AI-driven sustainability in telecom: Ericsson's vision for a greener 6G future

As a founding member of the Next G Alliance, Ericsson champions North American leadership in advancing next-generation networks, including 6G. Central to this effort is the ATIS Next G Alliance Green G Working Group, chaired by Ericsson, which is dedicated to embedding sustainability into the core of telecom innovation. Its mission is to ensure next-generation networks are high performing, energy efficient and environmentally sustainable.

Ericsson integrates AI into its networks, and is driven by its commitment towards sustainability, operational efficiency, and technological innovation. By leveraging AI and automation, Ericsson aims to optimize energy consumption, reduce CO2 emissions, and enable environmentally sustainable telecom operations. The company’s goal is to empower service providers with intelligent solutions that enhance network performance while aligning with global NetZero targets. [1]

AI technologies offer tools to optimize network operations. While AI is a transformative technology, the environmental footprint of its lifecycle could be significant. The growing adoption of AI is accompanied by substantial energy demands for applications that rely on the use of Large Language Models (LLMs) and Generative Al. This energy demand is majorly driven by data center operations and computational processes like model training and inference. As AI usage grows across multiple sectors, the operational energy demands are expected to account for a significant amount of AI's environmental footprint, making it critical to address this sustainability challenge.

The whitepaper, "Sustainable AI in Telecom: Promises and Challenges in 6G," published by the Next G Alliance Green G Working Group directly aligns with this mission. It explores the transformative potential of AI to drive sustainability in telecom while addressing its environmental impacts. By tackling critical questions—such as how to reduce AI’s carbon footprint and leverage its capabilities to create sustainable networks—the paper highlights strategies that align with Ericsson’s vision of sustainable innovation.

The objective of this whitepaper is to introduce the key concepts of “AI for Sustainability” and “Sustainable AI,” emphasizing the importance of balancing AI’s transformative potential for efficiency with the need to manage its environmental impacts. Sustainable AI and AI for Sustainability are two critical approaches that are shaping AI implementation. AI for Sustainability focuses on optimizing network operations to reduce energy consumption and improve resource efficiency, minimizing the environmental footprint of telecom networks. In contrast, Sustainable AI emphasizes the implementation of AI technologies to ensure sustainability in design, deployment, and use. Together, these approaches help in understanding AI’s dual role in mitigating its environmental impacts and driving sustainability across sectors.

Achieving sustainable AI deployment requires a focus on energy sourcing, particularly the integration of renewable energy into data center operations. Dynamically allocating workloads to data centers powered by renewables or scheduling high-demand tasks during periods of renewable energy surplus can significantly reduce emissions while supporting sustainability goals. These strategies enable AI systems to operate efficiently without compromising environmental objectives. While the white paper addresses overall energy consumption, this blog post delves deeper into the energy demands of the Radio Access Network (RAN).

The RAN accounts for over 80% of total energy usage in mobile networks, compared to approximately 12% for the core network [2]. As part of the shift from 5G to 6G, energy consumption in data centers is expected to rise due to the migration of RAN components and network functions to centralized infrastructures. In this context, AI plays a critical role in optimizing power usage, reducing operational inefficiencies, and driving sustainability across the rapidly evolving telecom ecosystem. [3]

Energy consumption in mobile networks

Techniques to reduce energy consumption of AI/ML

Model simplification

Nature inspired models tend to be smaller, sparser and computationally efficient. This research paper [4], applies predictive modeling and intelligent data pre-processing techniques for traffic forecasting, network optimization and resource allocation that would be instrumental for operators in routing traffic in different regions.  Ericsson demonstrated this ultra-low power AI at MWC 2024, which uses a novel neuromorphic-AI-based approach for radio channel estimation and showcased the feasibility of low-compute and low-energy AI using AI-based radio receiver use-cases. The neuromorphic neural network central to this demonstration, functions like the human brain, where only the neurons detecting a change are active and no computations are needed for neurons in remember state.  The fraction of inactive neurons translates directly to an energy efficiency gain compared to traditional deep neural networks. [5] The development of such AI architectures in delivering impactful solutions is key for the future of wireless communication.

Distributed training  

By distributing the training process across devices, federated learning reduces the dependency on centralized data processing and the associated energy costs of transferring large datasets to data centers. This technique also enhances data privacy and supports real-time, energy-efficient model updates directly at the network edge. In cellular networks, distributed training frameworks like resource-aware federated learning adaptively allocate resources across base stations to reduce energy usage while maintaining model accuracy.

Algorithm optimization

Pruning: Reducing the size of neural networks by removing redundant parameters or connections without significantly affecting performance

Quantization: Lowering precision (e.g., from 32-bit to 8-bit) for computations and weights in AI models, thereby saving memory and reducing power consumption. These techniques have proven particularly effective for embedded AI systems, where computational and energy constraints are critical.

Dynamic resource allocation

AI systems often consume more power during inference than training because inference is performed repeatedly at scale, often in real-time. For example, when AI models are deployed in applications like search engines, voice assistants, or recommendation systems, each query or user interaction triggers the model, leading to millions or billions of inferences daily. Unlike training, which is a one-time process (albeit computationally intense), inference workloads require constant computation across distributed systems, amplifying energy consumption over time.

Evaluating AI’s sustainability involves conducting a comprehensive Life Cycle Assessment (LCA), which examines environmental impacts rising from embodied and operational emissions. Metrics such as energy consumption, carbon emissions, water usage, and resource depletion are assessed to identify hotspots and opportunities for improvement. Key Performance Indicators (KPIs) are essential for measuring AI’s alignment with sustainability objectives. KPIs provide specific benchmarks, such as energy consumption per inference or carbon emissions per training cycle, enabling organizations to track performance effectively. Together, these metrics guide decision-making and foster accountability in sustainable AI practices.

Ericsson AI-based energy optimization solutions

AI solutions can be applied to at different levels in the network to address discreet customer requirements. The following sections describe various Ericsson AI-based energy optimization solutions.

Energy Infrastructure Operations

Reducing energy consumption and decreasing site visits leading to lower CO2 emissions are top of mind for communications service providers (CSPs) today. Ericsson has developed an energy management solution, called Energy Infrastructure Operations [7], that leverages AI and advanced data analytics to optimize energy consumption across the network infrastructure. Using cutting-edge AI technology, the solution creates energy efficiencies on the radio network, where most savings can be achieved. The solution not only addresses site-related energy savings, but also operational efficiencies to enable less site visits to be performed, ultimately resulting in CO2 emission reduction across multiple layers. Depending on engagement-specific parameters, the solution can achieve ~15% decrease in energy-related OPEX, ~15% reduction in site visits related to passive infrastructure and ~30% reduction of energy related outages. The solution has been trialed with customers in Europe, Asia, Middle East and Latin America.

Intelligent RAN power saving solution

Ericsson has partnered with customer in the Middle East [8] to deploy a cutting-edge AI and ML solution to significantly reduce energy consumption across customer’s network operations, marking a significant advancement in the telecom industry’s efforts towards environmental sustainability. A successful proof-of-concept (PoC) was conducted where Ericsson's Intelligent RAN Power Saving solution, part of Ericsson's Service Continuity AI app suite, demonstrated about 20% on 5G daily power saving capabilities. The solution leverages a machine learning prediction model that continuously analyzes real-time network data. Through intelligent decision-making capabilities, it determines whether to deactivate, activate, or maintain network components based on the data and activity in neighboring cells. This enables precise energy management and operational efficiency, in addition to resulting in reductions in carbon dioxide emissions and operating costs.

AI MIMO sleep

Uncovering energy saving opportunities at a granular or node level requires energy-metering features that support the energy analytics function. MIMO sleep is an efficient feature to maintain user experience while minimizing waste when lower capacity is sufficient. The problem was that manual configuration was required, which is both time consuming and less efficient. To address the issue, Ericsson developed AI-powered MIMO sleep [9], which automates parameter setting to reduce manual work and improve feature performance (both for KPIs and energy savings) at the node. An ML algorithm observes, predicts, and responds to user traffic. In a proof of concept, the ML algorithm was exposed to several weeks of traffic data at each site, then allowed control over Sleep Mode activation at each site. ML Sleep Mode management delivered 14% savings in energy consumption at each site, outperforming manual management. Customer KPIs were maintained during automated tower activation or deactivation.

Remote electrical antenna tilt

Looking ahead, we see that Reinforcement Learning (RL) approaches will further improve energy savings and network performance. RL is particularly useful in the type of dynamic, complex, and high-demand environments that constitute mobile networks. There are multiple ways that RL can be applied to networks in general and energy saving specifically. One example is the two successful trials Ericsson concluded applying reinforcement learning to remote electrical tilt of antennas (RET) [10]. At a glance it doesn’t seem that complicated but every time you tilt an antenna it changes the shape of the cell the antenna is in. This in turn affects the user experience of those served by that cell and the cells around it, further tilting of surrounding antennas has a cascading effect in the network. This makes it all the more impressive that in a single live network, when optimizing for reduced ERP, Ericsson and the partnering service provider caused a 20% decrease in DL transmission power without affecting performance.

Ericsson site energy orchestration

Ericsson's Site Energy Orchestration solution leverages AI and ML to provide intelligent energy management for Communication Service Providers (CSPs). This solution optimizes network energy consumption by using ML, AI RAN applications (rApps), RAN data, and external data interfacing to enable data-driven decision-making, predictive maintenance, and real-time energy orchestration. It helps CSPs implement peak load shaving and load shifting strategies, reducing operational expenses while improving sustainability. By dynamically orchestrating energy sources based on RAN demands and external conditions, CSPs can secure network performance, lower costs, and minimize their carbon footprint. This solution creates new revenues opportunities for CSPs by participating in energy grid balancing programs that support grid resilience and the transition to renewable energy sources.

Ericsson’s approach aligns operational performance with environmental responsibility, demonstrating how AI-driven energy management can drive both business and sustainability benefits. [11] 

Every hour is earth hour for networks

At Ericsson, we know that business can only thrive in a sustainable environment. But, of course, unlike Earth hour, mobile networks cannot be allowed to power off so we must all make sure that we optimize every hour. We take this challenge seriously and provide solutions across mobile networks and their lifecycles. We believe that this benefits us now and future generations.

The Next G Alliance Green Working Group has written this paper to address an urgent question: How can the telecom industry leverage AI to meet ambitious Net Zero and sustainability goals, while ensuring that AI technologies themselves are designed and deployed sustainably? By exploring the promises and challenges of AI in the context of 6G, this whitepaper provides actionable insights and underscores the importance of collective efforts to achieve a balance between innovation and sustainability

Through this work, the Green G Working Group aims to foster a deeper understanding of sustainable AI practices and catalyze industry-wide action to build a greener, more resilient telecom ecosystem.