- AT&T launches OTel 2.0, a more advanced family of open source telecom AI models trained on AMD hardware
- The carrier consumes more than 1 trillion tokens per month and runs over 100 generative AI models in production
- AT&T says data sovereignty means avoiding lock-in to any single chipset, AI model or development toolset
AMD ADVANCING AI 2026, SAN FRANCISCO — AT&T launched OTel 2.0, a new generation of its open source telecom AI models trained on AMD hardware. The carrier now consumes more than a trillion tokens per month across more than 100 generative AI models in production.
"We're excited to announce today the launch of OTel 2.0, which is a more advanced set of models that we have trained on AMD, and we're now making that available via open source across the broader industry," said AT&T CTO Jeremy Legg during an AMD Advancing AI 2026 keynote.
OTel — the Open Telco AI model family — is AT&T's answer to the problem that frontier models don't speak telco.
Telcos need models optimized to solve their unique operational and business problems. GSMA found only 16% of telecom generative AI deployments touch network operations, a gap the GSMA attributed to general models' inability to reliably parse network data, standards documents and vendor telemetry.
AT&T launched OTel 1.0 at Mobile World Congress in March as a founding contribution to the GSMA's Open Telco AI initiative. The effort produced 30 models across multiple sizes and architectures, including Google's open source Gemma, post-trained on a telecom dataset curated by GSMA, operators, vendors and universities. In under five months the models have logged more than 18 million downloads and sit atop the GSMA's Open Telco Benchmarks, outperforming far larger frontier models on telecom-specific tasks.
Running workloads on both AMD and Nvidia demonstrates independence. "We are big believers in data sovereignty, and data sovereignty for us means not being tied to a specific chipset, not being tied to a specific model or a specific set of development tools," Legg said. "You need as an enterprise to manage your data. Your data is your fuel."
Why independence matters
The practical payoff is bargaining power. An operator that can demonstrate production-grade results on multiple silicon platforms and multiple models, open and closed, is an operator that no supplier can hold captive on price. Legg said AT&T routes work across its portfolio of models to manage costs, and that discipline is holding token expenses in check even as consumption grows by double digits.
The scale of that consumption is considerable. Beyond its trillion-token monthly run rate, AT&T transcribes and mines more than 300,000 customer calls a day, applies AI to fraud detection and customer care, and uses models to determine optimal cell tower placement and RAN configurations. The company is now moving past discrete use cases to what Legg called agentifying entire workflows. AT&T is rebuilding end-to-end processes in HR and finance, including cash forecasting and employee onboarding, around AI agents.
That workflow framing is where Legg located the industry's most common failure mode.
"I'll start at a place that I don't think everybody starts from here, which is the human. You need world-class teams in order to do this, and you need people that think in workflows," he said. "There's an enormous amount of what I call racing from stoplight to stoplight. Folks are going from zero to 100 and then they hit the next stage of the workflow, and then there's a red light, and they haven't worked their way all the way through that."
Other telco AI activity
In addition to working with AMD, AT&T is developing edge AI infrastructure with Nvidia and Cisco. Internationally, SK Telecom, SoftBank, Orange, KDDI and Swisscom are contributing to the same GSMA effort behind OTel, while also pursuing their own telco-specific models. And Nvidia and Huawei are contributing to the GSMA initiative as well, with synthetic data pipelines and curated datasets, respectively.
AT&T has said its generative AI investments return roughly 2x in free cash flow, affecting about $1 billion annually — a counterpoint to research finding that most enterprises see no return on AI spending.
Legg also spoke during a segment of a keynote where AMD launched Helios, its first rack-scale AI system, built on open Ethernet networking and claiming 30% better token economics than Nvidia's competing rack. AMD also introduced the Instinct MI350P, an air-cooled GPU designed to run models up to 260 billion parameters inside existing enterprise data centers without facility upgrades — the class of hardware AT&T's on-premises open-model strategy is built to exploit.
Read more about the AMD Advancing AI 2026 conference:
Schneider Electric and AMD release Helios reference design for 246 kW AI data center racks