Background image: 123rf

AI-RAN now has measurable radio gains, but the investment case changes sharply with where capacity is constrained. Software on the installed RAN has the clearest near-term economics, while accelerated platforms depend on paid utilization, falling inference prices and how long the hardware remains competitive.

AI-RAN is starting to produce numbers that an operator can put into an investment model. SoftBank and Ericsson recently tested AI link adaptation on SoftBank’s commercial 5G network and reported average gains of about 10% in spectral efficiency and downlink throughput across the evaluated locations. Ericsson says the feature runs on baseband hardware already in the network. Nokia is also reporting substantial gains from AI-assisted radio algorithms and is building a more ambitious architecture around accelerated computing. The operator decision now depends on where those gains produce cash and how much new hardware has to be bought to get them.

Let’s start with a cell that has just crossed the operator’s capacity threshold. Assume the operator normally upgrades when busy-hour load reaches 90% and the cell is now at 91%. AI improves spectral efficiency by 10%. The equivalent traffic now consumes about 83% of the enlarged capacity, so the upgrade can move out.

How far it moves depends on traffic growth. At 10% annual growth, that extra headroom lasts roughly a year. If the planned upgrade costs $200,000 and the operator’s cost of capital is 10%, moving the spending out by 1 year is worth about $18,000 today.

This is a more useful way to think about a 10% radio gain. On a lightly loaded cell it may produce no near-term cash benefit. On a cell sitting just above an investment trigger it can change the capital plan. The percentage is the same. The value is created by the location of the cell on its capacity curve.

Ericsson’s result makes this case interesting because it does not require dedicated AI hardware. It is sold as a software subscription and runs on Ericsson silicon already in the installed base. That gives operators a relatively simple first use of AI-RAN. Find the cells where the improvement changes a real investment decision and compare the value of those changes with the software cost.

I think the subscription model deserves more attention than it has received. Assume AI saves $18,000 of present value on one constrained cell but almost nothing on several lightly loaded cells nearby. Whether the software can be bought and deployed economically around the cells that create value becomes part of the return. Ericsson’s public material tells us that AI in RAN is a subscription, but not enough about how that subscription is priced and scoped to answer the question. I would want that information before extrapolating a network-wide return from the trial result.

Nokia says it has already produced more than 20% combined spectral-efficiency gains from AI-assisted algorithms on existing platforms. Those gains include different techniques such as link adaptation, channel estimation and carrier aggregation. Nokia is also moving the software onto accelerated merchant silicon, including NVIDIA Aerial, and says that architecture can support much larger gains over time. Its 50% target for 2027 and more than 100% for 2028 are roadmap targets rather than demonstrated commercial-network results.

This affects the investment case. If an operator is choosing between a conventional baseband that already runs useful AI algorithms and a more expensive accelerated platform, the accelerated hardware cannot be credited with every percentage point of AI improvement. The relevant number is the additional radio value produced by the accelerated option over the best software-enhanced alternative. We do not yet have a clean public comparison that establishes that number.

The accelerated platform can still make sense for another reason. It gives the operator general-purpose compute that can do other work when the RAN does not need it.

A normal RAN refresh is the easiest place to examine that idea. Assume an operator is already going to spend $20 million replacing baseband equipment. An accelerated alternative costs $25 million. The economic hurdle is the extra $5 million, plus any difference in power, operations and future hardware replacement.

The first workloads I would put against that premium belong to the operator itself. SoftBank has already run internal AI workloads alongside RAN processing on its AITRAS platform before extending the system to external users. If an operator can move suitable work from an existing cloud bill onto compute that would otherwise sit unused, there is an identifiable saving rather than a forecast of a future market.

To test how far that idea can go, I use a recent Northeastern University paper by Gabriele Gemmi, Michele Polese and Tommaso Melodia, A Techno-Economic Framework for Cost Modeling and Revenue Opportunities in Open and Programmable AI-RAN. The paper was accepted for ICCCN 2026 and builds a joint cost and revenue model for the same investment question. It estimates the GPU capacity left after RAN demand and asks whether LLM inference revenue can repay the extra cost of an accelerated platform.

The timing problem appears immediately in their model. Security systems, customer-service systems and many network applications do not conveniently wait until midnight. The model finds its largest conflict between RAN and LLM demand between roughly 14:00 and 18:00. Internal compute only creates additional value when its timing can fit around the RAN or enough capacity has been reserved to run both.

External workloads make the economics harder. A factory processing large volumes of local video may have a reason to buy nearby compute because moving all of the raw data elsewhere costs transport capacity or adds delay. Robotics and other physical AI applications can have similar requirements. Those are plausible edge customers, but they are not evidence for the 8× return claimed in the academic model.

The study models something different. It uses consumer LLM demand derived from BurstGPT traces and ChatGPT usage assumptions. It values the resulting tokens at a prevailing cloud inference price. The authors explicitly say they have not included any premium for running inference at the edge.

Commodity LLM inference gives the operator a large potential market, but the RAN location may provide little advantage over a purpose-built data centre. An industrial workload can place a higher value on location, but it will typically care more about guaranteed capacity. The operator has to decide which market its spare compute is really intended to serve before assigning a price to it.

The amount of paid utilisation required can become demanding quickly. Assume accelerated RAN costs $4 million more than the dedicated alternative and the operator wants to recover that premium over 5 years. At a 10% cost of capital, the annual capital charge is about $1.06 million.

With 200 GPUs there are about 1.75 million physical GPU-hours in a year. At $1 of contribution margin per paid GPU-hour, roughly 60% of every available hour would have to earn money just to cover the hardware premium. At 50 cents of contribution margin the requirement rises above 100%.

The RAN still needs some of those hours. Maintenance takes some. Capacity kept in reserve for service guarantees takes some. The business depends on paid utilisation of the hours that remain, not the headline utilisation of the hardware.

Pooling can increase the number of useful hours. SoftBank and Nokia are developing an architecture in which Nokia’s External Compute Engine can broker computing resources from multiple operators while SoftBank’s orchestrator assigns capacity across RAN, internal AI and external AI workloads. A larger geographic pool gives the scheduler more differences in traffic patterns to exploit.

The study then gives us a useful way to test how good that business could become. It compares NVIDIA Aerial with a FlexRAN alternative in a dense urban Milan deployment and calculates returns as high as 8.2× on the incremental GPU-platform investment. The headline deserves attention because the model shows that shared compute can plausibly pay for a much more expensive RAN platform under some conditions.

The core assumption is easy to explain without the notation used in the paper. Imagine that better models allow a GPU to perform twice as much useful inference next year. If the market price for that inference falls by half, the GPU still earns roughly the same amount per hour. Its productivity doubled while the unit price halved. Now let the GPU become twice as productive while the market price falls by a factor of 2.5. You are earning about 80% as much as before.

The paper does consider this relationship. When falling token prices are exactly offset by improving model efficiency, its overbuilt scenario returns 8.2×. When prices erode moderately faster, the return falls to 2.7×. Push the imbalance further and it falls to about 1.3×. At the paper’s most aggressive price erosion, the return drops to 0.51×. The authors describe this relationship as the decisive parameter in their results.

This also changes how I read the paper’s overprovisioning result. The model considers one network sized for current RAN demand and another that buys enough capacity on day 1 to handle RAN demand at the end of the 10-year period. The second approach creates much more compute to sell during the early years.

At the favourable price-efficiency assumption, the return rises from 6.7× in the launch-sized network to 8.2× in the overbuilt network. The underlying business does not depend on overbuilding, since the smaller deployment still produces 6.7×. The price that a GPU-hour can continue to command is much more important to whether either investment survives.

There is still a capital timing issue with the larger scenario. The paper compares $9.68 million of Aerial hardware with $5.87 million of FlexRAN hardware, both sized for the future requirement. The resulting $3.80 million difference is treated as the additional investment that AI revenue must repay.

A conventional operator could instead add FlexRAN capacity as traffic grows. Starting with the paper’s roughly $0.95 million launch requirement, growing capacity by its assumed 20% a year, and discounting the additions at 10% gives a present value of about $3.6 million. The effective funding difference against $9.68 million spent up front is then closer to $6.1 million.

Future AI revenue should be considered similarly. The paper’s favourable case starts at about $31,000 a week, rises toward $80,000 around year 5 and then declines as the RAN consumes more of the shared GPUs. Discounting a reconstruction of that cash-flow curve at 10%, while using the staged conventional investment, takes the return to roughly 3× rather than 8.2×. The exact figure depends on reconstructing the published curve, so I would treat 3× as an approximation rather than a new result. The economic point is that the favourable case can remain attractive after financing is included, but the headline multiple is not the return an investment committee would calculate from discounted cash flows.

Power needs the same treatment. The paper uses $0.04/kWh even though the Eurostat reference it cites does not support that number. Its current public software defaults to $0.20/kWh. A 1.2 kW server at the paper’s 1.5 PUE consumes 1.8 kW at the facility. If it ran continuously for a year, electricity would cost about $631 at 4 cents per kWh and about $3,154 at 20 cents. The right number is the operator’s actual contracted power cost, but the sensitivity belongs in the investment case.

The accelerator itself creates one final timing problem. RAN equipment is bought against a long network lifecycle. The market value of AI compute can change much faster. The paper holds the GPU platform through a 10-year horizon. Its most favourable revenue curve does not peak until about year 5. An operator should therefore test whether the same accelerator would still be competitive at that point or whether another hardware purchase is required.

I would keep that risk separate from falling token prices. Better models can allow an old GPU to perform more useful work, which helps. New GPUs can simultaneously become much better than the old one, which hurts its competitive position. An investment model that includes only the first effect will make a long-lived accelerator look safer than it may be.

If I were advising an operator today, I would sequence the investment around how much has to be assumed. I would start with AI software on the installed RAN where a measured improvement changes a real capacity decision. I would want to understand the subscription economics at those cells before scaling it across the network.

At a normal hardware refresh, I would price the accelerated option against the best AI-capable dedicated alternative. I would give credit first to radio improvements that can actually be attributed to the accelerated platform and then to internal AI workloads we already know can use the spare hours.

I would add external edge compute where there is an identifiable customer with a reason to pay for location and a service requirement the shared platform can meet. I would be more cautious about financing years of future RAN capacity on the assumption that commodity AI demand will keep the GPUs occupied at today’s economics.

The decision model I would take to an investment committee would contain very few speculative numbers. It would show which RAN upgrades move, what the software costs, what incremental radio gain requires accelerated hardware, which existing compute bills can be displaced, which external workloads are contracted, the hourly overlap between those workloads and the RAN, the power price, and when the accelerator is likely to need replacement.

The remaining unknown is the one I would watch closely. Operators are considering a long-lived network asset whose second source of revenue comes from a market where the price of compute is falling extremely quickly. If model efficiency improves at least as fast, shared AI-RAN compute can remain productive for a long time. If market prices fall faster, the revenue case can disappear even while GPU utilisation looks excellent. That is the assumption I would insist on seeing sensitised before recommending a large accelerated AI-RAN deployment.

Sources |

• Ericsson, SoftBank and Ericsson conduct Japan’s first trial of Ericsson AI in RAN on 5G commercial network

• Ericsson, The RAN gets smarter: Ericsson puts AI where it matters

• Nokia, The era of AI-native RAN starts now

• SoftBank, SoftBank and Nokia Demonstrate External AI Workloads on AI-RAN Infrastructure

• Gemmi, Polese and Melodia, “A Techno-Economic Framework for Cost Modeling and Revenue Opportunities in Open and Programmable AI-RAN”, accepted for ICCCN 2026, arXiv:2603.28680