The GSMA is urging telecommunications operators to develop in-house AI models rather than relying on hyperscalers, citing a performance gap in frontier models for domain-specific tasks like 3GPP standards. The organization's Open Telco AI initiative aims to provide benchmarks and data sets to help operators build autonomous network capabilities and improve security.
The push for domain-specific AI models suggests that general-purpose LLMs from hyperscalers cannot yet manage the technical rigors of Level 4 and Level 5 autonomous networks. By developing in-house expertise, operators aim to secure data sovereignty and prevent service disruptions caused by external API changes or regulatory restrictions on proprietary models. This shift mirrors a broader trend in the streaming and connectivity ecosystem where infrastructure owners are reclaiming the software stack to optimize performance and reduce long-term licensing costs. Watch for whether these open-source foundations allow smaller regional operators in Europe and Africa to close the technical gap with U.S. and Chinese incumbents.
Ericsson has positioned itself as a leading counterweight to the hyperscaler-dependent model the GSMA is warning against. At MWC 2026, Ericsson networks chief Per Narvinger described AI-native RAN optimization as already delivering 10% spectral efficiency gains on algorithms refined over 30 years, arguing that purpose-built AI models running on custom silicon inside baseband units represent a fundamentally different approach from calling external APIs. The company demonstrated its link adaptation technology with Bell Canada in field tests during April 2025 and announced a similar collaboration with AT&T on Intel-based cloud RAN infrastructure at MWC. Narvinger projected that by the end of 2026, Ericsson will have 10 AI-ready radio products in market, each featuring neural network accelerators embedded in its Ericsson Silicon chips.
The GSMA's call for operator-built AI models intersects with a broader geopolitical conversation about technology supply chains in telecom. Nokia CEO Justin Hotard described the Europe-US relationship in telecom networks as one of co-dependence, noting that equipment may be assembled in one country, run software from another, and depend on chips designed elsewhere. That structural reality complicates the GSMA's vision of full AI sovereignty for operators, since even in-house models require training infrastructure and semiconductor supply chains that cross borders. Dell_Oro Group data cited in the same reporting shows that in markets where Chinese suppliers are restricted, operators often have only two or three realistic RAN vendor choices, concentrating leverage among a small number of Western equipment makers.
On the traffic and workload side, Ericsson's own research quantifies why the GSMA's timing matters. The June 2025 Ericsson Mobility Report found that generative AI currently represents only 0.06% of total mobile data traffic but carries a 26% uplink share compared to the typical 10%, a ratio that will strain network planning as agentic AI workloads grow. ChatGPT alone accounted for 60% of total AI traffic and 70% of all AI uplink traffic as of April 2025, with 546 million monthly active users. Ericsson's separate analysis of claims an 80% reduction in time spent on analysis and decision-making when systems handle routine network management, a figure that underscores why operators see domain-specific models as essential to reaching Level 5 autonomy without ceding operational control to third-party platforms.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source