Overview
- AT&T disclosed Thursday that it has started redirecting substantial employee and network AI queries away from proprietary models toward open-source and open-weight alternatives.
- The carrier said it processes about 45 billion AI tokens a day and currently runs roughly a quarter to 40% of queries on open models with a target to reach about 60%–80% over time.
- AT&T built a cache-aware AI gateway and smart routers that pick the cheapest adequate model for each query, and it launched an in-house open model called OTel 2.0 trained on more than 400 billion tokens using AMD GPUs and Microsoft Foundry.
- The company claims cost cuts of 80%–90% in targeted applications, while independent reporting finds savings up to about 56% on coding and advanced tasks when using model routing with a modest drop in quality.
- AT&T says running open models on its own systems keeps sensitive network and customer data in-house, and the firm is avoiding some Chinese open-weight providers while assessing security and compliance risks as the industry rethinks cloud and chip demand.