The telecommunications sector is experiencing a highly focused technological shift regarding artificial intelligence deployment. AT&T has officially launched a new AT&T Open Telecom AI model called OTel 2.0. Developed in collaboration with the GSMA, Microsoft, AMD, Dell, and Red Hat, this system is trained on 400 billion tokens specific to the telecommunications industry.
Alongside the model, the company debuted an AI Gateway capable of processing 45 billion tokens per day. This gateway utilizes cache-aware routing to decrease AI inference costs by up to 90%, reflecting a broader enterprise transition toward specialized models. The announcement arrived as the wider technology industry processed the implications of an agentic AI-based security breach at the platform company Hugging Face, highlighting that deploying AI at scale carries operational risks alongside efficiency gains.
The Architecture Behind OTel 2.0
At its technical core, OTel 2.0 is built upon Gemma 4 31B-IT, an open multimodal model created by Google DeepMind that handles text, image inputs, and video processed as frame sequences. AT&T trained this base model using 400 billion tokens explicitly selected from a wider pool of more than 1 trillion processed tokens.
This meticulous data curation ensures the system operates as a genuine telecom-specific AI, rather than a standard general-purpose system with a telecom label attached. According to the GSMA, the resulting system currently sits at the top of the Open Telco AI leaderboard. Notably, the top three performers on this industry benchmark are all domain-specific models, with no general-purpose models occupying the leading positions.
Addressing AI Inference Costs Through Intelligent Routing
AT&T’s Chief Data and AI Officer, Andy Markus, noted that many organizations default to utilizing the most powerful frontier models available, regardless of whether a specific task actually requires that level of sophistication. He argues this habit is largely unnecessary and highly expensive.
To resolve this, AT&T paired its AT&T Open Telecom AI with an intelligent routing system. Managing an average of 45 billion tokens daily is a current production reality for the company, and at that volume, model selection has direct financial consequences. Rather than sending every prompt to an expensive frontier model, the AI Gateway uses cache-aware routing to direct each task to the most cost-effective model capable of meeting required quality standards.
By weighing speed, cost, and expected output quality at each step, the gateway intelligently matches tasks to appropriate models. Markus confirmed that this routing process is already cutting AI inference costs by up to 90 percent, saving millions while keeping output quality completely intact.
The Shift From General-Purpose to Domain-Specific Models
The broader argument forming within enterprise technology is that domain-specific models will outperform general-purpose ones in specialist industries at a fraction of the cost. The GSMA’s stance on this subject is unambiguous, asserting that general-purpose models simply lack the necessary training data for network operations.
Why Telecom-Specific AI Outperforms Broad Models
In a publication accompanying the launch, the GSMA explained that while general-purpose models are advanced, their training sets barely cover specialized telecom standards, operating environments, or protocols. Consequently, when operators attempt to troubleshoot a live network fault or interpret an industry standard, the general-purpose models falter because they lack foundational exposure to the domain.
The trade association drew a direct comparison to manufacturing, healthcare, and financial services, which are already adopting domain-specific models. Telecommunications requires systems trained on its unique operational data and network configurations to support key use cases, such as product development, network configuration, and troubleshooting. Familiarity with specific terminology makes a measurable difference to output quality in these areas.
Open Source Strategies and Operational Flexibility
The launch of OTel 2.0 also emphasizes the importance of open-source frameworks. By building on an open base rather than a proprietary closed system, the AT&T Open Telecom AI framework avoids vendor lock-in. The foundational Gemma 4 31B-IT model allows the telecom training layer to be audited, shared, and adapted by other operators.
This flexibility is vital for enterprise operators who need to control model deployment across changing cloud and on-premise environments. For large operators facing massive token volumes, deploying a telecom-specific AI on efficient routing infrastructure handles the bulk of automated workloads while keeping AI inference costs manageable. Because these specialized systems can be significantly smaller than broad frontier models, they cost less to host, run, and fine-tune, fundamentally changing the economics of automated network operations.