AT&T has moved OTel 2.0 into full production, marking a major telecom AI milestone. The open-source model comes from the GSMA-led Open Telco AI initiative. It targets real operator workloads, not general consumer chat tasks.
The model uses Google Gemma 4 31B-IT as its base. It has 31 billion parameters, which define how AI learns patterns. That size stays far below frontier AI models. However, this choice helps operators run it in their own data centers.
Training required significant data preparation. AT&T and partners processed more than one trillion tokens. Tokens are small text pieces used during AI training. They reduced that pool to about 400 billion telecom-focused tokens.
The dataset includes standards and engineering material from major telecom bodies. Sources include 3GPP, ETSI, ITU, O-RAN, CAMARA, and TM Forum. This focus gives the model stronger telecom knowledge. It also reduces exposure to personal data sources.
Hardware choices make the deployment especially notable. AMD Instinct GPUs power much of the training and inference. The project uses AMD’s open ROCm software stack. This avoids dependence on NVIDIA and CUDA, which dominate many AI deployments.
Meanwhile, Dell supplies carrier-grade servers for on-premises deployments. These systems use AMD MI355X GPUs in local data centers. That setup supports sensitive network tasks requiring local control. It also gives operators a clearer production blueprint.
Large-scale training ran through Microsoft Azure using Microsoft Foundry. Around 430 AMD MI300X GPUs supported cloud training. This shows that major AI workloads can run outside CUDA-based environments.
Still, vendor dependence does not disappear completely. It may shift toward specific cloud platforms or open-source stacks. Operators must assess support, integration, and long-term maintenance before adoption.
AT&T also uses an “AI Gateway” to control model usage. The gateway caches answers and routes tasks between AI models. Simple jobs can stay with OTel 2.0. More complex tasks can move to larger frontier models.
This design can lower operating costs. AT&T says cached routing cuts inference costs by up to 90%. However, smaller operators may not match those savings. They may lack similar traffic volume and engineering resources.
The model already supports practical telecom workflows. It can summarize dense standards documents. It can generate compliant network configurations. It can also assist troubleshooting, runbook creation, and knowledge retrieval.
Security remains a serious consideration. Open AI models can introduce new attack paths. Prompt injection and model poisoning need careful testing. Telecom networks demand strict validation before automation touches live systems.
To support review, the project publishes model weights and documentation on Hugging Face. Weekly updates are also planned. That makes OTel 2.0 a moving platform, not a static release. For Telco AI, this deployment signals a shift from trials to operational reality.
