The financial side of AI-RAN can be divided into three different areas:
1) Improving the network’s own performance, where the value comes primarily from savings rather than new revenue.
a) Spectrum efficiency: If AI algorithms can, as Nokia estimates, even double data capacity per existing band, operators could achieve significant savings. The same amount of traffic can be handled more efficiently without a corresponding need to acquire more spectrum or build new base stations. This is particularly relevant in crowded urban environments. This is not really a question of whether computing is cheaper in a base station than in a large data center. RAN signal processing is done close to the network edge anyway due to real-time and latency requirements. More importantly, will the additional investment in GPU hardware yield sufficient benefits through better spectrum efficiency and network capacity?
b) Network programmability and flexibility: GPU-based architecture makes it easier to update and optimize network functions via software without the constant need to renew physical equipment.
2) Nokia’s own cost structure: savings in chip design
By transitioning from its own baseband chip development (ABIP was the last product) to Nvidia’s GPU platform, Nokia can reduce its reliance on in-house chip design and instead lean heavily on the scalable Nvidia ecosystem. This allows Nokia to reduce R&D costs related to in-house chip design, which likely constituted a significant portion of Radio Networks’ R&D expenses (totaling €2,076 million in 2025). This could be a structurally significant saving, which is particularly relevant given Radio Networks’ weakened operating profit margin (5.5% in 2024 and 2.8% in 2025). Of course, these savings will only materialize if a sufficient number of operators adopt Nokia’s solution.
On the flip side, moving to Nvidia’s platform also means dependence on Nvidia for pricing and availability. Nokia loses some of the bargaining power that its own chip provided, and part of the savings may erode over time in component acquisition costs if demand remains tight. But in Nokia’s situation, economic realities have likely forced it to make a virtue of necessity. On the other hand, dependence on Nvidia could decrease in the long run if GPU-based platforms become the industry standard and multiple compatible suppliers emerge on the market. In that case, Nokia wouldn’t be tied to a single supplier in the same way, even though it would have abandoned its own chip development.
3) Utilizing the base station as a platform for AI computing and applications
a) Leasing capacity: In my view, this is clearly the more uncertain part of the AI-RAN business case. In principle, idle GPU capacity during quiet mobile network hours could be leased to hyperscalers or other heavy computation-dependent companies, e.g., for LLM inference, billed based on the number of processed tokens, but in principle also for other computing workloads. In practice, there are plenty of challenges: base stations rarely have extra space or power for heavy computing units. Furthermore, a distributed base station network, where connections and resources vary by location, is a much worse alternative for many general AI computing tasks than a hyperscaler’s own centralized data center.
b) Low-latency edge computing: Autonomous vehicles on the ground and in the air can benefit from computing close to the user. However, these are unlikely to generate massive volumes in the next few years.
c) A platform for third-party developers (dApps): Nokia is opening up the lower layer of the RAN via the new E3 interface to partners like Cohere, who could build new applications directly on top of the base station. The monetization mechanism is not yet public, and this resembles Core Software’s (formerly CNS) Network as Code initiative, about which monetization hasn’t been hyped up much. One would assume that the anticipated revenues would be shared between application developers, operators, and Nokia.
CONCLUSIONS
The strongest AI-RAN investment thesis may not be that Nokia turns millions of base stations into small data centers. Network efficiency improvements and Nokia’s opportunity to reduce its own R&D spending on chip design do not require a separate external market for computing services; rather, their economic justification can arise simply from the spread of AI-RAN in Nokia’s core RAN business. Selling computing capacity and applications to outsiders, on the other hand, is an additional future option, as it depends on whether there will ultimately be enough paying customers for them.
There are also risks to keep in mind: if AI-RAN were to truly double spectrum efficiency, it could cannibalize Nokia’s traditional capacity expansion business, which could equal several years of traffic growth at a 20–30 percent annual rate. One possible mitigation measure would be a transition to usage-based or recurring pricing, tied to actual capacity produced rather than one-off software sales. On the other hand, regarding competition, if GPU-based platforms eventually become the industry standard, Nokia’s early-achieved strong customer position could help mitigate the impact of software standardization.
And that Nokia is getting to play with others too, provenly, with a view to the inevitably coming breakdown of Nvidia’s monopoly, in my opinion. I somehow have a relaxed feeling knowing that Nokia keeps everything open to everyone, having learned from mobile to avoid all signs of vendor lock-in.
So I wouldn’t be surprised if the stock drops tomorrow in all its illogic.