If you invest in $NOK you need to read this.
If you are interest in robotics, Phyiscal AI, edge AI. You need to read this.
But before you do that, please 'bookmark' and share.
On $NVDA call today:
“In the future, every single base station, every single radio network would become an AI-powered radio network.”
I just posted a substack article going over the implications of this statement..
You can read it here:
https://cruxcapitalgroup.substack.com/p/nvidia-just-told-us-something-important?r=6so16n
But let's take a step back first...
I wrote an article on AI RAN 6 weeks ago and I want to share it fully here now.
Enjoy!
When we think of AI Infrastructure we tend to think of giant data centers full of GPUs, hyperscalers spending tens of billions of dollars, a race to add more power, more fiber, and more capacity inside and between campuses.
And we are right! But, that’s not the entire picture.
AI is pulling intelligence closer to where data is created and where decisions need to happen fast. Cameras, robots, factories, connected machines, autonomous systems, live video, the physical world more broadly. Once that happens, the network around the data center becomes more consequential. And once that network becomes more consequential, the telecom footprint start to look like a possible compute surface.
One of the things I have been spending real time on lately sits outside the usual hyperscaler and optics conversation but could become a meaningful contributor to the AI buildout over time. It sits inside telecom infrastructure, touches the radio layer, and increasingly connects back into the broader optical, transport, and compute story as AI becomes more distributed.
It is also one reason I think Nokia is a really interesting play. Most coverage focuses on telecom, carrier budgets, restructuring, and just recently optical upside after Infinera. There is another layer forming here, and while it is still early, the proof points are getting more concrete quickly.
What I want to do in this piece is break down what this actually is, what is commercially grounded today versus what is still architectural, how recent operator tests changed my view of the maturity curve, and why the optical and transport side of the story may carry just as much weight as the radio side in the near term.
What am I talking about?
The Radio Access Network (RAN) is the part of the cellular system that connects your phone or device to the broader network. Radios, cell sites, baseband functions, and the software managing those wireless connections all live here. It is also one of the most physically distributed pieces of infrastructure in the world. Telecom operators already have sites spread across cities, industrial corridors, suburban footprints, and remote areas, with power, transport, hardware, and operations teams attached to those places.
Bringing AI into that footprint is a very different proposition from building another centralized AI campus. One model concentrates compute into a few giant locations. The other tries to make a distributed network smarter, more adaptive, and eventually more monetizable. That distinction is the core of what AI-RAN is trying to do.
Three ideas
There are really three bits we are talking about here. They are related but they have different timelines and different investment implications.
The first is AI for RAN. Using AI to improve how the radio network operates. So things like traffic optimization, energy management, better scheduling, faster issue detection, better use of spectrum, and more autonomous operations. Software making a complicated wireless network run better with less manual work. The incentive is already there because these networks are expensive, operationally heavy, and increasingly stressed by traffic growth. This is the most commercially grounded part of the story today, and the easiest for operators to justify because the value proposition is straightforwar with lower costs, better performance, and less day-to-day manual intervention.
The second is AI and RAN. Instead of treating radio workloads and AI workloads as separate universes, this approach puts them on the same underlying compute platform. Telecom sites already have distributed infrastructure. If those sites can handle both wireless functions and AI workloads simultaneously, the network footprint itself becomes more strategically valuable. This is where the NVIDIA angle becomes relevant, and where the proof points are starting to accumulate. The central idea is that the same physical telecom site can begin doing two jobs at once like running the mobile network and running AI compute. That is a fundamentally different way of thinking about what a tower site is worth.
The third is AI on RAN. This is where the telecom edge becomes a place where actual AI applications run like machine vision, robotics, industrial automation, real-time video inference, physical AI, and low-latency services that benefit from being processed closer to where the data originates. This is the version that sounds the largest and probably is the largest if it matures. It also carries the least near-term revenue visibility of the three. This is where the telecom network starts to look less like communications infrastructure and more like an application platform. If it develops the way bulls hope, operators are selling access to local compute close to the physical world alongside connectivity which is a fundamentally expanded business model built on the same physical footprint.
All three are worth understanding.
Why Nokia belongs in this story
Nokia is approaching AI-RAN from the inside. It already has the radio stack, the operator relationships, and the installed infrastructure base that give it a credible path into the category. AI-RAN gets adopted through existing vendor relationships, gradual software enablement, forward-compatible hardware, and operator trust and Nokia already sits inside all of those conditions.
Let’s hear it from the horses mouth.
The AirScale comment from Hotard is one of the most important lines in the whole story.
“If you buy an AirScale platform today, it’s going to be upgradable to AI RAN as we launch that platform. And so that’s the kind of opportunity where making the investment decision now and having clarity now as an operator, we think is particularly important.”
Operators are reluctant to spend heavily on today’s radio platform if a full replacement is required a few years later. Nokia is telling them the transition can happen gradually, which makes experimentation more realistic and lowers the friction around deployment. NVIDIA’s Ronnie Vasishta framed the broader shift in similar terms:
“Instead of upgrading networks in big, hardware-heavy cycles, we now have the opportunity to build them as fully software-driven systems. By running AI and radio access networks on the same accelerated computing platform, we make sure the network supports business needs, not the other way around.”
Hotard has also described where the category stands commercially:
“AI-RAN transforms RAN into a software-driven platform optimized for AI, and with NVIDIA and a growing ecosystem of partners we are progressing from validation to commercial deployment.”
Nokia already has the installed position, the product roadmap, and the operator dialogue to move this from concept into something more commercially durable over time.
Why this goes beyond the radio layer
If AI workloads become more distributed, if telecom sites begin carrying more intelligence, and if the network edge starts behaving more like a compute surface, the surrounding network becomes more consequential too. Transport, routing, optical capacity, and the physical ability to move more data between the edge, the cloud, and everything in between all get pulled into that shift. AI-RAN begins at the radio layer and the architecture around it follows.
Nokia’s David Heard at OFC:
“especially hyperscalers and neocloud players and even now out in the wide area network with service providers and mission-critical enterprise, they are buying road map because they are making plans, they’re buying data centers. They’re buying facilities. They’re planning HVAC right now.”
Hotard tied the optical and IP buildout to the same dynamic on the Q4 call:
“these are no longer the cloud computing systems that were built over the last 10 to 15 years. These are AI supercomputers, and AI supercomputers need higher bandwidth, richer connectivity. And we’re now seeing optical technology go into those and integrating and networking.”
Rob Shore, Nokia’s head of optical marketing, described the shift in how customers are thinking about optical innovation:
“Historically we’ve been focused for 30-plus years in the industry on building engines specifically focused on maximizing capacity per fiber. This generation is the first generation where we’ve really shifted. They want more cost-based and power-optimized solutions.”
That is the context for why Nokia’s broader network exposure belongs in this piece. A more distributed AI architecture requires the transport and optical infrastructure capable of supporting more distributed intelligence. The radio layer and the network underlay are being pulled forward together.
What is investable?
Separating nearer-term from longer-dated is the cleanest way to frame this.
The most commercially grounded part of the theme today is AI for RAN. Smarter operations, lower manual burden, better optimization, digital twins, and software that helps operators run their networks more efficiently. Hotard gave one of the clearest proof points on the Q4 call:
“We launched two new products in the quarter, including our agentic AI solution for event-driven automation management, which reduces network downtime by 96%.”
The economic value is direct and operators can justify it immediately. Better network performance and lower operating complexity with a clear line to cost savings.
AI and RAN is the next layer. Shared infrastructure, live operator tests, and gradual deployment through forward-compatible platforms are all making the story more credible. One T-Mobile test is the clearest proof point:
“The trial demonstrated concurrent AI and RAN processing on a single NVIDIA Grace Hopper 200 server using accelerated AI-RAN workloads, highlighting the ability to combine advanced radio access network functions with AI applications on a shared accelerated computing platform.”
The Indosat result added another live-environment confirmation:
“This milestone proves that AI and RAN workloads can run simultaneously on shared GPU infrastructure in a live operator environment, paving the way for distributed AI intelligence that makes 5G networks more efficient and intelligent and sustainable.”
This is still a developing category rather than a fully scaled revenue engine, though the proof points are accumulating faster than most expected.
AI on RAN is the longest-dated upside where the telecom edge becomes a true application surface for physical AI, machine vision, robotics, industrial automation, and low-latency inference. Nokia and SoftBank have already demonstrated one version of the monetization logic:
“Nokia and SoftBank demonstrated how spare AI-RAN compute capacity can be used to run third-party AI tasks. This integration marks a key step in transforming the RAN into an AI-enabled platform capable of delivering new AI services and revenue streams beyond connectivity.”
Elisa’s COO Sami Komulainen framed the longer arc well:
“AI-RAN is a key enabler for optimizing end-to-end network performance, enhancing service quality, and advancing toward AI-native 6G, as well as future agentic, robotic, and ultimately physical AI.”
Nokia gives us exposure to a theme with real proof points, a credible installed-base bridge, and enough supporting infrastructure to become economically meaningful if the architecture keeps moving in this direction. The nearer-term upside sits in optical, IP, and AI for RAN software. The longer-dated option sits in what the telecom footprint could become. Both are worth holding onto as the story develops.
The information provided is for informational purposes only and does not constitute investment advice, a recommendation, or an offer to buy or sell any securities. The author may hold a position in the securities mentioned. Readers should conduct their own due diligence and consult with a financial advisor before making investment decisions.
