Nokia is presenting its new AI-RAN platform as a practical route to extracting more capacity from spectrum operators already own. Built on Nokia’s anyRAN software and Nvidia’s Aerial system, it also represents something larger: an attempt to repair Nokia’s underperforming mobile-radio business through software subscriptions and external AI silicon.
The technology may prove significant, but its most ambitious performance figures remain targets, and a rival has already commercialised a different approach.
Nokia is turning radio capacity into software
Nokia says early work has produced spectral-efficiency gains above 20%. It is targeting 50% by 2027 and more than 100% by 2028, which could theoretically double the capacity available from existing spectrum. The latter figures are projections, not measured commercial results.
Pilots are expected near the end of 2026, followed by commercial availability in 2027. Operators would subscribe to the capability rather than wait for a complete network replacement. Deployment options include a GPU-powered card for existing AirScale sites, a standalone AI-RAN node and a cloud-server configuration delivered with partners.
That model could change radio economics. Spectrum licences and physical upgrades are expensive; software that improves utilisation creates recurring revenue for Nokia while allowing operators to delay some capital expenditure. It also gives networks a shared platform for radio processing and other AI workloads.
The launch is especially important because mobile infrastructure has been chief executive Justin Hotard’s most difficult portfolio problem. Nokia reorganised the business and pursued additional savings after acknowledging weak returns. Its October 2025 partnership with Nvidia included a $1 billion investment for roughly a 3% stake.
Using Nvidia silicon and CUDA may let Nokia reduce costly custom-chip development and concentrate on higher-value software. Investors have welcomed its AI and cloud direction, while Omdia estimates a cumulative AI-RAN opportunity above $200 billion by 2030.
The claim to be first depends on the definition
Nokia calls the product the industry’s first commercial AI-native RAN platform. That can be defended if the category means a GPU-accelerated architecture, but it does not mean Nokia was first to sell AI-based radio optimisation.
Ericsson began offering an AI-in-RAN subscription in June. It reports up to 20% higher downlink throughput and 10% better spectral efficiency across more than 15 live deployments. Its software runs on existing baseband silicon and requires no GPU, giving it an availability advantage and a different cost profile.
The contrast exposes the strategic choice. Nokia is betting that shared accelerated computing will enable larger gains and more flexible AI workloads. Ericsson is prioritising hardware independence and immediate deployment on installed equipment.
Nokia describes its wider ecosystem as Open RAN-compliant and includes merchant silicon, yet the advertised efficiency case currently depends on Nvidia’s stack. Nokia’s technology leadership has acknowledged that some Layer 1 software is tied to the underlying hardware. Open interfaces and supplier dependence can coexist, but operators should distinguish between them.
The practical comparison must include GPU cost, electricity, cooling, integration and operational skill—not just capacity gained. A 20% laboratory improvement may have very different economics at a busy urban site and a lightly loaded rural one.
Operators need proof from diverse live networks
Before committing, carriers should run pilots across different spectrum bands, traffic patterns and site configurations. They need independently observed performance at peak demand, alongside energy per delivered bit, latency, reliability and maintenance effort.
Subscription terms matter too. Operators should know which functions continue if a licence ends, how model updates are validated and whether workloads can move to alternative hardware. A platform that saves spectrum but creates expensive lock-in may simply exchange one scarce resource for another.
Nokia’s timing nevertheless makes sense. Mobile traffic continues to grow, spectrum is finite and conventional radio improvements deliver diminishing returns. AI-assisted optimisation can become commercially valuable even if the headline doubling takes longer than forecast.
This is best viewed as a comeback in progress. Nokia has selected a powerful partner, a recurring-revenue model and a problem operators genuinely need to solve. Ericsson has shown that the market already exists, while Nokia is arguing that GPU acceleration can expand it. The winner will be determined by verified network economics, not by which supplier draws the category boundary most favourably.
