NVIDIA Unlikely to Enter 6G RU Silicon, but Signals Possible Shift in RAN Architecture

Recent rumors have outlined that NVIDIA is exploring integrating purpose-built silicon for Radio Access Network (RAN) Layer-1 (Physical layer or PHY), due to the energy and computational overhead of calculating PHY in non-purpose-built silicon, in this case Graphics Processing Units (GPUs). If true, this will be pushing its strategy well beyond its current Artificial Intelligence (AI)-RAN and AI grid approach. Currently, NVIDIA’s AI-RAN strategy has been at the compute layer for the Distributed Units (DUs) and Central Units (CUs) and for non-Massive Multiple Input, Multiple Output (mMIMO) radios, while the mMIMO Radio Unit (RU) itself has traditionally been the untouched domain of Tier One infrastructure vendors, which have created custom, purpose-built Application-Specific Integrated Circuits (ASICs). This would represent NVIDIA’s first foothold in the hardware that transmits and receives radio signals.

This rumor likely stems from the fact that NVIDIA’s GPUs cannot handle high-order mMIMO PHY processing without making considerable compromises against custom silicon, certainly on energy consumption and heat dissipation. NVIDIA would likely need custom RU chipsets to address this challenge and it already has the partnerships to do so: it announced a strategic partnership with Marvell in March 2026, which is responsible for many custom RU chipsets, including Nokia’s ReefShark.

NVIDIA has not confirmed these plans to develop custom 6G RU silicon and has, in fact, denied them. Therefore, these reports are best understood as early stages of industry signaling, rather than defined strategy in NVIDIA’s telco roadmap. That said, the speculation may point to a latent need among operators for greater silicon choice and a wariness of replicating today’s lock-in with CUDA.

ABI Research does not expect NVIDIA to develop dedicated RU silicon, as this would go against its core model of software-defined platforms. However, if NVIDIA were to enter the RU market, it would challenge incumbent silicon vendors and the custom silicon model that has defined RAN hardware for decades as a programable alternative could reduce reliance on ASIC solutions.

Among vendors, Marvell is best positioned to adapt, given its existing relationship with NVIDIA and existing role in telco silicon. This points to a potential semi-custom model, rather than direct displacement. However, GPUs have not demonstrated the performance required for deployment at the radio level and efficiency remains a primary constraint as RUs are the most energy-intensive part of the network. This significantly limits the near-term viability of GPU-based RUs and in this context, any disruption would likely emerge gradually through hybrid architecture, rather than a direct replacement of existing RU silicon.

NVIDIA cannot manage the design and creation of custom RU silicon alone, it needs partners. Marvell can provide the design and integration and Intel/TSMC can manufacture the chipset, but NVIDIA will need an RU vendor to design a system around it. This would mean that NVIDIA forms partnerships with Ericsson, Nokia, and Samsung to create two parallel streams of products: one with vendor ASICs and one with NVIDIA custom silicon. However, it would not be in the interest of any of these vendors to cede control of their silicon roadmaps to NVIDIA, because this represents the most profitable part of their mobile network businesses.

Vendors and operators should not interpret these reports as an imminent disruption in the RAN market and ASIC-based RU roadmaps remain the most viable path into 6G, particularly given power and performance requirements. The key development to monitor is the evolution of the NVIDIA and Marvell relationship as a semi-custom approach that combines Marvell’s radio and Digital Signal Processor (DSP) expertise with NVIDIA’s compute platform, which is the most likely path as this concept further develops. Still, this would align with more long-term 6G timelines and deployments.

Therefore, operators should continue prioritizing flexibility at the DU and CU layers, where NVIDIA’s value proposition is already established, while maintaining optionality in RU vendor selection. NVIDIA’s position in the telecoms industry will likely deepen through these layers, rather than through direct entry into RU silicon, where custom ASICs remain dominant on power and cost grounds. Any shift toward custom RU and programmable architecture will emerge gradually, driven by proven AI-RAN economics and enterprise edge inference demand, though in the near term, it is unlikely to displace ASIC-based designs. However, this could change in the long term if new compute models such as token billing reduce their advantage.