Telcos step up to the AI model challenge – GSMA

• The GSMA’s Naresh Chouhan warns that frontier models still don’t perform well in telco environments
• But telcos such as AT&T and China Telecom are stepping up and making great strides with open-source models 
• More broadly, there’s still a disparity in telco AI spending trends, notes the GSMA man

When the GSMA launched Open Telco AI during MWC26 earlier this year, the pitch was simple: Frontier models are powerful but they don’t speak telco. Six months on, 18 operators have signed up to the industry initiative and there is some notable progress to share, according to Naresh Chouhan, a member of the GSMA’s AI technology and strategy team.

However, the core challenge still remains. “We’ve been monitoring the frontier models and seeing how they perform, and they’ve improved massively, but they still don’t perform well within the telco environment. They don’t understand telco language specifically well,” he noted. 

That’s a challenge that is best addressed by a broader technology ecosystem, noted the GSMA exec. “Open Telco AI is essentially an open resource for a wider community than just the mobile network operators”, with the aim of getting that broader community – including academia, the large language model (LLM) developers and more – involved in “improving AI specifically for telco use,” Chouhan explained during his presentation at the recent AI-Native Telco Forum in Düsseldorf. 

And clear progress is being made. 

The initiative’s first move was not to build models but to open up datasets. It created the Telco Capability Index, which measures AI model performance across an expanding set of telecom‑specific tasks, with the results feeding into the Open Telco AI Leaderboard that is hosted by open-source AI platform Hugging Face and which can be “used by the industry to evaluate AI in the first instance”, Chouhan explains.

Telco AI models rising

The progress made by telcos in developing useful open AI models for network operators has exceeded expectations, according to Chouhan. Just six months after the initiative’s launch, operators such as AT&T, China Telecom and SoftBank have risen swiftly up the leaderboard. “Where they hadn’t been benchmarking really well with their models [earlier this year], they’ve improved and refined and trained, so that now six out of the 10 top-performing models in our benchmarks are telco-originating,” noted the GSMA man. 

There are now scores of models openly available on Hugging Face “for people to learn from and optimise in their own way,” he added. 

A year ago the top of the leaderboard looked rather different: “Last year we had a lot more in the way of Gemini or Qwen coming through at the top end of the leaderboards. But now what we see is AT&T, China Telecom, SoftBank, Viettel and so on – and there’ll be more.”

Whereas frontier models from the major LLM developers were dominating the GSMA’s leaderboard at the start of this year, AT&T is topping the charts, thanks to the recent launch of its OTel 2.0 model, which was built using more than 400 billion tokens on AMD GPUs. The model’s capabilities include network engineering root-cause analysis, allowing it to interpret complex telco specifications better than generic AI models, Chouhan noted. 

To develop it, the US telco worked with the likes of Microsoft Azure and the GSMA, as well as AMD (it should be noted that AT&T and AMD have been heavily involved in Open Telco AI since it was founded). 

Other high-performing models developed by operators include those from Chungwha Telecom, China Telecom, SoftBank and Viettel.

“We are seeing real momentum and accuracy improving, but we are also seeing real use cases where operators are deploying AI and getting those operational savings,” noted Chouhan.

The need for interoperability

Around 18 of the Open Telco AI initiative’s 20 participants to date are operators, mostly large Tier 1s – that’s a starting position, Chouhan says, not a destination. “Where I think we can probably have the most impact is taking some of the great work and altruism of the Tier 1s that we’re working with and making that available to all,” he said. “Our focus should be about reaching out to the Tier 2s and eventually the Tier 3s in this world, and spreading that knowledge."

There is a second reason, beyond altruism, and that is interoperability. “At some stage there will be a need to look at some sort of standards around how systems may interact,” explained Chouhan. “Roaming might be looked after by an AI agent. So why not let them interoperate?”

He explained that the GSMA is set to hit its target number of signing up 20 telcos to take part in the project by the end of this year. “We’ll definitely exceed that,” he concludes. “We have proven a point, so we are now focused on diversification. That comes in two ways: Specialisation of models, and how these can grow across management, configuration or self-healing.”

At the same time, the initiative would benefit from more contributions across the broader community, as previously mentioned, in order to expand its data and help operators overcome challenges they face in adopting AI.

This includes an area that Chouhan says particularly excites him – small language models. These are tiny language models embedded on feature phones, running with no GPU and no connection, serving local dialects across Africa.

He adds: “A farmer having [the equivalent of] an assistant speak to them in their local dialect, without having an internet connection, is kind of mind-blowing. That’s providing real value.”

Spending disparity

While the Open Telco AI initiative is very focused on how to make AI more applicable to telcos, it’s also looking at how telcos use AI, and currently there’s a mismatch between network operator spending in general and how telcos are targeting their AI investments, noted Chouhan. 

He reminded the AI-Native Telco Forum audience that telcos spend billions of dollars each year on network operations, amounting to about a third (34%) of total operational expenditure (opex) on average. Yet, according to research from GSMA Intelligence (GSMAi), only 16% of generative AI (GenAI) spending has been directed at network operations, despite its share of total opex.

“There’s a disparity… What we’re trying to do is address it,” noted Chouhan.

GSMAi figures shared by Chouhan highlight that disparity. While just 16% of telco GenAI investments target network operations, the same proportion is spent on sales and marketing, and 47% is invested in customer care.

The Open Telco AI is seeking to address that disparity by exploring how AI can specifically help improve network operations efficiencies. One of the big questions to answer currently is just how much domain-specific data is needed to improve performance in a given use case through the use of a telco-specific AI model, said Chouhan. 

He also highlighted the potential gains and pitfalls from the use of agentic AI, a topic that has dominated AI-native telco conversations for the past year.

“We can see that agentic is going to play a big part when it comes to building autonomy. But the steps that we’re taking right now are inevitably going to lead to compound error,” Chouhan explained. “How can we take some steps towards measuring the impact and looking at containing it, and seeing what guardrails need to be faced?”

The answer, he suggests, is more objective evaluation frameworks, an area where a neutral industry body like the GSMA can play a key role. As a result, the Open Telco AI is working on an “evaluation framework as an open-source tool that anyone in the community can use to see what impact agents are having.”

- James Pearce, Editor, TelecomTV