AT&T has shifted 40% of its internal AI workloads to open-weight models, aiming to reach 60–70% usage within the next year. This transition has reduced the company's AI coding expenses by 56%, with only a 2% decrease in output quality. For specific complex tasks, cost savings reach 80–90% compared to proprietary alternatives.
The company currently processes approximately 45 billion AI tokens daily, marking a 5.6-fold increase from the 8 billion tokens processed a year ago. To manage these requests, AT&T utilizes an intelligent routing system based on LiteLLM and a custom AI gateway, which directs simpler tasks to models like Nvidia Nemotron, Meta Llama, and Google Gemma.
Beyond using existing models, AT&T developed OTel 2.0, an open-weight model tailored for telecom data. Created in partnership with the GSMA Open Telco AI initiative, the model was trained on over 400 billion tokens using AMD GPUs and Microsoft Foundry. The company continues to evaluate various open-weight options while reconsidering its investments in OpenAI and Anthropic.
