer look, especially when other telcos have chosen to work with frontier models from the likes of OpenAI and Anthropic for use in telco environments. Why start from scratch? The immediate takeaway is that when an operator at AT&T's scale processes 45 billion tokens daily, routing every query through a massive, generalized frontier model can be financially ruinous. To avoid this, AT&T has calculated that they are better off with a customizable, open-source model which will lower their token consumption bills and also produce more meaningful and telco-specific outcomes. Of course, not all operators are at the scale of an AT&T but the financial impact in proportional terms is still valid.  

A New Telco “Tokenomics” Playbook 

Telcos have key requirements for AI, cutting across inference costs, domain accuracy and silicon diversity. The best-known frontier models are best known for incredible reasoning and language interpretation tasks. However, feedback from operators suggests that frontier models still lack the foundational context for highly demanding and specific industry verticals like telecoms, where the model will need to parse multi-vendor network topologies, interpret proprietary telemetry data, or reliably diagnose core bottlenecks. As such, they are not the best fit for the highly regulated, highly vertical, data intensive telecoms domain. There is very little to no room for hallucinations as precision is non-negotiable. Not only do all outputs need to be highly accurate, they also need to be “affordable” in the sense that the process doesn’t consume a prohibitively large number of non-productive tokens.  

AT&T’s approach represents a structural overhaul of using AI within the telco domain. The playbook has the following highlights:

• Multi open-model usage: rather than relying on a single behemoth model, AT&T has deployed a cache-aware router that evaluates the speed, cost, and required quality of a prompt, dynamically routing it to the most cost-effective model for that specific task. It can even switch models mid-session during multi-turn conversations. AT&T claims that AI inference costs have dropped by up to 90%. This data point validates GSMA Intelligence’s recommendation in a recent Spotlight report “Telcos and the AI token economy: what could go wrong?”, where we highlighted the need for operators to route end-user token requests to the most efficient models with a view to managing overall cost efficiency. 
• Domain-Specific Open Models: In collaboration with the GSMA, AT&T has supported the Open Telco AI initiative, releasing the OTel 2.0 model with more to come. These models are trained strictly on curated, telco-specific datasets, with the initial dataset of approximately 15 billion tokens provided by the GSMA (sourced from standards bodies like the 3GPP, ETSI, CAMARA, ITU, O-RAN and TMForum, alongside a 10 billion token dataset from the “Telco Corpus” released by the GSMA and Pleias. This data was then combined with additional data from AT&T (in collaboration with RedHat, Dell, Microsoft Azure and AMD) to build the 400-billion token training set at the core of the OTel 2.0 model release. .
• Different models for different workflows - AT&T employed a diverse mix of open-source models rather than a monolithic approach for the initial training. For example, Microsoft's Phi-4 handled heavy data preparation and synthetic data generation, processing 700 billion tokens. For heavier reasoning workloads, they tapped OSS-120B. Post-training was done using Google Cloud's Gemma-4 (specifically the gemma-4-31B-it variant), achieving a 91.74% on the Open Telco AI Leaderboard

Sovereign AI is a real opportunity, and Telcos need not be the gatekeeper

AT&T’s approach is instructive for operators looking to service the broader Sovereign AI opportunity. A playbook created on open-source models and optimized for cost, wrapped with governance and security frameworks, can help operators retain strict control over proprietary data flows, protect enterprise IP, and build a pathway to truly Sovereign AI.

Telcos are positioning themselves as local champions for mission-critical alternatives to the leading frontier models. But to succeed, telcos do not need to compete directly with the frontier models. Rather, their focus should be on building affordable, trusted and secure networks that can be applied to the high-specific and sometimes even niche requirements of the business verticals in their home markets. Smaller sized but effective domain-specific models can do the trick. This will require a hybrid and highly flexible approach that balances inference costs, domain knowledge and accuracy and silicon diversity. 

Read more from the related report, Telcos and the AI token economy: what could go wrong?”.