AT&T Says It Has Processed Over 1 Trillion Tokens Training New Telecom AI Models With AMD
AT&T reports processing more than 1 trillion tokens on AMD-powered Azure compute while training new open telecom AI models using data from GSMA.
What AMD and AT&T announced
At AMD's Advancing AI 2026 conference, AMD, AT&T and Microsoft gave an update on the Open Telco AI initiative, an effort the companies launched with GSMA and other partners at MWC 2026. AT&T said it has now processed more than 1 trillion tokens through AI models hosted on Microsoft Azure and running on AMD Instinct GPUs. The company is also training additional open, telecom-specific models on AMD hardware, using managed Azure compute and data supplied by GSMA. AMD framed the update as evidence that the initiative is moving from an industry framework into practical, at-scale deployment.
Why AT&T wants telecom-specific models
According to AMD, general-purpose AI models can struggle with network data and telecom standards. Models trained on domain-specific data are meant to give telecom operators options that better match their needs for accuracy, performance and cost when interpreting network data or understanding telecom standards. AMD Newsroom said AT&T contributed models and telecom expertise, GSMA contributed data and benchmarks, and Microsoft Azure supplied the managed compute behind the effort.
The new OTel 2.0 model
As part of this update, AT&T said it selected 400 billion of the most useful tokens out of the 1 trillion it processed through Microsoft Foundry-managed compute, then used that smaller set to post-train a new model called OTel 2.0 on AMD hardware. AT&T describes OTel 2.0 as the largest and best-performing open-source model produced under the Open Telco AI initiative so far. Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD, said AT&T's progress shows what open collaboration, telecom expertise and the right compute can achieve together.
AT&T's reported efficiency numbers
AT&T also shared figures on how it runs AI day to day. The company says an intelligent, cache-aware gateway routes tasks to the right model and processes an average of 45 billion tokens daily. AT&T reports that this gateway approach has cut its AI costs by as much as 80%. These are AT&T's own reported figures, not independently audited results.
Key facts
- AT&T says it has processed more than 1 trillion tokens through Azure-hosted AI models running on AMD Instinct GPUs.
- AT&T selected 400 billion of those tokens to post-train the new OTel 2.0 model on AMD hardware.
- AT&T reports its AI gateway processes an average of 45 billion tokens every day.
- AT&T says the gateway has reduced its AI costs by as much as 80%.
- AMD, AT&T, GSMA and other industry partners launched the Open Telco AI initiative at MWC 2026.
What happens next
AMD said the work continues through further training, fine-tuning, evaluation, model publication and deployment testing. For AMD, the update is presented as a customer proof point for its open ROCm software platform and a compute lineup that spans data centers to edge devices.
Check your understanding
- About how many tokens does AT&T say its AI gateway processes on an average day?
- Whose data helped train the new OTel 2.0 model, alongside AT&T's own models and AMD's hardware?