Industry trade group – and Mobile World Congress organizer – GSMA launched an initiative targeted at accelerating the development and adoption of “telco-grade AI.”

The Open Telco AI program was initiated with a portal tapped with the grand scheme of allowing telecom “open models, data, compute, and tools to accelerate the development and evaluation of telco-focused AI models.” This is in contrast to what GSMA termed “frontier AI models” that “have advanced rapidly” but “continue to underperform on telecom-specific tasks.”

“Many general-purpose models struggle to interpret network data, understand standards documentation, or automate network operations with sufficient accuracy,” GSMA claims, noting this gap has limited telecom-based generative AI deployments in networks.

“Open Telco AI meets this challenge by uniting industry and academic partners to build the foundations of telco‑grade AI models, data, compute, benchmarks, and community,” GSMA proclaims. “Progress is tracked through the Telco Capability Index, which measures model performance across an expanding set of telecom‑specific tasks.”

Highlighted founding members included AMD and AT&T, with both cited for “making significant contributions” to the project.

AMD is providing its GPU compute capacity for model training, fine-tuning, inference, and evaluation, work to be done on AMD GPU platforms, cloud partner TensorWave, and open toolchains.

AT&T, for its part, is providing “open telco-models” developed and trained on “open, publicly available data” that is hardware and cloud agnostic. Andy Markus, chief data and AI officer at AT&T, added in a blog post that this work can build “confidence in more AI deployments in networks, and a more stable, reliable connectivity experience across the world.”

This mission was echoed by AT&T CEO John Stankey, who during an interview at this week's Mobile World Congress event AI was front and center in how the carrier was running and securing its network to provide both efficiency gains and a competitive edge.

“We're all running our call centers more efficiently using AI, getting better responses from our customers because of the choices we give them, and we're all writing more code faster," Stankey explained. "In our case, we're approaching nearly a 40% improvement in efficiency and effectiveness, and that not only is effective and efficient, but allows us to do more, bring more projects online than we've ever been able to do before that add value in the business."

While efficiency is a core tenet of adopting and implementing AI, Stankey said AT&T’s competitors were doing the same to the extent that “realistically, those efficiencies generally get competed away in the market,” and instead its efficiency gains were being used to take those savings and offer customers more value.

“That's the next step," Stankey said. "And so, using our data, our insights, what we know that somebody else doesn't know, I think that's the next great frontier. … You're starting to see really green shoots in areas like how we dynamically manage the wireless network to respond to loads and traffic, and what we're able to do to dynamically allow AI to realign antennas, move capacity in different areas.

Other firms touted as program contributors include KDDI, KPN, SK Telecom, SoftBank, and Swisscom, with “valued participants partners” including China Telecom, China Unicom, China Mobile, Deutsche Telekom, Google Cloud, IBM, Liberty Global, Telefónica, and Vodafone.

Building blocks

GSMA noted a handful of specific “building blocks” targeted for “co-creation” in the portal.

These include “telco models” that are open weight models designed to telecom tasks like network troubleshooting and standards interpretation. These models include work from AT&T; the Khalifa University-developed RFGPT radio-frequency language model; and a “large telco model” (LTM) from ADaptKey AI and built on Nvidia’s Nemotron AI models.

There is also an “open data” model that is a library of knowledge graphs, embeddings, and fine-tuning datasets of text, logs, and curated standards material from industry trade group GSMA, Huawei Technologies France, Khalifa University, Mantis NLP, NetoAI, Pleias, Purdue University, The University of Texas at Dallas, University of Leeds, Yale University, and Nvidia-sourced pipelines for generating synthetic data.

Other building blocks are for access to AMD and TensorWave-sourced compute and open toolchain for project training and inferencing open models; benchmarks for assessing model performance across seven telecom-specific metrics; and community resources for collaboration and engagement.

“And it’s not just about models. Open Telco AI will be a common space for operators, vendors, researchers, and developers to compare results, share learnings, and accelerate iteration that reflects how real-life networks operate,” AT&T’s Markus added about the group. “I encourage all companies with a stake in this effort to get involved. Participate in discussions, share what you’re learning and help guide the industry toward real AI value at scale. We’ll be participating and we hope to see you there.”

These efforts come as a number of similar AI-focused initiatives have recently hit the market, including an Nvidia-led effort just this week.

Analysts have pointed to the benefits AI can bring to the deployment and management of cloud-based open radio access networks (RAN), which are more complex orchestration challenges due to the disaggregated multivendor ecosystem. The use of AI could help close performance gaps for open RAN architectures compared with legacy RAN models.

Analysts are also predicting AI-RAN spend could top $10 billion by the end of the decade.