Nokia's GPU Gamble: What the AI-RAN Hardware Shift Means for Open RAN Operators
- October 12, 2026
- 8 mins
- Technology
- ai ran airscale gpus nokia open ran ran economics telecom
Nokia’s GPU Gamble: What the AI-RAN Hardware Shift Means for Open RAN Operators
Nokia’s reported plan to start shipping GPUs inside its AirScale radio access network chassis is the kind of news that makes procurement teams reach for their spreadsheets and strategy teams reach for their whiteboards. According to Light Reading, the Finnish vendor has been advertising a slimline GPU that can slot into an existing AirScale chassis originally designed for Marvell Technology’s silicon — with the same power constraints and, executives promise, the same price. It is a high-stakes bet that the economics of AI-RAN can be made to work inside the footprint operators already own.
For a skeptical executive, the announcement raises more questions than it answers. Below, we work through the ones that matter most, grounded in what Nokia, Ericsson, and the broader vendor ecosystem have actually said.
Why would Nokia put GPUs in a base station chassis at all?
Because the alternative — asking operators to rip out and replace their radio sites — is a non-starter. Nokia’s pitch, as reported by Light Reading, is that the GPU is a drop-in component for an existing AirScale chassis. That means the same power envelope, the same physical form factor, and, critically, the same price as the Marvell silicon it is designed to coexist with.
The strategic logic is straightforward. Nokia CEO Justin Hotard framed it in terms of capital efficiency: “What we anticipate with AI-RAN is that it gives customers an opportunity to reduce the ongoing hardware investment, or to elongate that,” he said. “Then they can get the benefit of ongoing enhancements once they deploy AI-RAN software.”
That is a software-first argument dressed in hardware clothing. If the GPU can be upgraded through software and support new AI algorithms for years — unlike an ASIC, which is fixed at manufacture — then the operator’s investment is protected against the rapid obsolescence that has historically plagued RAN hardware.
Is this actually a departure from Nokia’s earlier AI-RAN messaging?
Not a departure, but a refinement. Nokia has spent much of 2026 positioning itself as an AI-RAN leader with a merchant-silicon-friendly, software-first story. The company has already launched what it describes as an industry-first commercial AI-RAN platform, expanded field trials to eight operators, and partnered with NVIDIA on commercial AI-native RAN. The GPU-in-AirScale plan is the hardware corollary to that software narrative.
The nuance matters for operators. Nokia is not saying every base station needs a GPU tomorrow. It is saying that when AI-RAN workloads justify acceleration, the chassis can accept it without a forklift upgrade. That is a materially different proposition from vendors selling AI-RAN as a greenfield architecture.
What does the physics say about how far this can go?
This is where skepticism is warranted. As Ericsson’s head of RAN compute, Michael Begley, explained at MWC Barcelona, the size of a model or the complexity of an algorithm will always be constrained by the transmission time interval. The TTI imposes a hard limit on how long a base station has to perform critical tasks. Any calculation that misses the TTI is worthless because radio conditions will have changed.
In other words, not every AI workload belongs in the RAN. Some are latency-tolerant and can live in the core or the edge cloud. Others — beamforming, scheduling, interference management — are hard real-time and must execute at the radio site. The GPU-in-chassis approach only makes sense for that second category, and only if the power and thermal budget can absorb it.
Nokia’s promise of parity on power and price is therefore the crux of the bet. If it holds, the addressable market for AI-RAN acceleration expands dramatically. If it doesn’t, operators will treat the GPU as a niche option for high-value sites rather than a default configuration.
How does this compare with what Ericsson is doing?
Ericsson has taken a different tack, emphasizing RAN compute platforms and software subscriptions. Its AI-in-RAN subscription launch and its RAN compute platform landing at NTT Docomo suggest a strategy built around selling intelligence as a service layer on top of existing hardware, rather than pushing new silicon into the radio site.
That divergence is healthy for the market. It gives operators a genuine architectural choice: accelerate at the edge with GPUs, or centralize intelligence and keep the radio site lean. The right answer will vary by spectrum position, site density, and service mix. Operators in dense urban markets with high cell counts — think India, where Jio has highlighted population densities of 25,000 per square kilometre in Mumbai and up to 100,000 per square kilometre during major religious events — may find edge acceleration more compelling than operators with more distributed footprints.
What is the Open RAN angle here?
Open RAN’s promise has always been about disaggregation and choice. A GPU that slots into a proprietary AirScale chassis is, on its face, a proprietary move. But the broader AI-RAN ecosystem is moving toward open interfaces that could make acceleration more portable.
The E3 interface, championed by the AI-RAN Alliance and vendors like Cohere, is designed to let third-party dApps access RAN data and compute resources. If Nokia’s GPU supports those interfaces, the hardware becomes a platform rather than a lock-in. If it doesn’t, operators should ask hard questions about portability before they commit.
There is also the question of merchant silicon. Nokia’s willingness to build around Marvell and now GPUs suggests a pragmatic, multi-silicon strategy rather than a single-vendor bet. That is consistent with the merchant-silicon confidence Nokia has signaled in its recent AI-RAN messaging.
Will operators actually pay for this?
The honest answer is that nobody knows yet, and the vendors are hedging accordingly. Nokia’s Hotard framed the value as reducing or elongating hardware investment — language that acknowledges operators are not looking for new capex lines. They are looking for ways to sweat existing assets harder.
The counterargument comes from operators themselves. Deutsche Telekom has publicly pushed back on the idea that AI traffic will inflate RAN capex, arguing that uplink reality and traffic patterns don’t justify it. If that view prevails, AI-RAN acceleration becomes a targeted tool for specific high-value use cases rather than a blanket upgrade.
The middle ground is probably where the market lands. Operators will deploy GPU-accelerated AI-RAN at sites where spectral efficiency gains translate directly into revenue or cost avoidance — dense urban hotspots, enterprise campuses, venues — and leave the rest of the estate on conventional silicon until the software roadmap justifies broader deployment.
What should a skeptical exec ask Nokia in the next meeting?
Five questions worth putting on the table:
- What is the actual power delta? Parity on paper is one thing; sustained thermal performance in a live site is another. Ask for field data, not lab data.
- What is the software roadmap commitment? The value proposition depends on years of algorithm upgrades. Get that in writing, with performance milestones.
- Which interfaces are supported? If E3 or equivalent open interfaces are not on the roadmap, the GPU is a proprietary accelerator, not an open platform.
- What happens to the Marvell silicon? Coexistence is fine, but operators need to understand whether the GPU is additive or a migration path that strands existing investments.
- What is the total cost of ownership over five years? Compare GPU-accelerated AI-RAN against a conventional refresh plus centralized AI. The answer will differ by market, but the model should be transparent.
Does this change the Open RAN competitive landscape?
It raises the stakes. If Nokia can deliver AI-RAN acceleration inside an existing chassis at parity on power and price, it strengthens the case for its end-to-end proposition against both traditional RAN vendors and Open RAN challengers. If it stumbles on execution, it opens the door for vendors offering centralized AI-RAN architectures and for open-source stacks that decouple acceleration from the radio vendor.
The broader trend is unmistakable: AI-RAN is moving from concept to commercial reality. The AI-RAN Alliance now counts 132 members and 33 demos. Dell’Oro projects $35 billion in cumulative AI-RAN revenue by 2030. Nokia’s GPU gamble is one of the clearest signals yet that the hardware supply chain is repositioning for that future.
The bottom line
Nokia’s plan to ship GPUs in AirScale chassis is a pragmatic bet that AI-RAN economics can be made to work within the constraints operators already live with — power, space, and capital discipline. It is not a revolution; it is an option. And options are exactly what skeptical executives should want.
The question is not whether AI-RAN will matter. It is whether the hardware to deliver it can be deployed without breaking the business case. Nokia is betting it can. Operators should hold them to it — with data, contracts, and a clear-eyed view of where acceleration pays and where it doesn’t.