On October 30, 2026, NVIDIA announced that AI agents have been introduced into SETI (the Search for Extraterrestrial Intelligence) observations. Breakthrough Listen and the SETI Institute have deployed the agents in routine observations at the Allen Telescope Array (ATA), a group of radio telescopes in northern California. Using only natural-language instructions, researchers can adjust the telescopes' pointing, calibrate observational data, and run AI-based analysis. The platform that processes the large volumes of observational data in real time is NVIDIA Holoscan. According to NVIDIA, the new system can process 2.4 terabits of data per second, and it succeeded in using AI to detect the Crab Nebula pulsar in real time. However, being able to process large amounts of data quickly is a separate matter from being able to accurately identify artificial signals originating from extraterrestrial civilizations. The demonstration targeted a known natural object, and no signal from an extraterrestrial civilization has been discovered.
Real-time GPU analysis and AI-agent telescope control
Analyzing the large volume of signals captured by a radio telescope in real time requires feeding the observational data into computing hardware at high speed, running the necessary signal processing, and managing the telescope's status and computing resources at the same time. In this system, three main components share the work: DAQIRI handles data ingestion, CyberEther performs GPU-based signal processing, and Nexus oversees the instruments and processing stages.
| Role | Component | Main function |
|---|---|---|
| Data ingestion | DAQIRI | Ingests observational data directly into GPU memory |
| Signal processing | CyberEther | Processes signals on GPUs and visualizes the results |
| Observation and processing management | Nexus | Oversees instrument status, computing resources, and processing stages |
According to NVIDIA, CyberEther, developed by the SETI Institute, performs signal processing using CUDA-X libraries such as cuFFT and cuBLAS. Nexus, meanwhile, functions as a web-based observation management system that centrally manages instrument status, computing resources, and the flow of data processing. By giving natural-language instructions, researchers can monitor data loss and check spectrograms, which show how a signal's frequency components change over time. The system also supports changes to signal-processing blocks, linking telescope operation with the analysis of observational data.
The technologies that make up the system are also introduced on Stelline's official website. The site demonstrates live data from the ATA being ingested into CyberEther over 100 Gigabit Ethernet (100 GbE) for real-time correlation and visualization. Stelline also describes its "Coyote" class servers as carrying four NVIDIA RTX 6000 Blackwell GPUs per unit, with up to 800 gigabits per second of fiber-optic input and up to 256 terabytes of NVMe storage. The mechanism for ingesting data directly into GPU memory and the high-capacity storage for retaining observational data can be used together, depending on the application.
The 2.4 terabits per second figure that NVIDIA announced this time, on the other hand, is the performance of a system using RTX PRO 6000 Blackwell GPUs. NVIDIA says AI inference runs on facilities installed near the telescopes. However, the detailed measurement conditions, the observation time, and the number of GPUs used across the entire system have not been disclosed. Nor can the number of servers needed to process 2.4 terabits per second be determined from the Coyote server specifications alone.
From a 2024 real-time processing experiment to integrating observation control and AI analysis
This is not the first effort to use AI to process observational data from the ATA in real time. In its announcement on October 8, 2024, NVIDIA reported the results of an earlier experiment using IGX Orin and Holoscan. In that experiment, 28 ATA antennas were pointed at the Crab Nebula, and more than 90 billion data packets were collected over 15 hours. According to NVIDIA, observational data equivalent to nearly 100 gigabits per second could be analyzed in real time. This result came from observations with an actual radio telescope, not from a simulation.
What stands out in the latest announcement is that, in addition to this kind of real-time processing, it integrates telescope operation by natural language, instrument adjustment, calibration of observational data, and AI analysis. Rather than having researchers carry out each operation and processing step separately, the system lets them carry out a series of observing tasks through an AI agent. It can be seen as an effort to automate observation itself, not just to process large volumes of data. However, because the equipment and measurement conditions differ between the 2024 experiment and the current system, the rate of performance improvement cannot be calculated from the published data-processing figures alone.
What did the real-time pulsar detection prove?
NVIDIA cites the real-time AI detection of a pulsar in the Crab Nebula as a demonstrated result of this announcement. A pulsar is a type of rapidly rotating neutron star formed at the end of a star's evolution, and it emits periodic electromagnetic waves. What was detected this time was a signal from such a known natural object, not an artificial signal emitted by extraterrestrial intelligence. NVIDIA describes it as the first real-time AI detection of a pulsar, but this does not mean that an extraterrestrial civilization has been found.
The announcement also gave no detailed measurement results, such as observation time, number of detections, or recall. False-positive rates and comparisons with results obtained without AI have not been published either, so it is impossible to judge how much AI improved signal-detection accuracy. The integration of the observation system has been reported in NVIDIA's corporate blog and Stelline's technical materials, but no reference has been given to a peer-reviewed paper verifying this configuration or to independent replication results.
Looking ahead, NVIDIA says the research team is evaluating open models, including Nemotron. It says these are envisioned for cases where observational data must be processed on-site and for applications requiring rapid responses. It also has an eye on eventually applying the system to observatories operating on the far side of the Moon without relying on communications with Earth. However, these are evaluations currently under way and future concepts, and they do not mean that deployment of the system on the Moon has been decided.
This initiative uses AI agents to link radio telescope operation with the analysis of large volumes of observational data, advancing the automation of observing work. To judge how much it will contribute to the SETI search, however, its detection capability under different observing conditions, its false-positive rate, and the reproducibility of its results will need to be verified. As such scientific verification progresses, it should become clear how effective AI-driven observation automation can be at finding important clues among the vast number of radio signals arriving from space.