AT&T is running 45 billion tokens per day through a cache-aware AI Gateway that evaluates cost, speed and expected quality at every turn. It routes work to the cheapest model that can handle it and can switch models mid-session. Result: AI cost reductions of as much as 90%, already saving millions.
That is the public part. The more important part is what AT&T does next: it is using its own operational traces to train a domain-specific model, OTel 2.0, on hardware it controls.
Two numbers that matter more than the headline
- Over 1 trillion tokens processed for training through Microsoft Foundry Managed Compute.
- Over 400 billion tokens used to post-train OTel 2.0 on AMD GPUs, using GSMA data.
AT&T is one of the first companies to use Managed Compute at scale. The model is not rented. It is an open telco model trained on sector-specific data and run on AMD hardware that can flex between cloud and on-premises.
What this means for TokenOps
Traditional FinOps asks: what did we consume and what did it cost?
TokenOps has to answer a harder set of questions:
- What workload is consuming the intelligence?
- What business outcome is it producing?
- What is the lowest-cost execution path that still meets the quality, latency and risk threshold?
The router answers the third question at runtime. TokenOps answers the first two and closes the loop: normalize the consumption, attribute it to the operation and outcome, and use those economics to govern the next unit of AI capital.
Own the traces, not just the route
Routing is table stakes. The deeper move is what the operation does with the data it produces.
AT&T is converting live operational experience into an owned asset: traces, routing policy, a post-trained open model, and the learned execution path. That is the pattern we described in Owners, not renters — rent intelligence, but own sector-specific data and software assets.
The gateway does not just save tokens; it produces a curated stream of real tasks, outcomes and quality signals. That stream becomes training fuel for a model that speaks the company's own language and runs on hardware it controls.
This is where AI at build time, certainty at run time and Autonomous RunTimes (ART) come in. The economics have to be governed at runtime, and the runtime has to enforce the boundaries between rented frontier inference and owned, specialized execution.
The lesson for other enterprises
If you are routing to save tokens, you should also be learning from those tokens.
The ingredients are not exotic: operational data, execution history, and a clear idea of which outcomes matter. What AT&T adds is the discipline to treat traces as a capital asset and to build the machinery — gateway, training pipeline, hardware choice, economic ledger — around them.
Source and interpretation. AT&T, “The Tokenomics Equation: Balancing Cost and Performance,” July 23, 2026. AT&T reports the 45 billion tokens per day volume, 90% cost reduction, OTel 2.0, over 1 trillion tokens processed for training, and over 400 billion tokens used for post-training on AMD GPUs. This article is Proverify.ai's interpretation of their implications for enterprise TokenOps; AT&T does not endorse Proverify.ai or TokenOps.