AT&T targets $4 billion in savings through AI-driven tokenomics strategy

AT&T’s pivot from experimentation to a structured 'tokenomics' architecture addresses the primary barrier to telco AI: high inference costs. By decoupling data from underlying legacy hardware and utilizing edge-based processing, the operator is transforming its network from a passive pipe into a programmable engine capable of self-optimization. For the broader ecosystem, this sets a benchmark for how 5G Standalone infrastructure can be monetized through operational efficiency rather than just consumer data plans. Strategists should track AT&T’s transition from general-purpose LLMs to specialized internal SLMs as a signal for the next phase of enterprise AI efficiency.

AT&T's tokenomics strategy has evolved significantly since the initial announcement, with the operator now processing an average of 45 billion tokens per day, up from 27 billion. Mark Austin, VP in AT&T's Data Office, told Fierce Network that the operator expects to route 25% to 40% of its AI tokens through open models this year, targeting 65% to 70% ultimately, while keeping overall AI costs relatively flat despite the fivefold increase in daily token volume. The company has built a cache-aware router within a LiteLLM gateway that scores thousands of benchmark tasks to decide when to escalate from open models to premium options, factoring in pricing across cloud environments and cache retention costs.

On the open-model front, AT&T released OTel 2.0 in July 2026, a fine-tuned iteration of Google's Gemma 4 trained on over 400 billion tokens using AMD GPUs through Microsoft Foundry's Managed Compute service. AT&T is one of the first companies to use Managed Compute at scale through Microsoft Foundry, processing over 1 trillion tokens for training, with GSMA providing 15 billion raw tokens of telco-specific data to ground the model in industry standards. OTel 1.0, released in March, has been downloaded approximately 28 million times, while OTel 2.0 has reached around 5 million downloads, signaling broad industry interest in telco-specialized open models.

The operational mechanics behind AT&T's cost reduction rely on edge telemetry tokenization and intelligent model routing rather than a single architectural change. AT&T's approach converts verbose network logs into standardized integer keys at the edge before forwarding them to a message broker, achieving an 80-85% payload reduction that lowers cloud storage and compute expense. For BSS workloads, cryptographically signed security tokens carry identity and billing context from mobile apps to live agents without additional database checks, cutting 52 seconds from average handle time and saving approximately $4.3 million annually in a large call center. The operator is also scanning its roughly 150 production AI use cases to identify those older than one year that can be shifted from closed-source to open models, a pool representing about 25% of current token volume.