Bits & Orbits Weekly - Issue 7

The partnership marks a shift in Satellogic's business model away from one-off imagery sales toward a data-as-a-service approach, where partners including SynMax build the customer-facing AI. At the same time, Satellogic focuses on collecting data and operating satellites. Through a revenue-sharing model, Satellogic outsources analytics to firms closer to customer needs while keeping its own costs competitive.

SpaceKnow CEO and Co-Founder Jerry Javornicky said his customers have been constrained for more than a decade by the cost and frequency of underlying imagery, and that Satellogic's economics and the Merlin constellation change that calculus. The two companies are already collaborating on joint customer pursuits and product development, with additional offerings and milestones expected around the first Merlin launch later this year.

On July 3, two robotics preprints appeared on arXiv, each applying reinforcement learning or neuro-symbolic methods to autonomous aerospace guidance, navigation and control. One targets coordinated spacecraft inspection in orbit; the other addresses autonomous landing under uncertainty.

Patrick Quinn, Bala Prenith Reddy Gopu, George M. Nehma and Madhur Tiwari address multi-agent reinforcement learning (MARL) for controlling groups of inspection spacecraft around an orbiting target. Prior MARL inspection work rewarded agents for reaching a finite set of predetermined inspection points. The authors instead develop a generalized reward function informed by analysis of 3D reconstructions of inspected objects in orbit. Because the reward can evaluate any number of images at arbitrary locations, trained agents gain full control over when images are collected. The 13-page paper is cross-listed under Multiagent Systems and integrates a published correction to the original manuscript.

NEUROSYMLAND, from a team led by Weixian Qian and Xi Zheng, is a neuro-symbolic landing-site assessment system built to run on edge-constrained hardware. It constructs a probabilistic semantic scene graph from onboard visual input, then scores candidate landing regions against symbolic constraints for terrain flatness, obstacle clearance, and spatial consistency. NEUROSYMLAND validates on aerial platforms rather than spacecraft, but the problems it solves are the space domain's. Autonomous descent under uncertainty, real-time landing-site and hazard assessment on compute-constrained edge hardware, and disturbance-robust terminal guidance are the same challenges facing planetary and lunar landers — the kind of onboard, vision-based hazard avoidance that steered Firefly's Blue Ghost to its 2025 lunar touchdown. Edge-deployable GNC proven on inexpensive UAV testbeds is a well-worn on-ramp to flight-qualified spacecraft autonomy, which is why aerial-robotics preprints keep surfacing on the same arXiv track the space community watches.

On June 30, the University of Southern California announced that its student-built MAVERIC satellite is set to launch in July aboard a SpaceX Falcon 9 rideshare mission. MAVERIC is a 3U CubeSat developed at the USC Space Engineering Research Center within the USC Information Sciences Institute in Marina del Rey. More than 60 students and faculty took the satellite from concept to flight over two years, supported by a gift from Positron Capital Management.

The mission centers on imaging methods for on-orbit servicing, in which one spacecraft approaches another to inspect, repair, maintain, or refuel it. Using two cameras, MAVERIC will produce 2D and 3D images meant to give operators a perspective of nearby spacecraft during close-proximity operations. "Being able to watch what's happening and step in when needed during close-proximity operations helps build trust in autonomous systems," said David Barnhart, research professor of astronautical engineering at USC Viterbi and co-founder and director of the Space Engineering Research Center.

The imaging system also supports industry partner Planetary Systems AI, which is conducting its first on-orbit demonstration of AI-powered software. The company will use MAVERIC imagery to train machine learning models and evaluate processing space-based data directly in orbit, reducing the need to transmit raw data to Earth. "Taking our AI systems into orbit aboard MAVERIC and partnering with USC was two years in the making," said Cindy Chin, CEO and founder of Planetary Systems AI.

MAVERIC also tests a navigation approach that departs from standard practice. Rather than reaction wheels, the CubeSat uses Earth's magnetic field to adjust its orientation in orbit, and researchers will analyze flight data, apply AI-based reinforcement learning on the ground to refine the navigation software, and upload improved algorithms back to the spacecraft. The mission additionally evaluates low-cost magnetic field sensing, testing whether inexpensive CubeSats can collect high-quality data to support higher-fidelity global space-weather measurements.

On June 30, Orbital Compute, Inc., an LA-based startup five months old, filed with the Federal Communications Commission for permission to deploy up to 100,000 AI-focused data-center satellites in low Earth orbit. At full scale, the constellation is designed to deliver 10 gigawatts of computing power.

The satellites would fly at altitudes of 500 to 850 kilometers, with each 100-kilowatt-class spacecraft acting as a high-density server rack powered by a solar array and connected to third-party constellations via optical inter-satellite links. Orbital plans sun-synchronous orbits so the fleet sits in uninterrupted sunlight. For scale, the 10-gigawatt target matches the total new electricity capacity added to the entire United States power grid last year.

The company was founded by Euwyn Poon, who previously started the dockless electric-scooter firm Spin and sold it to Ford. "The demand for AI compute is outrunning what we can reasonably build on the ground — we're short on power, land, and water all at once. Space solves all three," Poon said. Orbital has closed a $5 million pre-seed round (reported in Issue 4). The company plans to launch a single-GPU demonstration payload on a SpaceX Falcon 9 next year to test how Nvidia chips withstand space radiation, followed by its first full-scale satellite, Orbital-1, in 2028.

On July 1, AIAA's Aerospace America published "Inside the space data race," a feature by Jonathan O'Callaghan surveying the emerging orbital data center (ODC) sector and the companies racing to move AI compute off the ground. The piece maps active demonstrators and the far larger constellations now filed with regulators, and pegs the market at $39 billion by 2035 per market intelligence firm BIS Research.

Two ODC demonstrators are already on orbit. Starcloud-1, launched in November 2025, carries an NVIDIA H100 chip to train a Google AI model, and last May the Chinese firm ADA Space launched the first 12 of a planned 2,800-satellite constellation for on-orbit AI processing. The feature traces the surge in interest to January, when Elon Musk tweeted that SpaceX would launch 1 million AI satellites; the company's FCC filing specifies a constellation of "up to" 1 million — we covered SpaceX confirming the 'Starmind' name and that filing in Issue 6. "Elon tweeted, and here we go. We have a new reality," says Jaroslaw Jaworski, CEO of Luxembourg-based Edge Aerospace to Aerospace America.

The largest filed proposals the feature cites are 1 million satellites from SpaceX, Orbital Compute's 100,000-satellite filing (above), 88,000 from Starcloud — whose GPU constellation we mapped in our Issue 3 — and 51,600 from Blue Origin. The draw is power and heat: the International Energy Agency estimates terrestrial data centers could consume as much power as all of Japan by 2030, and a dozen U.S. states are now weighing curbs on ground facilities, says Lori Gordon, a systems director at the Aerospace Corp. running an ODC feasibility study.

The obstacles the feature lays out are physical. The ISS generates just 75-90 kW from its football-pitch-sized arrays, so gigawatt-scale ODCs would need enormous panels, and heat cannot convect in vacuum, forcing large infrared radiators. Collision risk compounds the problem: Starlink's roughly 10,000 satellites performed 300,000 avoidance maneuvers in 2025, and University of Birmingham astronautics professor Hugh Lewis predicts sun-synchronous ODC constellations could push the field toward "half a billion maneuvers a year." Launch economics also lag — Nanyang Technological University's Yonggang Wen says prices must fall to $200-$300 per kilogram, which "you will need Starship" to reach.

The feature also catalogs adjacent bets and skeptics. Florida's Star Catcher set a 1.1-kW optical power-transmission record beaming a laser across a shuttle runway, and Overview Energy signed an April agreement with Meta to supply up to one gigawatt of power from space. Sophia Space proposes ultra-thin "TILE" satellites that radiate heat directly (we covered Sophia's TILE compute modules and its 2027 Apex demo booking in Issue 6). In May, ESA contracted Edge Aerospace to study ODCs, and Aetherflux rebranded as Cowboy Space after raising $275 million, targeting a one-megawatt data center by the end of 2028 (both were surfaced in the Pre-Release Issue 0 of Bits & Orbits). NVIDIA's Dion Harris cautioned that ODCs "are more likely to complement terrestrial infrastructure than replace it."

On July 2, perovskite solar panel startup Verde Technologies said it is shifting its commercialization focus to space, moving away from its initial rooftop market. The company is betting that its thin-film perovskite material can help power orbital data centers and other large constellations, which it is targeting as its initial market.

Till next time,