- GSMA warns telcos that today’s frontier AI models still fall short on telecom-specific tasks
- Open telco AI models could be critical to autonomous networks
- AT&T, SoftBank, China Telecom and others are already using GSMA’s Open Telco AI work to fine-tune models, benchmark performance and build in-house AI muscle
For operators betting AI will help automate their networks, GSMA’s Director of AI Technologies Louis Powell has a warning: the biggest models on the market still don’t really understand telecom.
That’s something the GSMA – in collaboration with operator members like AT&T – is trying to change with its push toward open telco models. And its efforts could prove critical both for the future of autonomous networks and network security.
The GSMA has spent the past two years benchmarking AI models against telecom-specific tasks, and the results show a stubborn gap. Frontier models may be improving quickly overall, but Powell said its research has found they’re not getting much better at parsing 3GPP standards, troubleshooting RAN issues or handling the domain-specific complexity operators need for autonomous networks.
Mind the AI model gap
This gap has real implications for telcos and has shaped how they’ve gone about chasing the AI opportunity. Powell noted that while deployments in customer service, marketing and backend services have taken off “only about 16% of deployments are on the network.”
“We saw a gap between what these models were capable of in general areas and what was happening in the telecoms area, and we were very concerned that the industry wouldn't make the most out of this opportunity,” Powell said. He added that GSMA was also “slightly concerned that the supply chain of AI is very, very narrow, and we don’t’ want everybody outsourcing everything to the hyperscalers.”
According to Powell, closing the model gap is critical for operators looking to achieve Level 4 and Level 5 autonomous networks. “You're not going to get there…if the models fundamentally aren't accurate enough or efficient enough to get there,” he pointed out.
Powell added that open options will also allow operators to tackle emerging security and sovereignty challenges in a way they just can’t with closed models. He noted a recent incident involving a rogue OpenAI model was solved by open models from Hugging Face, and also highlighted the challenges presented by recent efforts to curb access to certain proprietary models. Both scenarios represent real risks for operators.
Open models, he explained, can help prevent security issues like the one OpenAI faced because they allow operators to program in their own guardrails in a way that’s not possible with closed models. And owning a fine-tuned or post-trained model makes it less likely that access will suddenly be pulled as it might be with a frontier model.
Interest in open models among telcos has been steadily building in recent years. A recent Nvidia survey found 89% of telcos said open-source models and software are important to their AI strategy. For comparison, just 40% of respondents in the 2025 version of the same survey said they planned to use open-source tools. The percentage of operators who said they plan to develop AI solutions in-house also rose from 37% in 2025 to 42% in 2026, with another 38% stating they planned to be directly involved in codeveloping solutions with a partner.
Building AI muscle
To help operators get started down the open AI path, GSMA created the Open Telco AI initiative. Launched at MWC Barcelona earlier this year, the program provides operators with benchmarks to test how models perform on real telecom tasks, data sets and purpose-built models, and can even help them secure GPU capacity for model training and fine tuning. The idea is to give operators more control over their AI destiny.
AT&T has very publicly seized on this program to build its open source OTel 1.0 and 2.0 models, using GSMA data in the latter to ensure it deeply understands telecom standards. But Powell said other operators like SoftBank and China Telecom are also using its benchmarks to evaluate proprietary models they’ve built on open-source foundations.
But it’s one thing to ensure a model speaks telecom – for example, getting it to understand that a “cell” refers to a geographic area covered by a base station and not a human cell. It’s something else entirely to actually apply it to solving network problems.
On that front, Powell said Huawei has created an open data set called TeleLogs, which is designed to help train models for root cause analysis. “That took a lot of work,” Powell said, noting that compiling data about incident details, KPIs, resolution steps and ground truths across vendors “is not easy.”
All of these efforts are designed to help operators build what Powell described as the AI muscle needed to allow them to control their own fate.
“Data availability is hard. And AI expertise – [there are] very few AI researchers working out of telecoms businesses outside of the U.S. and China. So, if you’re in APAC, in parts of Europe, in Africa, you have very little knowledge of how to build these models,” he said.
GSMA is acting as a sort of an AI personal trainer for operators, helping them build muscle and get good at the (somewhat boring) foundational skills behind model building before they go off and try to create sexy AI products for the commercial market.
“I think it’s very hard for telcos who want to monetize AI if they can’t do AI themselves,” he said. “This is a great way to do it: get good at building models, solutions, tooling, agents, and then they can look to see about what they can commercialize to outside parties.”
Read more about telco AI and open models:
Open models are driving AT&T’s AI ‘tokenomics’ strategy
GSMA seeks to tailor AI models for telco requirements
Together AI expects IBM-backed AI capacity to sell out before launch
Open weight AI vs open-source AI: What’s the difference?
Telcos are increasingly going DIY when it comes to AI