Column originally published on the website of the Argentine Center of Engineers.
For the Eng. Marina Rosso Siverino, president of the CAI's Technology, Innovation and Digital Transformation Commission. Second installment of the series on Artificial Intelligence and next-generation networks.
In a previous note we stated that Artificial Intelligence (AI) is no longer a competitive advantage to become a condition for operational survival in next-generation networks. No network today can manage, with traditional engineering methods, the combination of critical traffic, thousands of IoT sensors, and real-time transmission that current infrastructure demands, let alone do so within an OPEX budget and sustainable energy consumption.
The solution lies in self-configuring, closed-loop systems that detect degradation before the user notices and self-repair, using a combination of machine learning and deep learning for prediction, and reinforcement learning for real-time decision-making. Added to this are generative AI as an assistant to engineers and expert systems that automate basic rules.
That diagnosis has a concrete institutional counterpart. At the Mobile World Congress in Barcelona in March 2026, GSMA launched Open Telco AI, an alliance that brings together operators, infrastructure providers, AI developers and academic institutions with a specific goal: to build AI models trained specifically for the telecommunications business, an “operator-grade AI”, instead of continuing to rely on generic, general-purpose models.
The organization's diagnosis concludes that the most advanced AI models on the market are not significantly improving their ability to analyze 3GPP standards, solve RAN problems, or handle the specific domain complexity required by an autonomous network. Ultimately, only the 16% of GenAI's implementations in telcos is effectively applied to the network; The bulk of it is concentrated in customer service, marketing and back-office functions, that is, on the periphery of the business and not in its core value.
The gap is also reflected in the economic destination of the investment. A GSMA Intelligence survey of 700 operators The report noted that only between 151 and 201 of AI deployments had an external revenue target; the rest remained focused on costs and efficiencies. In Latin America, out of 57 cumulative deployments, 911 pursued internal objectives, and only 91 targeted products or revenue. The business exists, but it is still much closer to cost savings than to a new revenue stream.
The real underlying issue: who gets the margin
Louis Powell, director of AI technologies at GSMA, He publicly stated that It's very difficult for telcos to monetize AI if they can't make it themselves.. In other words, without their own development capabilities, telcos remain consumers of other companies' technology, with no room to build differentiated services or new revenue streams. The organization's response is Open Telco AI, which combines open models contributed by various stakeholders, along with an open data library and a public benchmarking leaderboard. It's worth examining each of these models, as they concretely illustrate what "tailor-made AI for telcos" means:
- OTel, from AT&T. It is the initiative's flagship model and tops the Open Telco AI leaderboard. The latest version, OTel 2.0, is a language model with 31 billion parameters and nearly 400 billion telecom-specific tokens—selected from over a trillion processed—combining standards documentation from seven organizations (3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN, and TM Forum) with AT&T's own data developed in collaboration with Red Hat, Dell, Microsoft Azure, and AMD. Its first version, OTel 1.0, was already trained on GSMA data; the key improvement of version 2.0 is demonstrating that a "domain-specific" model can be much smaller than general-purpose models and still be more accurate in tasks such as network troubleshooting, configuration, and product development, resulting in significant savings in computing and inference costs.
- TeleLogs, by Huawei. This is a synthetic dataset created by the NetOp team at the Huawei Paris Research Center to train and evaluate language models for the automatic diagnosis of 5G network faults. It simulates drive tests with parameters from base stations, antennas, and beamforming, along with user metrics. In other words, it's the study material that a model uses to learn why a specific network is failing, and it's already being used in the GSMA's Open Telco AI leaderboard.
- The Large Telco Model (LTM), by Nvidia and AdaptKey AI. It's a 30-billion-parameter model built on Nvidia's Nemotron 3 family, which the consulting firm AdaptKey AI fine-tuned using open telecommunications datasets, including industry standards and synthetic network logs. It's specifically designed for autonomous network tasks: detecting faults in multi-layered, multi-vendor environments, planning rollback routes, and validating the impact of a change before and after implementation. Its key differentiator is that operators have complete visibility into the training process and can fine-tune it with their own operational data without exposing it to third parties.
The benchmarks themselves, however, place a concrete limit on the enthusiasm: in TeleYAML, the test that translates instructions into valid network configurations, The best models scored less than 30 points. GSMA concludes that, for now, the safest path is not to delegate control to a single LLM, but rather to combine foundational models with specialized components, scheme validations, and human oversight. For critical infrastructure, accuracy should be measured by task, not by the overall eloquence of the model.
Adoption numbers reflect the shift towards openness: according to an Nvidia survey By 2026, 891% of operators consider open source a key element in their AI strategy (compared to 401% in 2025), and 42% are already planning to develop solutions in-house. SoftBank and China Telecom, which use GSMA benchmarks to evaluate their own models, complete the picture of an initiative that has moved beyond mere rhetoric. Mobile World Congress 2026 itself showcased this leap from demo to implementation: operators like Orange, Deutsche Telekom, and T-Mobile demonstrated AI already integrated into their networks live, not just as a proof of concept.
The operational incentive is also quantifiable. The operators consumed approximately 290 TWh of electricity in 2023 —around 11,330 TW of global consumption— and, for most, networks accounted for more than three-quarters of electricity use. In a sample of 19 operators, energy intensity per unit of data fell between 101,330 TW and 301,330 TW annually between 2019 and 2023. But GSMA warns that 5G densification can increase total consumption even with more efficient equipment; hence the value of features such as AI-optimized sleep modes and dynamic shutdown of idle capacity.
The other side: the gap between discourse and the engine room
Measured with a common taxonomy, the gap is less anecdotal and more visible. A 2025 TM Forum global survey The survey placed only 41% of participants at level 4 of autonomy; 171% reported level 3, 31% level 2, 36% level 1, and 12% level 0. In that survey, the stated goals for 2026 were much more ambitious: 23% expected to reach level 4 and 46% level 3. The gap between the current state and the target confirms that the main obstacle is not the availability of models, but the underlying infrastructure on which they must operate: legacy and fragmented BSS/OSS systems, outdated inventories, data lacking common context, and resistance to replacing existing systems. Without addressing this underlying issue, no AI model can deliver on its promises.
An Argentine reading: scale and concentration
In Argentina, the challenge is no longer abstract. ENACOM registered 1,887 5G sites by the end of 2025, This represents a growth of approximately 591% compared to 1,190 six months earlier. However, 1,306 sites—691%—were concentrated in the City and Province of Buenos Aires; Córdoba and Santa Fe accounted for another 249. This asymmetry matters: training and operating network AI involves not only optimizing the highest-traffic nodes, but also maintaining quality, energy efficiency, and predictive maintenance across vastly different geographies, densities, and transport links. For local operators, the advantage will not come from copying a global model, but from fine-tuning it with their own data and with governance that ensures every automated recommendation is traceable, reversible, and secure.
The business opportunity presented by GSMA is well-founded: without their own models, operators risk repeating with AI what they already experienced with the cloud, ceding control and profit margins to third parties. But this opportunity depends on telcos first streamlining their internal infrastructure and resisting the temptation to sell autonomy they don't yet possess. As noted in our previous article, the fundamental condition remains the same: before AI can be deployed massively across critical infrastructure, it must "speak the telco language" with the precision and operational security that language demands. The question that remains for the industry is not whether AI will transform the telecommunications business, but who will capture that value: the operators who invest in their own capacity, or the hyperscalers who already have it.
Fountain: CAI