Nokia, together with Nvidia, has announced the industry’s first commercial AI-RAN platform, calling it the biggest shift in radio networks in decades. AI-RAN – RAN stands for Radio Access Network, the gear linking phones to the broader network – is not a mere generational upgrade: it embeds AI compute directly into the access infrastructure, with the stated goal of doubling network capacity.

The move is more radical than it appears. For years, the telecom world has seen network virtualization (vRAN, Open RAN) as a path to reducing dependency on proprietary hardware. Now, the injection of Nvidia GPUs – the computational heart of the platform – turns the radio base station into an edge AI processing node. It means the same infrastructure handling voice and data traffic can run LLM inference, process real-time video streams, or fine-tune lightweight models without ever leaving the operator’s perimeter.

To grasp why this matters beyond telecom circles, just look at what happened in data centers. The arrival of GPUs shifted the center of gravity from specialized mainframes to software-defined architectures built on accelerated commodity hardware. AI-RAN promises to do the same for the mobile edge: radio units stop being closed boxes and become programmable compute servers capable of hosting third-party AI workloads.

Winners and losers. The first winner is Nvidia, which opens a new market of thousands of edge sites to equip with GPUs. Nokia, for its part, positions itself as the software orchestrator of an ecosystem previously dominated by a handful of radio hardware vendors. The losers are traditional suppliers that lack the flexibility to integrate AI acceleration into their stacks, and perhaps the hyperscale cloud providers, because the ability to process sensitive data directly on the mobile network makes centralized solutions less attractive when latency and privacy matter.

Sovereignty and data control. AI-RAN has deep implications for anyone evaluating on-premise LLM deployment. A mobile network embedding programmable GPUs essentially becomes a distributed, self-hosted infrastructure, able to keep data within the physical and jurisdictional boundaries chosen by the operator. In regulated sectors (GDPR, healthcare, defense), this means offering AI services without traversing external public networks, with granular control over the processing chain. For enterprises, the option to rent AI-RAN capacity from telcos could become an alternative to both cloud and traditional on-premise servers.

The second-order effect is a shift in market incentives. If mobile operators also become providers of AI compute, the clear separation between connectivity and cloud services dissolves, creating direct competition with AWS, Azure, and Google Cloud precisely on low-latency inference and data residency. It is no accident that Nokia and Nvidia are emphasizing that AI-RAN can double network capacity: a concrete argument to convince telco CFOs to invest in hardware that both improves network performance and enables new revenue streams.

The missing piece, which will prove decisive, is Total Cost of Ownership (TCO). Edge GPUs consume power and require maintenance. Compute density per watt and the ability to orchestrate mixed workloads (network + AI) will determine whether the model is economically sustainable or remains a technology showcase. But the direction is clear: network infrastructure is getting ready to become the physical layer of distributed AI, just as the debate over digital sovereignty and cloud costs grows ever more heated.