NVIDIA says its open source 5G and 6G research tools have passed 1 million downloads. In a November 6, 2026 post on the NVIDIA blog, the company says global developer adoption of NVIDIA AI Aerial and Sionna grew 10x over the past year, and that more than 8,000 research and academic institutions use these platforms. It then lists five results from universities and an operator.
This is a vendor roundup of partner work. It is still useful, because it shows what AI-for-wireless research looks like in practice, and because every tool named is something you can download. This post separates what NVIDIA reports from what you can check.
TL;DR: the five results NVIDIA highlights
| Who | What they did | Reported result |
|---|---|---|
| King's College London | Robot arm with camera linked to a Jetson AGX Orin running Sionna | Task completion time down by up to 52.6% in fault-injected tests |
| SoftBank | Neural receiver inside Aerial CUDA-Accelerated RAN, tested with a commercial 5G phone | Higher uplink throughput at all 15 tested settings, up to roughly 60% gain with interference |
| ETH Zurich | Tuned neural receivers with local measurements on the Aerial Testbed | Block error rate cut by more than one-third vs selected pretrained versions |
| University of Leeds | Portable 5G network-in-a-box on a DGX Spark with a digital twin | 45.6% lower RAN power vs a full-power baseline |
| CTTC | GPU-accelerated channel models in the 5G-LENA simulator | Simulation time from about three hours to 10 to 20 minutes |
What are Aerial and Sionna?
NVIDIA AI Aerial is the company's wireless software. NVIDIA says it open sourced it about a year ago, and the post links the Aerial CUDA-Accelerated RAN repository on GitHub. A radio access network (RAN) is the part of a mobile network that connects phones to the core, and running it on GPUs lets operators put AI models in the same box as the radio signal processing.
Sionna is an open source toolkit for 5G and 6G research. NVIDIA says it supports software-defined 5G networks and edge AI processing. The Aerial Testbed, used by ETH Zurich, is also named, though the post does not say whether it is open source and gives no direct link.
The post gives no version numbers or license terms for any of these tools. Read the repositories for both before you build on them.
What do the five results show?
King's College London: wireless links that help a robot
King's College London connected a camera-equipped robot arm to an NVIDIA Jetson AGX Orin running Sionna. In fruit-packing and cube-stacking tests with deliberate faults, NVIDIA says task completion time fell by up to 52.6% and task success improved, compared with sending periodic images. The idea is that the network decides what to send based on what the robot task needs, and not on a fixed schedule. It connects 6G research to the robotics work we cover, such as Long-WAM, NVIDIA's world action model.
SoftBank: a neural receiver on a real phone
SoftBank integrated a neural receiver into Aerial CUDA-Accelerated RAN and tested it with a commercial 5G smartphone. With interference present, NVIDIA says it delivered higher uplink throughput than a conventional receiver at all 15 tested settings, with a maximum gain of roughly 60%. A neural receiver replaces hand-designed signal-processing blocks with a trained model. The phone test matters because lab-only receivers often fail on real hardware. The post does not say which interference conditions were used, so the 60% is a best case and not an average.
ETH Zurich: tuning with local measurements
ETH Zurich used the Aerial Testbed to tune neural receivers with measurements taken locally. NVIDIA says this cut block error rates by more than one-third compared with selected pretrained versions. The post points to an ETH Zurich paper on arXiv for details. Note the comparison: "selected pretrained versions", not every baseline.
University of Leeds: a 5G network in a box
Leeds built a portable 5G network-in-a-box powered by an NVIDIA DGX Spark, with a digital twin to validate decisions. Its system predicts demand 10 minutes ahead to guide base station sleep. NVIDIA reports a 45.6% reduction in RAN power consumption versus a full-power baseline while keeping device connectivity. Energy is the main running cost of a mobile network, so sleep scheduling is a serious target. A full-power baseline is a generous comparison, though. Operators already use simpler sleep schedules.
CTTC: simulation in minutes, not hours
The Centre Tecnologic de Telecomunicacions de Catalunya ran GPU-accelerated radio channel models on Aerial CUDA-Accelerated RAN inside its open source 5G-LENA network simulator. NVIDIA says simulation runtime dropped from about three hours to 10 to 20 minutes, depending on the integration, while preserving expected statistical behaviour and network performance. CTTC describes the project on its own site: CTTC accelerates 5G network simulation.
Why does an AI company care about 6G research?
Telecom networks are a large, power-hungry compute market, and NVIDIA wants GPUs in them. The 6G story is also an AI story: researchers expect the next generation of mobile networks to use trained models inside the radio itself, for receivers, scheduling and energy saving. If those models are developed on NVIDIA tools, they run most easily on NVIDIA hardware.
That strategy shows in other NVIDIA moves. We covered the Apollo and Brightspeed plan to turn telephone buildings into an inference network, which rests on a report and not an announcement. The same open-source-first playbook appears in NVIDIA's AI-for-science pledge in Europe, in the C-BRAIN dementia research agents, and in the SETI Allen Telescope Array work. Free tools bring researchers in, and researchers' papers become the case for the hardware. For the wider launch context, see our summary of the GTC Berlin keynote.
What is a neural receiver, in plain terms?
A phone sends data as radio waves. The base station has to turn the noisy signal that arrives back into bits. Classical receivers do this with a chain of hand-designed steps: estimate the channel, equalise it, demap the symbols, decode the error-correcting code. Each step uses a formula that engineers derived under assumptions, such as a particular noise model.
A neural receiver replaces part of that chain with a trained network. It learns from examples what real signals look like, including interference and distortion that the formulas model badly. The promise is better throughput when conditions are messy. The cost is that the model must run fast enough to keep up with the radio, which is why GPU-accelerated RAN software matters, and why researchers need open tools where they can inspect and change the whole pipeline.
That is also why two of the five results compare against a conventional receiver or a pretrained model. The honest question for any neural receiver is not whether it wins in one setup but whether it keeps winning when the environment changes. ETH Zurich's approach of tuning with local measurements speaks to exactly that: a model trained elsewhere may need adjusting for the site where it runs.
How should you read these numbers?
- They are NVIDIA's summaries. Each figure is a headline from a partner project. Only some link to a paper or partner page.
- Baselines differ. "Conventional receiver", "selected pretrained versions" and "full-power baseline" are not the same kind of comparison.
- Best cases are quoted. The 60% and 52.6% figures are maxima ("up to", "maximum gain").
- Lab is not network. A single phone, one testbed or a simulator is far from a live operator network with thousands of cells.
- No licenses in the post. For commercial use, read the repository licenses first.
None of this makes the results wrong. It means the useful move is to open the linked papers and repositories and check the conditions.
How to try the tools
- Start with the Sionna repository if you want a research toolkit for link-level and system-level simulation on a normal GPU machine.
- Read the Aerial CUDA-Accelerated RAN repository if you work on real radio software. Expect specific GPU and network-card requirements, which the repository documents.
- Use the NVIDIA AI Aerial developer topic page, linked from the NVIDIA post, for overview material.
- For a small physical setup, the Leeds design uses a DGX Spark. Our DGX Spark price and memory note covers that hardware.
What this means for builders
Most readers will never touch a base station. The takeaway is wider. Trained models are moving into places that used to run on fixed algorithms: signal receivers, power scheduling and network simulation. Each of the five results swaps a hand-tuned component for a learned one, or speeds up a slow simulation with GPUs. If you work in embedded, edge or robotics systems, the same pattern is coming for your stack, and open toolkits like Sionna are the cheapest way to learn it.
Related reading
- Apollo and Brightspeed: NVIDIA inference network in telephone buildings
- NVIDIA SETI Allen Telescope Array agentic AI
- NVIDIA Europe AI-in-science pledge
- NVIDIA and C-BRAIN open-source dementia research agents
- GTC Berlin 2026 keynote: everything announced
- NVIDIA DGX Spark price and memory
- Long-WAM: NVIDIA world action model
Figures reflect NVIDIA's blog post dated November 6, 2026. Repository details and licenses may change.