At the 2026 Mobile World Congress (MWC) in Barcelona, the telecom industry confronted a pivotal question: how to integrate AI into wireless networks.

This debate is reflected in different strategic approaches: Nvidia supports the use of software-defined, distributed graphics processing units (GPUs), turning the network into a platform for both telecom and AI use cases. In contrast, Intel and Ericsson prefer a central processing unit (CPU)-based approach to manage network workloads while staying within power limits.

AI-RAN compute architecture

A key part of Nvidia’s strategy is the artificial intelligence radio access network (AI-RAN). Their hardware model uses the Aerial RAN Computer, which relies on Nvidia GPUs to handle both regular network functions and AI tasks.

In an interview with EE Times at MWC 2026, Ronnie Vasishta, Nvidia’s SVP of telecom, detailed the company’s architectural vision. “We’re a platform company,” Vasishta stated. “Our platform is a compute platform that also includes software stacks and libraries, and it addresses many different vertical implementations on top of that platform”. He noted that telecommunications represent a vertical that plays “a very important part to play in the overall infrastructure of deployment of AI.”

To support this setup, Nvidia spent eight years building Layer 1 software for the radio access network, called Aerial. The Aerial platform is the foundation for the Aerial RAN Computer, letting partners such as Nokia create their own software on top of the hardware. Vasishta described this approach as a “five-layer cake,” where the infrastructure and chip layers support a network of specialized companies.

GPUs versus CPUs for the new RAN

Nvidia and Nokia favor a GPU-based approach, while Intel and Ericsson support CPUs with built-in AI acceleration using the Xeon 6 platform.

At MWC, Intel executives argued that telecom workloads rely primarily on inference, not model training, so using external accelerators makes things more complex and power intensive. In Barcelona, Ericsson reinforced this CPU-focused approach by showing a Cloud RAN solution running on Intel Xeon 6 servers that handles core network functions without extra hardware accelerators.

Vasishta contested the premise that GPUs are fundamentally incompatible with the power constraints of cell sites. “A GPU-based system in terms of performance per watt can be incredibly efficient,” he stated.

Vasishta said that Nvidia’s Arc Aerial RAN Computer, made for cell towers, runs within a 300-watt power limit. He called this setup “as efficient as any purpose-built implementation” because GPUs deliver high computing power with less energy. He also argued that running the same tasks on a CPU would be “wholly inefficient.”

Monetizing the software-defined network

By adding graphics processing units, telecom operators can generate new revenue by offering AI-powered edge computing services to enterprise clients. Integrating AI into the radio access network creates opportunities for operators to provide specialized services, such as real-time data analysis and automation, which can be monetized.

Vasishta explained that AI enables operators to monetize use cases such as security and retail analytics through video language models. For example, network-connected cameras can actively analyze footage and trigger alerts, such as notifying staff about unusual behavior in stores or monitoring restricted areas, enabling operators to offer these value-added services to enterprise clients.

He listed ways operators could earn revenue, including public safety monitoring, retail inventory management, and asset tracking—for example, lost luggage. Additional monetization options include services for automating warehouse robots and supporting self-driving vehicles with real-time tracking and rerouting, all enabled by the network’s AI capabilities.

Ultimately, Vasishta stated, Nvidia envisions “this radio access network access points being the connectivity fabric for AI inference.”

Advancing agentic AI and autonomous networks

At MWC, Nvidia announced new efforts to develop agentic AI for managing complex software-defined networks. This technology moves networks beyond following predefined workflows, allowing them to reason independently, understand what operators want, and make decisions automatically.

Furthermore, Nvidia launched a new open-source telecom model built on its Nemotron 3 architecture. AdaptKey AI fine-tuned this 30-billion-parameter model using industry standards and synthetic logs to better understand telecom terms. The model is intended to help manage network operations, such as fault detection and planning fixes.

The company also introduced “Blueprints” for orchestrating agentic AI. These frameworks allow operators to build AI agents using proprietary operational data.

One blueprint focuses on intent-driven energy efficiency in the radio access network, integrating VIAVI simulation platforms to test energy-saving policies before deployment in live environments. Another blueprint manages network configuration through multi-agent orchestration, working with frameworks such as the BubbleRAN Agentic Toolkit to handle demand surges.

Partnerships and commercial deployments

The AI-RAN architecture is now moving from lab tests to real-world use through partnerships with global telecom companies. Nvidia announced a joint effort with industry leaders such as Booz Allen, BT Group, Cisco, Deutsche Telekom, Ericsson, MITRE, Nokia, SK Telecom, SoftBank Corp., and T-Mobile US to build future wireless networks on software-defined platforms.

Several operators have reached technical milestones with Nvidia’s technology. T-Mobile US is running live field trials using Nvidia’s AI platform and Nokia’s software, supporting commercial apps alongside regular 5G connections.

SoftBank completed a 16-layer massive MIMO field trial on a fully software-defined 5G network using Nvidia platforms. Indosat Ooredoo Hutchison in Indonesia used the technology for pre-commercial testing, including an AI-powered 5G call that remotely controlled a robotic dog.

Software deployment is growing as well. Cassava Technologies is using Nvidia’s blueprints to build an autonomous platform that improves multi-vendor mobile networks in Africa. At the same time, NTT DATA is rolling out a network configuration blueprint with a top operator in Japan to handle user demand spikes after service outages.

Navigating the transition to 6G

Industry leaders view investments in software-defined 5G networks as a step toward 6G. Nvidia presents its MWC announcements as laying the groundwork for an AI-native 6G standard.

Vasishta defined the transition to 6G as a departure from static hardware cycles. “5G Advanced and 6G should all be run on the same infrastructure,” he stated. “[It should be] software upgradeable.”

Since 6G is being developed while AI is changing quickly, operators need flexible systems. “You don’t want to have a fixed hardware platform that cannot accept the new changes. And that’s what AI native means,” Vasishta said. He added that an AI-native network is defined by “the ability to integrate AI into that infrastructure and that hardware.”

“We’re at a pivotal moment,” said Srini Gopalan, CEO of T-Mobile US. “In the U.S., we’ve laid the foundation with 5G Advanced and AI-native networks where intelligence lives inside the network. As 6G becomes the backbone of the AI era, telecom will serve as the nervous system of the digital economy, enabling autonomous systems and intelligent industries at scale and unlocking new value for customers and businesses alike.”

Vasishta predicted that future network traffic will be more unpredictable, with bursts of data from video language models, self-driving cars, and educational tools. To handle this, he said networks must “dynamically adjust to different environments, different number of demands on the network, and for that, you need AI as well.”

He advised telecom providers not to wait for official 6G hardware before upgrading their infrastructure, comparing it to buying electronics. “It’s like you don’t wait for a TV because the next model is better… to watch TV,” Vasishta said. “Buy something that can be software upgradable, and that’s what the AI native platform is.”