AI-RAN Trials Signal 6G’s AI-Native Future

This September in Madrid, the 3GPP, the global telecommunications standards body, will convene to chart the migration path from current 5G networks to the forthcoming 6G standard, which is expected to begin its rollout by 2029. The decisions made there will shape how deeply AI becomes embedded within 6G radio equipment and the broader network architecture.

The stakes are considerable. If the industry missteps, network operators could be saddled with costly AI-enhanced wireless infrastructure that fails to cut expenses, lacks proven durability in real-world outdoor conditions over years of use, and offers little advantage over cheaper, existing technology.

John Strand, a Copenhagen-based industry analyst who has advised telecom providers worldwide for three decades, points to a $1 billion deal last October as the clearest sign of AI’s growing influence. That deal saw GPU giant Nvidia invest in Nokia, the Finnish telecom infrastructure firm. According to Nvidia’s announcement, the company envisions the chips inside Advanced 5G and 6G radio access networks (RANs) resembling the servers that power ChatGPT more than specialized telecom hardware.

Nokia, Nvidia, and T-Mobile have already conducted proof-of-concept trials of this vision. Earlier this year, at T-Mobile’s innovation lab in Seattle, a single radio and an Nvidia server simultaneously managed a live 5G connection while running AI tasks such as video streaming and captioning. The demonstration worked, though it was controlled and limited to one site. Nokia’s timeline calls for broader commercial trials later this year, with a full rollout targeted for next year.

For Nokia, the deal could prove advantageous. Strand notes that RAN prices have fallen for 25 years, so adding premium AI hardware allows the company to command higher prices and “increase the entry barrier for competitors to move into this market.”

Nokia is not the only player pursuing an AI-driven 6G future. Ericsson, the Stockholm-based infrastructure maker, has been running its own AI-in-the-network trials with T-Mobile since early 2025. In June, it began selling a software upgrade that embeds AI into existing radios and base stations without requiring new hardware. Huawei, the Chinese equipment giant, has also moved in this direction, launching AI tools this year that enable networks to diagnose and repair themselves automatically.

Despite these test cases, Strand says he has yet to hear a compelling justification for AI-RAN, one that would convince operators to increase RAN spending while potentially raising energy consumption. Wireless engineers have spent decades pushing radio networks close to their theoretical efficiency limits. Kim Kyllesbech Larsen, chief technology and information officer at United Group, a telecom operator based in Hoofddorp, Netherlands, says it is difficult to see the economic rationale for deploying AI algorithms broadly across 6G RANs. He notes that companies like Seattle-based Opanga Networks already deliver some of the efficiency gains AI-RAN promises using current network architecture, with no new hardware required.

Larsen acknowledges that the trials conducted by Nvidia, Nokia, and others represent important engineering milestones and demonstrate that the concept works. But they fall short in key ways. The trials show that AI computing workloads can share space and power with the RAN, but they do not prove that AI must be deeply integrated into the RAN’s core architecture. That is a different, harder problem, he says, and one that demos so far have not resolved.

“AI-RAN will ultimately have to demonstrate that it delivers capabilities and/or economics that cannot be achieved by simply making existing RAN architectures smarter,” Larsen says. “Until that is proven, operators should evaluate AI-RAN pragmatically, rather than assuming that a new architecture is automatically a better one.”

The risk, Larsen adds, is not just unproven hardware. It is also coordination. Imagine several capable managers all trying to improve the same business, each judging success by a different metric. One minimizes energy use, another maximizes speed, a third maximizes coverage. Individually, each makes reasonable decisions. Together, they can end up working against each other. The same risk applies to a network run by multiple independent AI systems with no clear chain of command, Larsen says.

The fundamental question for 6G, Larsen argues, is not whether AI belongs in the network, but where inside the network AI algorithms should actually run. “It’s important to distinguish between AI hardware and AI algorithms,” he says. He has seen compact machine learning algorithms that boost wireless network efficiencies without requiring heavy-duty GPUs. These algorithms execute in microseconds to a few milliseconds and are well suited to more modest chips.

Larsen ran his own simulations of potential 6G AI functions to determine how quickly each needs to react. Some require microsecond-level responses, while others have more leeway. Only the fastest-reacting functions, he found, need to run on hardware at the tower itself. That means some of the speedups AI might enable can be computed off-site, “rather than assuming every cellphone tower will become an AI data center.”

Merouane Debbah, a researcher at Khalifa University’s Digital Future Institute in Abu Dhabi, has run some of the only live trials testing AI hardware in real radio conditions. He says the debate over 6G’s AI future is not simply about placing AI compute at the tower versus relying on distant cloud data centers. “My expectation is not that 6G will place one giant AI model inside every tower,” Debbah says. “A more credible architecture is hierarchical and heterogeneous: very small models inside radios and basebands for hard, real-time decisions; more capable models at edge sites; and large foundation or agentic models at regional or central levels for reasoning, planning, and coordination.”

As 6G standards begin to take shape, Larsen says the expectations placed on them by network designers will mature as well. “Ultimately, I trust good architecture more than individual hardware components,” he says. “AI-native RAN will succeed by placing the right intelligence in the right place, with the appropriate authority, latency, and safeguards, rather than trying to make every part of the network equally intelligent.”