Artificial intelligence agents steer space observatories: processing 2.4 terabits per second cuts years of analysis to hours
Radio astronomy observatories have abandoned storing tons of raw data before examining it, opening the way for real-time processing systems managed by software agents.In the Allen Telescope Array in northern California, the “Breakthrough Listen” initiative and the “SETI” Institute have begun operating artificial-intelligence agents that steer the telescope, calibrate its instruments and apply mathematical models to incoming signal streams directly, all via a single operational command phrased in natural language.
Radio telescopes have traditionally generated massive volumes of sensor data, creating a persistent operational bottleneck for researchers seeking to detect meaningful signals amid the flow of astronomical observations. To alleviate this choke point, the team adopted the “NVIDIA Holoscan” platform and its associated “Daceri” library, a framework that pulls terabytes of data per second directly into GPU memory, enabling real-time processing as the signal streams rather than filling hard-disk media and accumulating it.
The “Nexus” desktop coordination system serves as the hosting environment for the software agent, providing it with a unified source of truth about device status, computing resources and workflow paths.This architecture allows researchers to monitor missing data packets, read live spectral plots and rewrite signal-processing blocks using direct textual commands. Complex calculations, however, are performed by the “Cyber Ether” framework developed by the “SETI” Institute, relying on accelerated computing libraries such as fast Fourier transforms and basic linear algebra operations.
The system’s on-site processing capacity reaches 2.4 terabits per second, using “RTX Pro 6000 Blackwell” processors, with inference performed right next to the telescopes. This configuration enables the agent to reposition the telescope’s dishes, focus on the clean signal, detect and correct data errors, and plot results instantly, compressing a workflow that previously took humans weeks or years into just a few hours. The approach has already demonstrated its practicality by delivering the first AI-assisted instantaneous detection of a rotating astrophysical burst in the Cancer Nebula, while the team moves toward evaluating open models such as “Nemotron” for operating future offline observatories.
We see in this technical model a direct message for infrastructure and data teams in the Gulf and Egypt, especially those working in satellite-downlink stations, geological sensing and radar monitoring.The problem this system solves is not astronomy per se, but the challenge of ultra-dense data streams that burden observation networks and seismic-survey projects for the region’s energy sector. Continuing with the legacy model of first archiving raw data and then moving it to the cloud for batch processing later raises cloud-storage costs and creates a temporal gap that prevents real-time intervention.
The genuine transformation needed locally consists of building edge-processing architectures directly next to the sensors, replacing conventional software pipelines with agent environments capable of correcting the signal path instantly via natural-language commands. This shift moves the skill set required of regional systems engineers from managing cloud databases to mastering real-time processing libraries and edge graphics computing, so that operational decisions are based on what processors capture in fractions of a second rather than waiting for delayed analysis cycles.