GSMA and ITU sign letter of intent on open telecom AI

GSMA’s non-binding letter backs open telecom models, with no funding or timeline

In sum – what to know:

• Non-binding framework – The letter sets out three workstreams but carries no binding obligations or dispute mechanism.
• Open-weight telecom models – The OTel 2.0 model passed more than 30 million downloads by October, though downloads don’t measure deployment.
• Supply-chain control – Louis Powell argues that only four or five large players own telecom intelligence, leaving operators exposed if access is withdrawn.

The GSMA and the International Telecommunication Union have signed a letter of intent to collaborate on “trusted, open AI for telecoms”, built around the GSMA’s Open Telco AI program. The letter was signed at MWC26 Doha, however it is non-binding, and neither organization has published things like funding or a timeline just yet.

The underlying point is that operators shouldn’t depend on a handful of large providers for AI intelligence, especially in networks where errors are costly. The signing comes as ITU member states debate another AI mandate at PP-26, the agency’s plenipotentiary conference in Doha.

What the letter of intent covers

A letter of intent sits below a memorandum of understanding and carries no binding obligations or dispute mechanism. This one sets out a framework for collaboration on Open Telco AI, the GSMA program developing telecom models, data and benchmarks.

The first workstream covers telco-grade AI for network operations, infrastructure and standards. The second covers telecom-specific data assets, and the third covers training, deploying and evaluating the OTel open model family. They also plan to align technical standards, open data interoperability and evaluation frameworks. The stated aim is “secure, transparent and high-precision AI across global networks”.

Neither organization has disclosed funding, staffing, deadlines, a governance structure or licensing terms. GSMA chief regulatory officer John Giusti signed for the GSMA. Seizo Onoe, director of the ITU’s Telecommunication Standardization Bureau, was pictured with the signatories. His bureau would handle any standards work, an indication of where this could lead, but not when.

Why operators want open telecom models

The GSMA’s case starts with performance. AT&T’s Mark Austin said general-purpose models answered telecom-standards questions correctly only about 60% to 70% of the time, and the GSMA’s Louis Powell said gains in math and coding didn’t carry over to telecom tasks. AT&T says its telco-trained models reach roughly 90% on the same kind of benchmark, though that’s a vendor-reported figure.

The headline model is OTel 2.0, a version of Gemma 4 31B-IT post-trained on 400 billion telecom tokens, the text fragments models process. It tops the Open Telco AI leaderboard, and OTel here means Open Telco, not OpenTelemetry, the separate observability framework. The GSMA says downloads rose from 16 million in May to more than 18 million by late July and more than 30 million by October.

Downloads aren’t deployment value, and the commercial numbers are less flattering. GSMA Intelligence estimates that AI contributes roughly 2.5% of operator revenue, with deployments split 62/38 between cost-saving and revenue objectives. Only around 16% of telecom generative-AI deployments target network operations.

For the GSMA’s Louis Powell, that makes openness a supply-chain issue. He argues that only four or five big players own intelligence in the telecoms industry, leaving operators exposed if a provider withdraws access. An operator that owns a fine-tuned model retains more control over continuity and guardrails.

There are limits. The OTel paper says the resource is English-only and primarily text-centric, while its main evaluations use unseen portions of its own corpus rather than fully independent benchmarks. The paper calls these open-weight models. The weights can be downloaded and adapted, but the full training resources aren’t necessarily available under open-source terms.

ITU’s role amid its own AI debate

The ITU was already involved in much of this work, as it hosts the AI Telco Troubleshooting Challenge on its AI for Good site and is a partner on TelcoRetrieve, which tests whether models can ground answers in telecom standards, and AI Agents for RAN, which tests automated agents in a radio access network simulator. The two challenges have a combined €55,000 prize pool and were developed with the GSMA, ETSI, IEEE, ITU and Huawei. The earlier troubleshooting challenge attracted 1,255 participants from 104 countries.

The letter therefore formalizes an operational relationship rather than creating a program from scratch. It also sits alongside, but separate from, the ITU’s Open Telecom Agent-Based Intelligence initiative, or OTAI, which focuses on interoperability between AI agents.

The politics are harder to separate. At PP-26, member states are revising Resolution 214, the ITU’s AI mandate. A US proposal would replace “artificial intelligence” with “super intelligence” and bar the ITU from supporting policy, governance or ethics frameworks. Bogdan-Martin was re-elected unopposed for a second term on November 12, with 164 of 173 votes.

Open Telco AI’s participants include China Telecom, China Unicom and China Mobile, while Huawei is a sponsor and data contributor. Whether this technical collaboration can stay insulated from the fight over who controls telecom AI, and what the ITU may say about it, remains unresolved.