Nvidia puts wireless networks and genAI on the same platform

Telecom networks have always run on rigid, single-purpose gear locked away in cell towers. You buy expensive proprietary hardware, plug it in, and it only ever handles radio traffic.

Nvidia is now pitching a radically different idea: turn those wireless towers into GPU-powered AI hubs.

Nvidia rolled out NVIDIA AI Aerial, a suite of accelerated computing hardware and software designed to bring AI-RAN (radio access network) technology into the wireless mainstream. The pitch is simple: telecom operators can run radio signals and generative AI workloads on the exact same infrastructure.

Instead of just routing packets, cell sites could simultaneously process 5G and 6G signals while powering edge AI for factory robots, self-driving cars, and local assistants.

Digital twins for entire cities

AI Aerial arrives as a full development stack rather than just a batch of cards.

At the core is the Aerial CUDA-Accelerated RAN library, which lets partners deploy virtualized wireless workloads straight on Nvidia hardware. On top of that, Nvidia added AI Radio Frameworks built on PyTorch and TensorFlow, including Sionna, a neural network link-level simulator for 5G and 6G research.

Perhaps the coolest part is the Aerial Omniverse Digital Twin platform.

It models real-world geography, buildings, and signal reflections so operators can simulate wireless deployments with physical accuracy. Engineers can stress-test an entire city network across thousands of base stations inside software before ever bolting hardware to a rooftop.

Nvidia is also backing the tech with serious industrial muscle. The company teamed up with T-Mobile, Ericsson, and Nokia to build the AI-RAN Innovation Center, while partners like SoftBank, Fujitsu, Samsung, and Dell are jumping into the broader ecosystem.

Why it matters

Telecom operators spent hundreds of billions building out 5G, only to watch hyperscalers capture almost all the generative AI upside. AI Aerial gives telcos a realistic way to turn their vast footprint of base stations into revenue-generating edge AI nodes, lowering total ownership costs while selling low-latency compute to local enterprises.

For Nvidia, this is the classic playbook in action. Just like it took over corporate data centers, Nvidia is trying to crack open telecommunications—a famously closed, proprietary market—and absorb it directly into the CUDA universe.

That said, telecom sites care about power efficiency and milliwatts in ways hyperscale data centers never had to. Whether general-purpose GPU stacks can truly beat custom telecommunications silicon on energy economics remains an open question, and Ericsson's cautious performance evaluation shows the industry is testing the numbers closely.