AI Agents Now Steer the Allen Telescope Array to Hunt for Alien Signals
SETI just handed the keys of a massive radio telescope to AI agents that process 2.4 terabits of data per second in real time.
Research · Source: NVIDIA Blog
What happened
Breakthrough Listen and the SETI Institute just put AI agents in charge of the Allen Telescope Array in Northern California. Researchers can now steer the massive radio telescope using simple natural language requests. The agents handle the heavy lifting. They calibrate the antennas, fix data errors, and run AI models on live streams. A web-based orchestrator called Nexus acts as the agent harness. It gives agents a single source of truth for the instrument state and compute resources.
The system relies on NVIDIA hardware and software to process data in flight. A library called DAQIRI pulls sensor data straight into GPU memory. This skips the old method of writing everything to hard drives first. The SETI Institute developed a framework called CyberEther to do the math using NVIDIA CUDA-X libraries. The setup runs on premises using RTX PRO 6000 Blackwell GPUs. It handles a massive 2.4 terabits of data every second right next to the instrument.
The pipeline already proved it works. The automated system achieved the first real-time AI detection of a pulsar in the Crab Nebula. Agents compress what used to take humans weeks or years into mere hours. Now the team is testing open models like NVIDIA Nemotron. They want to see if these models can run offline for future observatories on the far side of the moon.
Key facts
- 2.4 terabits — Data processed per second by the on-premises system
- RTX PRO 6000 Blackwell — NVIDIA GPUs running the inference next to the telescope
- DAQIRI — Holoscan library pulling data straight to GPU memory
Why it matters
We are moving from AI that analyzes dead data to AI that operates live physical infrastructure. Processing 2.4 terabits per second in flight means the bottleneck is no longer storage speed. Builders can now design systems that react to massive sensor streams instantly. You do not need to save data to disk if your agent can understand it directly in GPU memory. This architecture changes the game for any industry dealing with high-frequency sensor data.
This changes how we build autonomous scientific instruments. If an agent can steer a telescope and clean its data locally, it can operate in extreme isolation. The SETI team is already planning for lunar observatories where cloud connectivity is impossible. This pushes the edge of edge computing. We will see more autonomous agents deployed in environments where data locality and fast turnaround are the only things that matter.
For builders
Process sensor data in flight
Writing massive data streams to disk kills performance. Use tools like Holoscan to pull data straight into GPU memory. Infrastructure builders who master in-flight processing will win contracts for high-throughput sensor networks.
Build natural language hardware controls
The SETI orchestrator lets researchers adjust processing blocks and steer antennas with plain text. Hardware startups must stop building clunky dashboards. The new standard is an agent harness that translates prompts into physical actions.
Deploy open models for offline edge
Cloud APIs are useless on the moon. The SETI team is testing open models like Nemotron for offline use. Founders building for remote industrial or space applications must optimize open weights for local inference.
My take
We are building AI agents to listen for alien AI agents. That sounds like pure science fiction, but the tech stack is already running live in California. If an autonomous agent can steer a radio telescope and spot a pulsar in real time, your enterprise software has absolutely no excuse for being slow.
Original reporting: NVIDIA Blog. This is my rewrite and opinion.