NVIDIA Eyes GPUs for 6G Radio Units: A New Frontier for AI-Native Open RAN

On a windswept rooftop in a European test city last spring, engineers huddled around a prototype massive MIMO radio unit. The air hummed with the low whine of fans struggling to cool custom ASICs as beamforming algorithms strained to track dozens of users across a crowded spectrum. One engineer muttered about power draw spiking beyond projections. Across the ocean, in a Santa Clara lab, a different team was sketching something radical: replacing those fixed-function chips with programmable GPUs.

That contrast captures the moment NVIDIA finds itself in. Having already made inroads into the central and distributed units (CUs and DUs) of the RAN with its AI Aerial platform and Grace Hopper superchips, the company is now signaling that GPUs could become essential deeper in the stack—right inside the radio units (RUs) that sit closest to the antennas. For Open RAN operators eyeing the leap to AI-native 6G networks, this development could rewrite the rules of disaggregation, efficiency, and vendor flexibility.

From DU to RU: Extending the AI-RAN Playbook

NVIDIA’s AI-RAN strategy has long focused on software-defined RAN running on accelerated computing. Products like the Aerial RAN Computer (ARC) family, built on GB200 NVL2 and other platforms, already support Open RAN, vRAN, and cloud RAN deployments. Partners have demonstrated carrier-grade performance across 5G bands, with benchmarks showing strong reliability and the ability to run AI workloads alongside radio processing.

The new wrinkle targets the RU, where Layer 1 (PHY) processing for advanced massive MIMO and beamforming has traditionally relied on ASICs. In simpler 4G or basic 5G radios, low-PHY functions live comfortably in the DU. But as antenna counts climb—32x the compute demand in 5G Advanced versus legacy systems, and potentially 1,024 elements in ultra-MIMO 6G scenarios—the RU must handle sophisticated low-PHY tasks to avoid fronthaul bottlenecks and performance hits.

NVIDIA argues that programmable GPUs, optimized for constrained environments like those already proven in automotive and robotics (under 100W and high temperatures), can meet these demands while delivering the flexibility AI-native networks require. This aligns with broader industry shifts toward shared infrastructure for AI and RAN workloads, enabling new monetization at the edge.

Why Massive MIMO Changes the Equation

Massive MIMO is the linchpin. In these systems, beamforming and other low-PHY functions split between DU and RU. Conventional RUs embed ASICs for this slice of processing. Moving to general-purpose silicon opens the door to CUDA developers—already numbering in the millions—and NVIDIA’s open-sourced Aerial CUDA-accelerated RAN libraries.

For Open RAN, where standardized interfaces like O-RAN’s 7.2x aim to enable multi-vendor interoperability, a GPU-based RU could reduce the friction of pairing different silicon providers. It sidesteps the need for vendors to expose proprietary algorithms across interfaces, a barrier that has limited commercial multi-vendor massive MIMO deployments to date.

Intel’s approach, focused on CPUs like Granite Rapids for virtual RAN with lookaside accelerators, has avoided RU silicon entirely. Samsung continues to rely on homegrown ASICs or demonstrated full L1 on AMD CPUs in some cases. NVIDIA’s move, if realized, would give operators a unified, software-evolvable platform from DU to RU—potentially accelerating the shift from disaggregated Open RAN foundations to truly AI-native architectures.

Power, Economics, and the ASIC Dilemma

Skeptics rightly point to power consumption. RUs can account for up to 90% of a mobile network’s energy use. GPUs carry a reputation as power hogs, but NVIDIA emphasizes purpose-built embedded variants suited for telco constraints, not data-center monsters.

Economically, the case strengthens as RAN spend stagnates. Global operator RAN capex hovered around $35 billion recently, down from higher peaks, making custom ASIC development for niche telco volumes increasingly hard to justify. Programmable platforms that evolve with 6G standards and support concurrent AI applications become more attractive, especially when paired with partners like Marvell (which NVIDIA has backed and which is enhancing Octeon processors for GPU integration).

This could benefit Open RAN operators by lowering barriers to innovation, allowing smaller players or open-source efforts to contribute to RU-level intelligence without massive silicon investments.

Implications for Operators and the Open RAN Ecosystem

If GPUs enter the RU, Open RAN’s promise of multi-vendor, software-first networks gains new momentum. Operators could deploy unified platforms supporting both 5G workloads and emerging 6G AI features, with AI-driven optimizations running natively across the stack.

Challenges remain: real-world power efficiency at scale, thermal management on towers, and ensuring interoperability through O-RAN interfaces. Yet the trajectory points toward greater convergence of AI and RAN, building on existing AI-RAN Alliance momentum and collaborations with operators worldwide.

For a technical audience, the takeaway is clear—watch the RU. What once seemed off-limits to general-purpose computing is now squarely in NVIDIA’s sights, potentially unlocking the next phase of scalable, intelligent Open RAN.

Sources

• https://www.lightreading.com/6g/nvidia-has-a-radical-new-ai-ran-plan-a-6g-radio-unit-chip
• https://www.nvidia.com/en-us/industries/telecommunications/ai-ran/
• https://ai-ran.org/
• https://blogs.nvidia.com/blog/software-defined-ai-ran/
• https://nvidianews.nvidia.com/news/nvidia-and-telecom-industry-leaders-to-develop-ai-native-wireless-networks-for-6g