AI-RAN and 6G Infrastructure: Moving to Software-Defined Radio
ExtremeSpec is reader-supported: some links on this page are affiliate links, and as an Amazon Associate ExtremeSpec earns from qualifying purchases at no extra cost to you. Learn more.
AI-RAN and 6G infrastructure allow telecommunications operators to run AI and radio functions on a shared software-defined platform. This architecture holds as long as software-defined radio achieves performance-per-watt parity with proprietary silicon. Telecommunications network architects and 6G research scientists should act on this, while consumer electronics hobbyists can skip it.
Who should act on AI-RAN and 6G Infrastructure and who should skip it
Readers who should act on AI-RAN and 6G Infrastructure
- Telecommunications Network Architects should transition from proprietary hardware to software-defined AI-RAN because software-defined radio has reached parity with custom silicon.
- 6G Research Scientists should adopt NVIDIA’s Sionna and Aerial libraries to access a full simulation and over-the-air testing environment.
- Infrastructure Project Managers should evaluate the 100 gigawatts of power available in existing telecommunications infrastructure for AI compute.
- Network Operations Engineers should implement agentic flows like the AI Telco Engineer to achieve a 3.6x latency improvement in network troubleshooting.
Readers who should skip AI-RAN and 6G Infrastructure
- Consumer Electronics Hobbyists should skip this topic and focus on general PC hardware reviews as this infrastructure is industrial in scale.
- Mobile App Developers should skip this topic and refer to mobile development documentation, as this focuses on RAN architecture and 6G physics.
What to do next about AI-RAN and 6G Infrastructure
- Download the Sionna library from GitHub to perform differentiable link-layer simulations for 6G research.
- Join the 6G Developer Program to access shared over-the-air testing environments with over 8,000 other members.
- Evaluate the MGX 2U short for base station deployment to verify if the RTX Pro 4500 supports your required cell and frequency capacity.
- Implement the AI Telco Engineer agentic flow to achieve a 3.6x latency improvement in equalizer optimization.
The Shift from Proprietary Silicon to Software-Defined AI-RAN
AI-RAN represents a move away from traditional radio access networks that rely on proprietary equipment built on custom silicon. Previous systems required specific hardware for every function, but software-defined radio now allows AI software to run on the same infrastructure as the radio access network. Think of software-defined radio like a universal translator for hardware, decoupling the intelligence of the network from the specific brand of silicon it runs on.
What is AI-RAN?
AI-RAN is an AI-enabled Radio Access Network that uses software-defined architecture to run AI and radio functions on the same infrastructure. This allows operators to use the same AI stack for both network operations and the radio access network, reducing the need for specialized silicon.
NVIDIA identifies the tipping point for this transition as the moment software-defined radio achieved performance-per-watt parity with proprietary systems. Why did it take so long to reach this point? The transition required eight years of dedicated CUDA kernel exploration and software development to optimize how GPUs handle signal processing. This journey established the foundation for an AI-native stack where innovation can occur at every layer of the software stack.
Tapping into 100 Gigawatts of Distributed Power
Telecommunications infrastructure currently contains roughly 100 gigawatts of power that can be leveraged in the intelligence era. This power is distributed across 100,000 sites, including mobile switching offices, mobile exchanges, and points of presence. Because large land power assets are becoming harder to find, these existing distributed assets are becoming a primary factor for AI compute.
NVIDIA suggests that the 100,000 sites mentioned by Jensen Huang represent a distributed infrastructure capable of leveraging 100 gigawatts of power, which means the infrastructure is designed to support a wide distributed compute footprint. Using this existing power avoids the need to build new energy infrastructure for AI factories.
The Hardware Foundation: MGX and DGX Spark
NVIDIA and its partners have developed the Aerial RAN computer family to enable the deployment of radio access networks across the world. These systems allow GPUs to sit at a base station, addressing concerns about power availability and cost at the edge. The family includes several models tailored for different roles in the infrastructure.
NVIDIA has not announced specific pricing for the MGX 4U or DGX Spark systems; check with Super Micro and NVIDIA for current quotes. The following table identifies the primary hardware members of the Aerial RAN computer family:
The Aerial RAN Computer Family
| Model | Manufacturer | Key Feature |
|---|---|---|
| MGX 4U | Super Micro | CRAN Support |
| MGX 2U | Dell | Distributed Cloud |
| MGX 2U short | Quanta | Contains RTX Pro 4500 |
| DGX Spark | NVIDIA | 6 cells 100 megahertz 4T4R |
Democratizing 6G Research with Open-Source Tools
NVIDIA created the Sionna library to democratize 6G research by providing a link-layer simulator that is differentiable. Researchers can use Sionna to simulate and test wireless algorithms in a full simulation environment. By placing the library on GitHub, NVIDIA has enabled over a million downloads from the corporate and academic research communities.
What is a link-layer simulator?
A link-layer simulator is a tool used to simulate network connections at the fundamental layer of the OSI model. It allows researchers to test how data moves across a wireless medium without needing to build physical hardware for every test.
The 6G Developer Program currently includes over 8,000 members and provides a shared environment for over-the-air testing. These tools allow researchers to move from theoretical papers to real hardware testing. Combined with the Aerial software stack—a 5G stack accelerated on GPU—these tools provide a complete system for 6G innovation.
Orange’s Nemotron speech-to-text pilot provides real-time translation for emergency services
Orange has demonstrated a Nemotron speech-to-text pilot that provides real-time translation in emergency services. This application uses a full-duplex model that recognizes tone and semantics, providing a latency-sensitive solution for firefighters. The pilot demonstrates how AI can be integrated into live intelligence platforms at the edge.
Cisco and BT are using Nemotron post-trained on site data to automate network troubleshooting. A site engineer can take pictures of a new installation and use an agent to determine if the installation is correct immediately. This process replaces a manual evaluation that typically takes over three weeks, significantly increasing the productivity of operators.
Nokia projects a 50% spectral efficiency improvement by next year and a 100% improvement the year after, effectively doubling the spectral efficiency of current systems in two years. These gains are achieved through the AI-RAN architecture, which allows the network to become an AI service provider.
A Network Engineer’s Inspection Workflow
Suppose a network site engineer in the UK needs to verify a new installation before leaving a site. The engineer takes pictures of the installation and calls up an agent that uses a Nemotron model post-trained on site data. The agent immediately identifies if the installation is correct, eliminating a 3-week wait for evaluation. If the engineer did not have access to the post-trained model, they would be forced to wait for a manual review of the photos before completing the job.
What NVIDIA and Nokia have not yet confirmed about Aerial RAN and spectral efficiency
- NVIDIA has not announced specific pricing for the MGX 4U or DGX Spark systems; check with the manufacturers for current quotes.
- Nokia has not announced a specific date for when the 100% spectral efficiency improvement will be fully realized.
- NVIDIA has not confirmed the exact date for the next generation of the Aerial RAN computer family.
Frequently Asked Questions
What is AI-RAN and how does it differ from traditional networks?
AI-RAN is a software-defined radio access network that allows AI software to run on the same infrastructure as the radio access network. Traditional networks use proprietary equipment built on custom silicon, while AI-RAN uses a software-defined architecture to run multiple functions on a single stack.
How much power is available in existing telecommunications infrastructure for AI?
There is roughly 100 gigawatts of power in today’s telecommunications infrastructure that can be leveraged for the intelligence era. This power is distributed across approximately 100,000 sites including mobile switching offices and points of presence.
What hardware can be used to run AI-RAN at the edge?
The Aerial RAN computer family, including the MGX 4U from Super Micro, the MGX 2U from Dell, the MGX 2U short from Quanta, and the DGX Spark, are designed to run AI-RAN. These systems demonstrate that GPUs can be cost-effectively and efficiently deployed at base stations.
How can I use Sionna for 6G research?
To use Sionna for 6G research, you can download the differentiable link-layer simulator from GitHub. It allows you to simulate and test wireless algorithms in a full simulation environment and even perform over-the-air testing.
Based on NVIDIA’s From AI Infrastructure to the 6G Future.
