New alliance including AT&T and AMD aims to fix the reliability gap in telecom AI
In sum – what we know:
- Open Telco AI – Global initiative announced at MWC 2026 to develop telco-grade artificial intelligence through open collaboration.
- Founding partners – AT&T providing open models and data; AMD supplying compute capacity and GPU platforms.
- The problem – Current generic AI models lack the precision and reliability required for highly regulated telecom network operations.
The GSMA wants to make AI a little more telco-focused. The association took the wraps off of the Open Telco AI at Mobile World Congress in Barcelona, with a pretty clear pitch. General-purpose AI models, however impressive they’ve become, aren’t quite getting the job done when it comes to telecom network operations — an environment where reliability, complexity, and precision are incredibly important. Right now, only 16% of telecom GenAI deployments touch network operations at all.
Louis Powell, Director of AI Initiatives at GSMA, has emphasized that current AI models do not yet “speak telco.” It’s an apt way to frame it. Telecom networks are some of the most demanding, heavily regulated environments out there, and bolting a general-purpose model trained on internet text onto real-time network troubleshooting or dense standards interpretation just doesn’t get very far. The GSMA isn’t trying to build proprietary tools here — instead, it’s positioning itself as a convener, pulling together operators, vendors, AI developers, and academics around shared models, data, compute resources, and benchmarks.
That said, telecom has seen its share of ambitious consortium efforts that never quite delivered.
“The telco world is notoriously slow, and such federations have been launched – and fallen apart – at previous occasions,” noted Johan Ottosson, VP Strategy at Arelion, in an interview with RCR Wireless. “So if the initiative can learn from its history as well as hyperscaler success, i.e. stay focused, start small and generate progress, and standardize from there, it’s got good chances of adding real value.”
Partner contributions and technical resources
The two founding partners each bring something the other doesn’t. AT&T is open-sourcing a family of telco models trained on publicly available data, built to be hardware- and cloud-agnostic. The goal is to show that AI can deliver meaningful results across different scales of compute — not just at the hyperscaler end of the spectrum. On top of the models themselves, AT&T is contributing datasets, evaluation frameworks, and real-world testing expertise.
AMD is handling the compute side, including cloud infrastructure through its partner TensorWave. That gives participants access to the training, fine-tuning, inference, and evaluation resources they need without getting locked into any single vendor’s stack.
The model ecosystem is already broader than you’d expect from a launch event. It includes RFGPT, a radio-frequency language model out of Khalifa University, along with AdaptKey AI’s Large Telco Model (LTM), which is built on Nvidia Nemotron. Open datasets aggregated from leading universities and technology partners fill out the rest of the technical foundation. Now, whether a consortium model can realistically keep pace with hyperscalers pouring massive R&D budgets into proprietary AI remains to be seen.
“For training purposes, the first question is about the size and quality of the dataset,” said Ottosson. “If a consortium can federate this, it solves a unique problem that a large budget just can’t solve with sheer money.” Put another way, the collective data advantage of a broad telecom coalition might be the one asset that no amount of spending can replicate.
Benchmarking and industry participation
Perhaps the most concrete piece of the initiative is the Telco Capability Index — a new framework for tracking how well models perform across an expanding set of telecom-specific tasks. There’s already a leaderboard evaluating performance on seven benchmarks, with plans to grow from there. This kind of structured evaluation is critical. Without clear, standardized ways to measure progress, it becomes nearly impossible to separate real improvement from noise.
Early community engagement looks promising. The “AI Telco Troubleshooting Challenge” has pulled in over 1,000 registrations, and the “Agentic Challenge” is pushing participants to tackle real-world telecom AI problems. These challenges generate practical solutions while simultaneously stress-testing the open models and tools under conditions that resemble actual operations.
The supporting partner list is extensive, spanning major operators and tech companies like Google Cloud, Deutsche Telekom, China Mobile, Vodafone, Telefónica, IBM, Liberty Global, and others. That kind of breadth adds credibility at launch, though the real signal will be sustained contribution over time.
There’s also a legitimate question about the speed at which these open models can graduate from lab settings into live, mission-critical network environments. Ottosson suggested a pragmatic way forward. “Maybe the first step doesn’t have to be autonomous, but to augment humans and improve time-to-response, or suggest better pre-emptive responses. This can go a long way, and the findings will help us to solve the next set of challenges.”
