LITEON and Nvidia Push AI-RAN Forward
LITEON Technology is accelerating the commercialization of AI-native radio access networks (AI-RAN) by integrating its open radio units with NVIDIA's AI Aerial platform. The move, showcased at MWC Barcelona 2026, aims to build more scalable and efficient RAN architectures using GPU acceleration, signaling a major shift in how telecommunications infrastructure is being built.
The architectural shift underpinning this collaboration is the move away from traditional RANs, which rely on specialized, purpose-built hardware, to a software-defined model. This new approach utilizes general-purpose GPUs on commercial off-the-shelf servers, enabling a single, unified infrastructure to handle both complex AI workloads and RAN signal processing. The parallel processing capabilities of GPUs are uniquely suited for the intense computational demands of modern wireless communication and AI. This results in significant performance and efficiency gains; one proof-of-concept demonstrated that GPU-based processing could be 12x to 30x more energy-efficient than CPU-based execution for certain workloads. This helps reduce the total cost of ownership (TCO) for virtualized RAN (vRAN) architectures. NVIDIA's Aerial platform is a full software and hardware suite designed to build these AI-native networks. It includes the now open-source Aerial CUDA-Accelerated RAN libraries for developing Layer 1 and Layer 2 functions, alongside the Aerial Omniverse Digital Twin, which allows for physically accurate simulations of entire city-scale networks before deployment. LITEON's role is providing the critical O-RAN compliant hardware, specifically the sub-6 GHz and millimeter wave Open Radio Units (O-RUs). These units adhere to the O-RAN 7.2x fronthaul interface standard, which is designed to reduce integration complexity and deployment barriers for network operators. This AI-RAN model is already moving from labs to live field trials with major operators like T-Mobile U.S., SoftBank, and Indosat Ooredoo Hutchison. In performance benchmarks, ecosystem partner SynaXG demonstrated a fully software-defined AI-RAN achieving 36 Gbps throughput with under 10 milliseconds of latency on a single NVIDIA GH200 server. This convergence of RAN and AI workloads transforms base stations into distributed edge data centers. The infrastructure can run AI inference for new services during periods of low network traffic, turning a traditional network cost center into a potential revenue-generating platform for low-latency edge AI applications.
