AT&T routes 40% of AI workloads through open-weight models, cutting coding costs by 56%

AT&T is quietly pulling off one of the more aggressive enterprise AI pivots in recent memory. The telecom giant now routes roughly 40% of its internal AI requests through open-weight models, and it’s not slowing down. The company has set a target of 60-70% within the next year.

The math behind the decision is hard to argue with. AT&T has slashed AI coding costs by 56% while absorbing only a 2% decline in output quality. For certain complex workloads, the savings are even more dramatic, with cost reductions hitting 80-90% compared to closed-model alternatives.

The tokenomics of telecom-scale AI

AT&T’s AI consumption has grown at a pace that would make any CFO nervous. The company now processes approximately 45 billion tokens per day, up from around 8 billion just a year ago. That’s roughly a 5.6x increase in twelve months.

AT&T calls its approach “tokenomics,” and the logic is straightforward: not every AI request needs the most expensive model in the room. The company uses an intelligent routing system built on LiteLLM and a custom AI gateway to match tasks with the appropriate model. Simpler requests get directed to open-weight options like Nvidia Nemotron, Meta Llama, and Google Gemma. More complex tasks can still tap into closed models when the quality threshold demands it.

Open models closing the gap

AT&T executives have stated that open-weight models have narrowed the performance gap with proprietary alternatives, which has prompted the company to rethink its investments in offerings from Anthropic and OpenAI.

AT&T isn’t just consuming open models. It’s building them. The company launched OTel 2.0, a post-trained open-weight model specifically designed for telecom data. The model was developed in association with the GSMA’s Open Telco AI initiative, trained on over 400 billion tokens, and built using AMD GPUs and Microsoft Foundry.

The broader enterprise shift

AT&T isn’t operating in a vacuum. The move toward open-weight models reflects a growing trend across enterprise AI adoption, driven by three concerns: cost efficiency, customization, and data sovereignty.

The company is also keeping an eye on Chinese open-weight models, though it hasn’t adopted any. Security and compliance concerns have kept those options on the evaluation bench rather than in production.