Nokia’s GPU Gamble: What Dealing Out GPUs in AI-RAN Means for Open RAN Operators

Nokia is reportedly preparing to ship GPUs inside its AirScale baseband chassis, according to Light Reading. The move, described as a “high-stakes AI-RAN gamble,” would see Nokia slot a slimline GPU into hardware originally designed for Marvell silicon. The GPU must operate within the same power envelope and at the same price point, while supporting software upgrades and new AI algorithms for years—unlike fixed-function ASICs. CEO Justin Hotard framed the strategy as a way to help operators “reduce the ongoing hardware investment, or to elongate that,” with benefits accruing once AI-RAN software is deployed.

This is not just another vendor announcement. It is a direct challenge to the assumption that AI in the RAN requires a rip-and-replace cycle. For Open RAN operators, it raises urgent questions about architecture, procurement, and the economics of intelligence at the edge. Here are five things this means for operators.

1. The Hardware Refresh Cycle Gets a Software Escape Hatch

For years, operators have been told that AI-RAN would demand new silicon—accelerators, purpose-built ASICs, and eventually 6G-ready radios. Nokia’s GPU-in-AirScale approach offers a different path: reuse existing chassis, swap in a GPU, and unlock AI capabilities through software.

That matters because RAN capex is under pressure. Ericsson’s Börje Ekudden recently noted that a large part of the vendor’s equipment shipped in 2025 and 2026 is already upgradeable for AI features and can be enhanced through software. Nokia’s move doubles down on that narrative: AI becomes a software subscription, not a hardware forklift.

But the constraints are real. The GPU must fit within the same power and thermal limits as the Marvell silicon it replaces. It must hit the same price. And it must support future algorithms without a hardware change. If Nokia can deliver on those promises, operators gain a rare thing: a hedge against premature hardware obsolescence.

2. The TTI Is Still the Hard Wall

AI in the RAN is not like AI in the cloud. As Ericsson’s Michael Begley explained at MWC Barcelona, the transmission time interval imposes a hard limit on how long a base station can take to perform critical tasks. Any calculation that misses the TTI is worthless because radio conditions will have changed.

That means AI-RAN cannot simply run large language models at the edge. It requires smaller, specialized models designed for specific tasks—beamforming, scheduling, interference management—and they must run deterministically within milliseconds.

Nokia’s GPU gamble only pays off if the GPU can meet those real-time constraints while sharing power and space with existing baseband functions. The company’s claim that the GPU will support “new AI algorithms for years” implies a software-defined approach to RAN intelligence. But operators should press for benchmarks: what latency, what throughput, what spectral efficiency gains, and under what load?

3. Vendor Lock-In Risk Shifts from Silicon to Software

Open RAN’s promise is multi-vendor interoperability. But if AI-RAN intelligence runs on a proprietary GPU inside a proprietary chassis, controlled by proprietary software, the lock-in risk simply moves up the stack.

Nokia is not alone in this. Ericsson has launched AI-in-RAN subscription services. Samsung and KDDI have demonstrated AI-driven RAN gains on live 5G. SK Telecom is building a national AI-RAN test network. The common thread: vendors are racing to own the AI layer, even as they embrace open interfaces at the lower layers.

For operators, the lesson is to demand open APIs and standardized interfaces for AI applications—what the industry calls dApps or rApps. The O-RAN Alliance’s E3 interface, for example, is designed to let third-party AI applications control RAN elements. If Nokia’s GPU supports such interfaces, operators retain choice. If it does not, they trade one form of lock-in for another.

4. The Business Case Depends on Uplink, Not Downlink

Much of the AI-RAN hype focuses on downlink throughput and spectral efficiency. But the real constraint may be uplink. As one recent analysis noted, the 4GHz AI case has a way to go because uplink capacity is limited by physics—device power, antenna size, and spectrum.

AI applications that run on the RAN—predictive scheduling, anomaly detection, energy optimization—may not require massive uplink bandwidth. But AI applications that serve customers—video analytics, AR/VR, autonomous vehicles—will. Operators need to model both sides of the equation.

Nokia’s GPU could help by processing uplink-heavy AI workloads closer to the edge, reducing backhaul and core costs. But the business case will vary by spectrum holdings, cell density, and customer mix. There is no one-size-fits-all ROI.

5. The Competitive Landscape Just Got More Interesting

Nokia’s move puts pressure on Ericsson, which has its own AI-RAN ambitions but has emphasized software upgrades to existing hardware. It also challenges new entrants like Arm, which is pushing into vRAN with Kenyi, and Nvidia, which is eager to see GPUs in radio units.

For Open RAN operators, the upside is choice. If Nokia, Ericsson, and others all offer GPU-accelerated AI-RAN, operators can benchmark performance, power, and price. The downside is complexity: more options mean more integration work, more testing, and more vendor management.

The AI-RAN Alliance, now with 132 members and 33 demos, is trying to harmonize these efforts. But standards move slower than product cycles. Operators should engage with both.

What to Watch

Nokia’s GPU gamble is a bet that operators want AI-RAN without a hardware revolution. If it works, it could accelerate AI adoption in the RAN and extend the life of existing baseband investments. If it fails—due to power, price, or performance—it could reinforce skepticism about AI-RAN hype.

Either way, the conversation is shifting from “if” to “how.” Operators should ask vendors for concrete TTI-compliant benchmarks, open interfaces for AI apps, and clear upgrade paths. The winners will be those who can deliver intelligence without breaking the economics of the RAN.

Sources

• Nokia to start dealing out GPUs in high-stakes AI-RAN gamble
• AI can enhance 5G network performance by 30%: Ericsson
• The week in AI-native telco (3-9 Oct 2026)