If you wanted to encapsulate Ericsson’s topline messaging on networks at its pre MWC 2026 event it would be this: AI will change networks. That creates both challenges and opportunities. And the response is to use AI to manage that.
How will AI change networks? In a variety of different ways. We’ll see an increase in human to machine, machine to machine, agent to human and agent to agent traffic. One effect of this will be to stress the Uplink Channel, with more traffic heading up into the network from the device or human at the end of it. A typical smartphone today might have a 90/10 split between downlink and uplink traffic. For a smartwatch uplink traffic is 30%. AR glasses might be up to 40% on the uplink. Another knock-on will be to tax latency budgets, with certain use cases – for example those smart glasses again – needing very tight latency.
Jenny Lindqvist, Head of Business Area Cloud Software and Services, said the ability for AI agents to reason and act across physical and digital environments represents a “profound shift”. Even though AI is not new in telco, the applications we are now talking about such as robots and drones have those new requirements in latency, uplink requirements, resiliency and reliability that will take networks far beyond today’s “best effort” model.
On the upside, this creates opportunities for operators. CTO Erik Ekudden said that we’re at an inflection point where a combination of distributed cloud, advanced connectivity and physical AI underpins the next wave of mobile innovation, creating new use cases and taking operators into new sectors. The industry is creating a digital platform as a full stack, making sure networks are open for innovation for anyone to add value on top via APIs. Managing all this, assuring the SLA-based business models that underpin the new business cases, will not be possible with manual processes. Networks must be AI-powered and then AI-native.
Ericsson’s strategy is to be the best provider of networks for AI, and also using AI, leveraging 5G-Advanced and AI powered 5G to enable the move into new sectors such as government, mission critical industries and defence. All of this is about leveraging data in the network. It starts with 5G SA, moves into AI powered 5G, and it wants to enable networks to reach hyperscale as platforms, offering services to developers in a consistent way.
So that means that the structure of the network will have to change. Meeting these AI-driven changes in an efficient and sustainable way will require AI itself. To meet demands for greater artificial intelligence, the network itself will have to become artificially intelligent.
The AI-ready RAN
As Per Narvinger, Head of Networks, Ericsson, said, the answer is to work out where this will happen. For example, training could take place in data centres, inference at the edge. At the far edge that means the cell site itself.
That’s part of the reason why Ericsson launched a bunch of new radio units that it marketed as “AI-ready”. How so? Because they contain a neural network accelerator. And what is that? It’s a programmable matrix core within its beam forming chip for massive MIMO radios which can carry out matrix computations for RAN processing.
By leveraging this chip which is being used for Layer 1 beam forming in massive MIMO, Ericsson found it can also service other AI use cases. One early option it is pursuing is an AI powered receiver where the AI will do base channel estimation using the capability of the hardware.
Narvinger said, “So it turns out that the Ericsson silicon that we have been developing over years, by just putting in some neural network accelerators there, we get the silicon that is very, very well suited for running AI workloads.
“There are a lot of tasks in the radio, for example beam steering, channel estimation, trying to figure out conditions between the user and the antenna, that are very well suited for AI models.”
“When we talk about these neural network accelerators, within our Ericsson ASICs we add technology blocks so it gets more optimised for AI, more optimised for general matrix multiplication. And then by just adding some accelerators and adding a bit more memory, you can suddenly handle AI models in a very efficient way.”
Narvinger sees great potential for AI to increase performance in the RAN, even where there are highly optimised algorithms already in operation. One example he likes is a link adaptation model where Ericsson achieved a 10% increase in performance just by applying a small AI model to an algorithm that has been optimised by the industry for the past 30 years.
We have disaggregated the hardware and software, so we don’t need to make that choice today, to be honest
AI-RAN
Building a neural network accelerator to bring forward the AI-powered radio is not the same as developing an AI-RAN, as it is commonly understood in the industry. AI-RAN as a concept has mainly focussed on the use of GPUs to run baseband software and also AI workloads on a common architecture. If you are going to talk about AI for networks and networks for AI, where does the AI-RAN come into play?
For Narvinger, that calculation should be considered in terms of the overall debate about where best to place compute and intelligence in the network. For example, the link adaptation model doesn’t need a GPU, you can train it on a CPU on the existing baseband.
“With Ericsson silicon, we have synergies between the silicon that we put in the radio and what you put in the baseband compute. But as you know for baseband compute there are a lot of contenders to host the RAN applications, whether that be Nvidia and GPUs, or AMD, Qualcomm or Nokia.
“I think when it comes to the training of AI models, GPUs and Nivida are in pole position. But when it comes to inference or execution of the models, there are many players that are trying to provide a more optimised solution for that.
“With our customers we will figure out where we should be putting those workloads over time. But we have disaggregated the hardware and software, so we don’t need to make that choice today, to be honest.”
That said, and despite an Ericsson white paper last year which was ambivalent about GPU acceleration for Cloud RAN, Narvinger says, “You can assume there will be more announcements in that area, what is working on what. But it’s true that for commercial networks what we’re really running on is purpose built and x86. We have two big customers in the US leveraging Intel, a lot of other customers are using purpose built RAN.
“It’s quite early in the cycle. We are getting AI models out in the network this year, that’s what you will see in the roadmap.”
T-Mobile, one of the more AI-RAN forward operators, is “looking at the gains” Narvinger says, but it is not rolling out GPUs in the network now for RAN workloads.
“They will evaluate what’s best for them. If they say GPUs we will put our RAN software on the GPUs.”
Ericsson has said for years that its Cloud RAN strategy is to enable portability between platforms, by truly disaggregating its software from the underlying hardware. It reiterated that message at the briefing day. If customers see value in a GPU accelerated RAN, then the vendor stands ready. But it is fair to say it is not pushing for it.
Ericsson claims its hardware/software disaggregation means that it can or will be able to implement its RAN software on any underlying major hardware platform, using common software interfaces.
As for Nokia, which has received a billion dollar investment for Nvidia, which many think makes it more aligned with the AI-RAN GPU model, Narvinger said, “I think the customers are more trying to figure out what is the most TCO optimised way of building a network in the future. Then I have a competitor that seems to put themselves in one very firm position, but I think they would have to comment whether that is successful or long term.”
rApp as a Service
Ericsson’s other main launch at the event was of its rApp as a Service on AWS compute and cloud.
The two companies have developed a blueprint for making rApps available as a service from within AWS marketplace. The apps are connected to the Ericsson RIC via the R1 interface, and in theory to any other RIC over that same open interface.
Jean-Christope Laneri, Vice President & Head of Cognitive Network Solutions, said that the driver for this is to deliver a multi cloud product from Ericsson for customers that want to move to public cloud.
The blueprint leverages the AI services stack from AWS for lifecycle and for the agentic pieces that make different rApps collaborate together – with that reasoning capability coming from AWS.
Laneri said Ericsson is seeing a few customers embarking on the move to public cloud as part of their overall netops transformation, appreciating the elasticity and ease of operations.
At its core, the AWS implementation offers Ericsson another route to market for its RAN optimisation apps.
Laneri said that while Ericsson collaborates with all the hyperscalers on its infrastructure as a service and on-demand core network offering, “this is special because it is true as a service where we leverage native services from AWS to build the product.”
More AI in the network edge – the UPF
Elsewhere, Jenny Lindqvist said that the company is exploring the colocation of the core network UPF with AI inference at the edge.
“In our labs we are running our user plane function (UPF) together with Cloud RAN on a single server, and then in addition communicating with an AI application. Needless to say, that’s a very footprint efficient way to deliver those type of workloads very far in the edge.”
Lindqvist said co-locating the local packet gateway with the AI application, and potentially also RAN software, would enable the management of SLA-assured slices near the breakout point.
“In addition, we have the opportunity to talk to slicing capabilities for specific AI workloads to be then handled at the edge versus the more traditional non-prioritised traffic.”
