AT&T shifts AI usage from 25% to a targeted 70%-80% open models over time

AT&T’s reported daily token usage, 45 billion, highlights the scale of its AI operations. At this volume, even small per-token savings translate into significant cost reductions. Open models, which can be self-hosted or fine-tuned, offer a way to avoid recurring licensing fees tied to proprietary APIs. The move also suggests AT&T is prioritizing cost predictability over the convenience of managed services.

The shift from 25% to 70%-80% open models implies a multi-year migration. Open models require infrastructure for deployment, monitoring, and updates, which adds operational overhead. AT&T’s telecom-grade reliability demands will test whether open models can match the uptime and performance of proprietary alternatives. The transition may also involve retraining or adapting existing workflows to new model architectures.

Data control appears to be a key driver. Open models allow AT&T to keep proprietary data within its own networks, reducing exposure to third-party providers. This is particularly relevant for telecom, where customer data and network telemetry are highly sensitive. However, open models introduce new security risks, such as model poisoning or unauthorized access, which AT&T will need to mitigate with internal safeguards.

The announcement reflects a broader trend among enterprises with large-scale AI deployments. While startups and smaller firms often rely on proprietary models for ease of use, companies like AT&T are betting on open models to avoid vendor lock-in and align AI costs with their existing infrastructure. The success of this strategy will depend on whether open models can deliver consistent performance at scale.