NVIDIA says AI agents now steer the SETI Allen Telescope Array, calibrating it and analyzing up to 2.4 Tbps of live radio data from plain-language requests

Background and the problem

Radio telescopes produce huge data volumes. The usual approach writes everything to disk and analyzes it later, so researchers can wait weeks to learn whether an observation held anything worth following up, and steering, calibration, and packet-loss debugging depend on a few engineers who know the instrument. The Allen Telescope Array (ATA) at Hat Creek Radio Observatory in Northern California has 42 dishes of 6 m each and was the first radio telescope designed for SETI. NVIDIA's case study brings a coding-agent style of work to a real scientific instrument.

Architecture and how it works

According to NVIDIA's blog, the system has three layers. DAQIRI, a Holoscan platform library, pulls terabits per second of sensor data straight into GPU memory so it is processed in flight rather than written to disk first. CyberEther, a signal-processing framework from the SETI Institute, does the math with CUDA-X libraries such as cuFFT and cuBLAS. Nexus, a web-based observatory orchestrator, acts as the agent harness and gives agents one source of truth for instrument state, compute, and pipelines. Researchers can check packet loss, read live spectrograms, and rewrite processing blocks in natural language.

Benchmarks and reported results

Only two figures are given, both vendor-reported: up to 2.4 Tbps on RTX PRO 6000 Blackwell GPUs, and the first real-time AI detection of a pulsar in the Crab Nebula. NVIDIA did not publish detection accuracy, false-positive rates, or comparisons with the old pipeline, and "weeks or years compressed into hours" is a qualitative claim. No extraterrestrial signal has been found.

What developers can do with it

For agent builders, the useful part is the harness design: Nexus exposes hardware state and pipelines through one interface agents can read and write, much like giving a coding agent tools, memory, and checks. To build something similar, start with the NVIDIA Holoscan SDK and DAQIRI docs, and look at the open CyberEther project for signal processing. The team is next evaluating open models such as Nemotron for local or offline inference.