๐๐-๐ฅ๐๐ก & ๐ฆ๐ฝ๐ฒ๐ฐ๐๐ฟ๐ฎ๐น ๐๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐
AI-RAN impact on Spectral Efficiency
TELECOMMUNICATIONS
Amit Parashar
10/3/20263 min read
Operators have locked roughly $240 billion into spectrum licenses over the past three decades. That capital sits on the balance sheet as a fixed asset with one job: generate capacity. Yet most operators cannot multiply that capacity without bidding again at the next auction. AI-RAN changes that equation. If spectral efficiencyโmeasured in bits per second per Hertzโcan be doubled, then the same spectrum generates twice the capacity. Nokia's roadmap claims this is possible: more than 20% gains already demonstrated, 50% by 2027, and a complete doubling by 2028. The question I face in every solutioning conversation is whether that doubling feels real or aspirational when you account for the actual constraints of production networks.
Spectrum is the most capital-intensive, longest-lived asset on an operator's books. Unlike hardware, which refreshes every five to seven years, spectrum licenses lock in costs and constraints for a decade or more. Spectral efficiencyโhow many bits you can squeeze from each Hertzโis therefore the only sustainable lever to multiply capacity without fresh capex. AI-native RAN algorithms promise to optimize that lever in ways classical link adaptation and beamforming cannot.
The evidence supporting near-term gains is already in the field. Ericsson reports up to 20% higher downlink throughput from more than 15 live trials worldwide. SoftBank's first commercial validation in Japan showed up to 25% spectral efficiency improvement in dense cells, with average gains around 10% across multiple locations. A 7g.network benchmark covering SK Telecom, T-Mobile, and Vodafone confirms that 10 to 18% spectral efficiency improvements are achievable in operator networks today with telco-grade AI. Those numbers matter because they are field-verified, not lab claims.
Nokia's trajectoryโ20% proven, 50% by 2027, more than 100% by 2028โis best understood as a target co-developed with Nvidia and built into product planning. When I map current trial data to a full doubling, I see a credible direction but a substantial gap. The trials show what is achievable in controlled scenarios with dense cells, favorable signal-to-noise ratio conditions, and optimized traffic mixes. Doubling capacity across a mixed network of cell types, hardware generations, and real-world traffic requires several things to align simultaneously.
First, the AI models must remain compact, deterministic, and energy-efficient. Ericsson and other vendors are explicit on this point: telco-grade AI cannot be experimental or resource-hungry. Models must fit within latency and power budgets defined by existing basebands and uplink capacity. That constraint is real, and it means the most aggressive algorithms stay on the shelf in favor of ones that are reliable and auditable. Second, uplink load increases with more aggressive downlink scheduling. That is not free; operators must have headroom on the return link or accept tradeoffs in peak spectral efficiency. Third, verification at scale is non-trivial. Moving from a trial in one market to fleet-wide rollout requires independent benchmarking, A/B testing, and confidence that gains are repeatable, not artifacts of specific conditions.
The industry is moving in the right direction. Nokia and Nvidia are extending AI from the scheduler into beamforming, link adaptation, and interference management. The compact models are maturing. Operators are running real pilots with serious measurement. But I do not see evidence yet that 2x spectral efficiency will be the default across broad operator networks by 2028. More likely, we will see 10 to 30% gains become standard, with the best-optimized, densest cells approaching 2x in specific scenarios.
That distinction matters for how operators plan. If you are building capacity projections, separating what is proven today from what is roadmapped for 2027 and 2028 is essential. The 20% gains can inform near-term decisions. The 2x target is a useful benchmark for long-term strategy, but not a planning anchor until operators have verified it in their own networks.
#SpectralEfficiency #AI-RAN #TelecomNetworks #5G #NetworkCapacity