Welcome to Issue 19, where AI chips reach orbit as autonomy drops the ground
Last week orbital compute was about how these constellations get financed. This week it became silicon on the launch manifest, while onboard autonomy moved from assisting operators to cutting them out entirely.
On the Space4AI side, the hardware is going up. Google detailed the first in-orbit prototype of its Project Suncatcher research effort, carrying four of its Trillium AI accelerators and set to fly October 1 on a SpaceX rideshare with Planet. Days earlier, China launched Chaozhisuan-1, a satellite built to run AI on Earth-observation data in orbit.
On the AI4Space side, autonomy keeps stepping back from the ground. ESA's Earth-observation chat companion gained humanoid robots as conversational front-ends. NASA flew its autonomous precision-landing package on a drone, running simulated lunar descent maneuvers without a pilot. AstroForge went furthest, unveiling a deep-space mission that runs entirely on its in-house intelligence model with no commands from Earth after separation.
Both sides are converging on machines that operate in space without waiting on the ground.
Specifics below.
Did you know?
We have recently launched the Bits & Orbits Insights
Check out the growing set of independent dashboards, trackers, and maps on the
AI × Space frontier:
Space4AI
Google's Project Suncatcher MVP with four Trillium TPUs to launch on Falcon 9 Transporter-18
On September 24, 2026, Google detailed the first in-orbit test of Project Suncatcher, its research moonshot to run machine learning hardware in space. The prototype satellite carries four Trillium TPUs and roughly 1 kW of solar power, and is scheduled to launch October 1 on SpaceX's Falcon 9 Transporter-18 rideshare mission, developed in partnership with Planet.
Project Suncatcher, announced last year, is built on the premise that low Earth orbit offers near-constant sunlight, which Google says can generate up to eight times more solar power than the equivalent on Earth. The MVP is designed to answer a narrow question first: whether Google's TPUs can survive spaceflight and operate through the radiation and thermal extremes of orbit. Writing on the announcement, Travis Beals, Senior Director for Paradigms of Intelligence, framed the launch as a deliberate first step rather than an operational data center.
The team has already run the hardware through ground qualification. A launch subjects the spacecraft to about ten minutes of intense vibration and sustained acceleration up to 10 g, with individual components such as the TPU chips seeing 50 to 100 g, so engineers shook the satellite on all three axes to mimic launch frequencies. For radiation, Google tested Trillium TPUs in a proton beam at UC Davis's Crocker Nuclear Laboratory while running AI workloads and monitoring for bitflips; initial results showed the chips tolerated a total ionizing dose greater than what a five-year mission would deliver. Cooling is the other open problem, since a vacuum has no airflow and heat must be shed through radiators, so the team is combining heat pipes and radiators validated in a thermal vacuum chamber.
The orbital launch is meant to surface failure points that ground tests cannot, ahead of a 2027 milestone that puts two satellites in orbit to test free-space optical laser links between spacecraft. Future satellite designs are intended to carry dozens of TPU chips flying in clusters, communicating over high-bandwidth lasers at short range. Google likens that pointing problem to hitting a coin-size target from miles away while both ends are in motion.
China launches Chaozhisuan-1, an orbital AI compute satellite, on September 20
On September 20, 2026, CAS Space launched Chaozhisuan-1 on its Kinetica-1 rocket. The satellite is built to process Earth-observation imagery with onboard AI instead of sending the raw data to the ground.
TechTimes reported that the Kinetica-1 Y18 flight lifted off at 12:03 pm China Standard Time and placed nine satellites into a sun-synchronous orbit about 500 kilometers up. It was the rocket's 16th flight and China's 68th orbital launch of the year.
The spacecraft carries a high-resolution optical camera and an image-processing computer. Tom's Hardware, citing Digitimes, reported that keeping that processing in orbit is meant to cut cross-regional data handling from hours to minutes. TechTimes reported a 4-meter-resolution visible-light camera whose raw output would otherwise overwhelm limited ground-station windows. Under the usual bent-pipe model, the satellite relays imagery for processing on the ground. Chaozhisuan Beijing Technology, the developer, says onboard inference identifies objects, filters events, and flags anomalies, then sends only the results down, compressing response time from hours to minutes. TechTimes reported that this timing claim comes from the developer and has not been independently tested.
TechTimes reported that optical laser links let Chaozhisuan-1 take raw data from other satellites and return processed results, so it works as a shared compute node rather than a single camera. A containerized runtime lets models be updated or replaced by uplink after launch, and the satellite is designed to keep operating between intermittent ground contacts.
According to TechTimes, Chaozhisuan Beijing Technology was founded in July 2024 and announced publicly in May 2026. Chaozhisuan-1 is meant to be the first satellite of a constellation that could eventually number in the thousands, with no firm deployment timeline. A March 2026 industry alliance behind the program includes Zhipu AI, which rebranded internationally as Z.ai and was added to the US Commerce Department's Entity List in January 2025, and SenseTime, on the Entity List since 2019 and placed on a US Treasury investment blacklist on December 10, 2021. Other partners include Beijing Wenge Technology and Tsinghua University.
Tom's Hardware placed the launch in China's wider orbital-computing push, including the government's early-June approval of a Space Computing Industry Innovation Center, and contrasted the satellite with SpaceX's proposed AI1, which targets 150 kW of peak compute power.
AI4Space
ESA Φ-lab integrates two humanoid robots into EVE Earth Virtual Expert
On September 23, 2026, ESA Φ-lab announced that two humanoid robots, Ameca and Frankie, have been integrated as conversational front-ends for EVE (Earth Virtual Expert), the chat-based Earth-observation companion Φ-lab has developed with Pi School.
The two robots were developed by Frank Reply GmbH and are equipped with built-in cameras for real-time visual analysis of images on a screen, interpreting natural-language queries while EVE handles the underlying data analysis. Ameca handles satellite image searches, interpreting plain-language requests to locate and retrieve the most relevant imagery from EVE's database. Frankie focuses on vegetation analysis, breaking down concepts such as the Normalised Difference Vegetation Index (NDVI) by looking at a satellite image on a screen. EVE itself was built with support from Imperative Space, Mistral AI and Wiley.
The setup exposes three modes: an image-recognition game guided by Ameca, an NDVI-based crop-health inspector led by Frankie, and a "Behind the Scenes" mode. In that mode, a dedicated screen reveals EVE's operational workflow in real time as a user speaks, showing retrieval, reasoning, planning and tool calls as they happen. The system fuses a low-latency speech-to-speech AI model with EVE's agent capabilities and grounded data retrieval.
Daniel Fitzpatrick, Senior Manager at Frank Reply, said the conversational interface lets people request specific imagery, such as of their hometowns, and watch requests rendered into machine language via EVE's toolchains. Nicolas Longépé, Earth Observation Data Scientist at Φ-lab and EVE's Technical Officer, said the interfaces turn complex Earth-observation technology into natural dialogue without users needing to speak the specialised language of remote sensing.
Φ-lab plans a demo and hands-on session at the "Agentic AI for Earth Observation" workshop in Berlin in October, with a fully operational version of EVE to launch afterward.
NASA SPLICE autonomous precision-landing system completes simulated lunar descent in drone tests
On September 24, 2026, NASA published imagery from an August 27 test in which an Alta-X quadcopter drone flew its Safe and Precise Landing – Integrated Capabilities Evolution (SPLICE) guidance package near the agency's Armstrong Flight Research Center in Edwards, California. NASA's Johnson Space Center in Houston developed SPLICE, which completed simulated lunar descent and landing maneuvers during the drone testing.
SPLICE bundles several sensors and processors to enable autonomous precision landing on the Moon, Mars, and icy worlds. Terrain Relative Navigation uses a camera that compares live images of the surface against existing orbital imagery to fix the spacecraft's location. Navigation Doppler Lidar fires laser beams from four fixed-angle telescopes and reads the returns to estimate velocity and altitude, while a Hazard Detection Lidar builds a 3D map of the landing field to flag steep slopes and large rocks. The Descent and Landing Computer, a high-performance multicore processor, fuses all sensor data and computes a safe landing location.
The system is aimed at landing sites that were out of reach during Apollo, such as regions with hazardous boulders and permanently shadowed craters that hold high scientific interest. NASA states the same SPLICE technologies could later help land humans on Mars.
AstroForge unveils Autonomy-1, first fully autonomous deep-space mission powered by Solo transformer model
On September 21, 2026, asteroid-mining startup AstroForge, in an announcement by CEO Matt Gialich, introduced Autonomy-1, which it describes as the first space mission to complete its entire mission after separation without a single command from the ground. The mission is powered by Solo, AstroForge's in-house spacecraft intelligence model, and will remove the human operator from the loop entirely.
Solo is an intelligence layer that sits above the spacecraft's existing deterministic, physics-based flight software rather than replacing it. It ingests the spacecraft's state, identifies when something is off-nominal, and determines what should happen next, flying the vehicle without ground updates. AstroForge argues Solo can act on a fuller picture than any ground team, because deep-space communications limits how much telemetry can be downlinked, while Solo reads the data at the source onboard. The company frames the problem in cost terms: mission operations dependence on Earth makes up nearly one-third of its overall mission costs and does not scale to fleets of spacecraft.
Solo will first fly in shadow mode on AstroForge's next mission, DeepSpace-2, running onboard and processing real spacecraft data while its decisions are not executed by the vehicle. Autonomy-1 will be the first time Solo takes control. The mission will launch aboard the first flight of Stoke Space's Nova Pathfinder launch vehicle, and will transmit telemetry and science data one way, from the spacecraft to the ground, with no uplink commands.
Autonomy-1 will also carry COMPASS, a NASA Goddard heliophysics payload, so Solo will be responsible for coordinating the functions required to support a real scientific payload in addition to operating the spacecraft. AstroForge positions the work against prior milestones such as NASA's Deep Space 1 "Remote Agent" nearly three decades ago and OSIRIS-REx's autonomous touchdown and sampling, noting those systems limited autonomy to specific functions or mission phases while human operators retained responsibility for the mission itself.
Till next time,
Meta-beat Column of this week
Read also about the AI Pipeline that sits at the core, producing this Newsletter, including its ups and downs of this week: