Intelligent RAN Forum panel says AI-RAN efficiency pays, demand doesn’t yet
In sum – what we know:
- Where the payoff sits – Energy optimization, alarm-storm root-cause analysis and AI-assisted planning are where operators and vendors see measurable value from AI-RAN today.
- The forgotten costs – None of the panelists named hardware as the underestimated line item, pointing instead to the skills gap, legacy system integration and lifecycle management.
- A circular revenue case – Application developers won’t build RAN-specific applications until edge infrastructure exists, and nobody builds the edge without applications to fill it.
AI is already paying for itself in the radio access network through energy savings, fault diagnosis and capacity planning, according to a panel of operators and vendors at the Intelligent RAN Forum. Rob Hughes, head of wireless marketing at 1Finity (Fujitsu’s network infrastructure brand); Güneş Kesik, associate R&D director at Turkcell; and Gabriel Ionita, director of group partnering and business development at Deutsche Telekom all put the payoff in network operations.
Tom Camp, the Telecom Council executive director moderating the panel, framed the question as a split between measurable value and investment thesis. The discussion that followed was really about budget allocation: where the next dollar goes when the opex case already works and the revenue case still needs proof.
Ops first, revenue later
“We see the clearest value in energy optimization, better resource use, and operational automation,” Kesik said. Turkcell has built its own AI-driven optimization and network management tools, but Kesik called the current state “just a starting point.”
Ionita agreed, with day-to-day network operations first and cost savings second. He said pilots that don’t end in a customer-facing result should be rethought. “I’m not a proponent of doing pilots and doing network things that do not have an end result in customer improvement,” he said.
Hughes, from the vendor side, described the same operational priorities: power savings, root-cause analysis that cuts alarm-storm resolution from hours or days down to minutes, and AI-assisted network planning. He also recalled a performance figure from work with ecosystem partners on AI-for-RAN applications roughly 18 months ago, where he saw 20 to 50 percent cell-edge uplink throughput improvements. He described it as a memory of trial results rather than a product specification, with the range depending on conditions.
Camp then asked who owns the benefit, and there is no obvious answer inside a large operator. Kesik’s response was to move the accounting up a level. “Energy savings can reduce operational costs, but better use of capacity of our existing infrastructure can delay capex investments,” he said. “We need to evaluate the impact of AI at the overall network economic level. Maybe just looking at single organization or single KPI would not be enough.”
The benefit is spread across opex, deferred capex and workforce productivity, so no single budget holder owns it.
The panelists also conceded the limits of what they can prove. Camp asked whether customer experience metrics are being measured around AI-driven network improvements. “I would like that to be the case,” Ionita said. “That’s one of the asks.” Kesik said proving AI’s impact on customer experience is at very early stages, and that energy, capacity utilization and operational effort serve as stand-in KPIs for now. Hughes said operator skepticism is about maturity rather than performance. “You see some trials and some smaller deployments, but not large-scale deployments yet,” he said.
No panelist pointed to a published customer experience metric tied to AI-RAN.
Skills and legacy costs
Camp asked what the industry is underestimating about cost and complexity. None of the three pointed to hardware.
Hughes pointed to the skills gap. Operators have a good handle on what these deployments cost, he said, but AI-RAN requires RAN knowledge blended with data-center and IT expertise, and that combination is the part people understand less well. It is why he pushes small, live deployments over roadmap exercises: they expose the real architecture, business model and staffing requirements before anyone commits to scaling. “It’s important for people to get started in some real live deployments, even if it’s small scale, just to start developing a real feel of what exactly do these business models look like,” he said.
Kesik said compute, software, data integration and lifecycle management all add cost, and that networks are getting more complex regardless of AI, through new enablers and new vertical service expectations. The question is whether the trade is net positive. “The complexity we remove through automation should be greater than the complexity we introduce with AI,” Kesik said.
Ionita pointed to legacy integration. “The things that get forgotten are the integrating with the legacy systems,” he said. Operators carry decades of built-up infrastructure, and wiring those systems into any AI pilot is expensive.
Camp then moved to audience questions: do AI models survive vendor software updates, topology changes and new band additions? Hughes said the effort is real but that Open RAN experience carries over, with multi-vendor integration and lifecycle discipline developed in disaggregated environments applying directly to AI-RAN. Integration effort has dropped noticeably, he said, but the incremental work has not disappeared. Kesik said Turkcell already runs more than 30 vendors in its network, that interoperability comes from standards, and that programmability only pays when automation converts that flexibility into economic value.
Kesik described three things that separate investable innovation from the next upgrade cycle: measurable economic impact, repeatability and scalability.
Demand isn’t there yet
Camp put the CFO question to the panel. You get one incremental dollar for the RAN. Where does it go?
Ionita pointed to the network already in place. “Making better use of the network we already have,” he said, describing better efficiency from what is built, with the customer result as the end view.
Kesik said he would love to invest the next dollar in experimenting with areas that could create new revenue or new product lines for Turkcell. “But we are asked to be pragmatic,” he said. The dollar goes to AI-driven automation and closed-loop optimization — AI for the network rather than the network for AI — with investment in network-for-AI following once demand validates new use cases. With Dell’Oro projecting the RAN market broadly flat through 2030 but the mix shifting, the pressure to show returns on existing infrastructure is only going to increase.
Hughes argued for opportunity-led deployment tied to a paying enterprise customer: build the architecture and business model around one real account, then grow as revenue justifies it. He added that blending AI with private wireless embeds connectivity deeper into the enterprise relationship. “Your traditional business is strengthened if you can add more value by moving up the value chain to the customer,” he said.
Camp then asked whether the efficiency case alone is strong enough to justify the investment.
Ionita would not generalize. “We need a real customer problem. I believe we need a customer willing to pay. So enough demand,” he said. Kesik was more definitive: for AI-for-network specifically, operational costs are large enough that automation and energy efficiency justify further investment on their own. Hughes said the tangible results are there today, but that public-network operators are not ready to deploy at scale.
According to Hughes, hyperscalers are excellent at AI cloud, but European siting is constrained by power and water, data sovereignty is a real concern, and the direction of travel is compute moving toward the data rather than data crossing the world. Operators hold real estate, power, connectivity, spectrum and enterprise relationships — advantages if they choose to use them.
Kesik put a number on it — more than 30,000 Turkcell sites across Turkey as a potential distributed compute footprint. It only matters, he said, if verticals demand local execution for sovereignty, privacy or latency-critical work like automated driving or smart manufacturing. “It really depends on customer demand,” he said.
Ionita framed GPU as a service as “a focus market, and maybe even a partnership opportunity, rather than an attempt to replace the hyperscalers.” Camp noted that Deutsche Telekom invested heavily in GPUs, but the panel didn’t dwell on the specifics.
The panel agreed on where value shows up today and on the limits of what can be proven, but not on what would make the revenue case financeable. Hughes described the difficulty as circular: application developers won’t build RAN-specific applications until the edge infrastructure exists, and nobody builds the edge without applications to fill it.
The panel left open whether operators can wait for that demand to materialize while hyperscalers move into edge real estate that operators already own.
Table of Contents
