Welcome to Issue 9, where Orbital AI turns into running code
Last week, the orbital-compute story was about profile and regulatory pushback; this week it became running code on both sides of the ledger.
On the Space4AI side, the thread runs from what is already executing in orbit to what is being planned at national scale. Microchip released its free VectorBlox 3.0 SDK with sparse-network support, and Planetek put that inference stack aboard AI-eXpress-1 in orbit, while China's Nayuta Space unveiled Alaya, a 12,500-satellite gigawatt-class compute constellation.
On the AI4Space side, the models and autonomy are landing on the same trajectory. ESA's Hera verified its autonomous navigation stack through a deep-space software upgrade ahead of its November rendezvous with Dimorphos, and NASA's R5-S9 CubeSat launched to demonstrate onboard edge computing. The Space Force put $69 million behind TALOS, Slingshot's AI agent that plays the adversary in satellite-defense training, and Esri rolled out geospatial foundation models across ArcGIS.
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
Microchip releases free VectorBlox 3.0 SDK for satellite AI; Planetek Italia and AIKO validate it in orbit
On July 14, Microchip Technology (Nasdaq: MCHP) released version 3.0 of its VectorBlox SDK — free software that takes a trained image-recognition AI model and gets it running on Microchip's PolarFire chips, a family of low-power processors flying aboard satellites today. The SDK handles the entire journey from a trained model to a working chip, and both it and the accompanying CoreVectorBlox IP cost developers nothing.
The headline feature is support for so-called sparse neural networks. Many of the numbers inside an AI model are zeros, and a sparse model simply skips the math that involves them — so the chip does less work, answers faster, and draws less power. Microchip's own July 16 webinar put the gain at roughly twice the inference speed, and the compression technology behind it came in with Microchip's Neuronix acquisition. One chip can now also juggle several AI tasks at once, so jobs that used to require separate processors can share a single low-power device. "As AI models continue to grow in complexity, compression is becoming essential for deploying intelligence at the edge," said Shakeel Peera, who heads Microchip's FPGA business unit.
What lifts this beyond a routine software release is that two companies have already proven the stack in space. The first validator is Planetek Italia, which runs it aboard the AI-eXpress-1 satellite, deployed in 2025, through its SPACEDGE services line. Rather than waiting for imagery to reach analysts on the ground, the satellite works through its own pictures as it flies — detecting objects, interpreting scenes, and generating finished information products in real time — with the PolarFire chip providing the radiation resilience that continuous operation in Low Earth Orbit demands. "Leveraging VectorBlox acceleration on Microchip's PolarFire SoC enabled us to efficiently deploy advanced onboard AI pipelines for low-latency payload operations in orbit," said Vito Fortunato, who manages the SPACEDGE line.
The second validator is AIKO, which validated its clear_CHARLES software suite on the same platform — spotting clouds and ships in imagery directly on board the satellite. "The combination of PolarFire SoC and VectorBlox creates a powerful synergy for deploying AI-powered autonomy solutions directly in orbit," said Federico Fontana, AIKO's Head of Hardware Engineering. Microchip cited one further workload already flying on the platform: Spacecraft Pose Network v2, which works out another spacecraft's position and orientation from camera images — the judgement a craft needs for autonomous docking, debris removal, satellite inspection, and formation flying — running on PolarFire parts immune to single-event radiation upsets, with secure boot and anti-tamper protection.
China's Nayuta Space unveils Alaya, a 12,500-satellite GW-class orbital AI compute constellation
At a space-computing conference in Beijing held June 29–30, China's Nayuta Space — legally Qianyi Aerospace (Beijing) Technology Co. — unveiled Alaya, a plan to put 12,500 AI compute satellites into a sun-synchronous dawn-dusk orbit at 600–700 kilometres altitude and grow them into a gigawatt-class orbital computing network built on off-the-shelf AI chips. The name derives from ālayavijñāna, the Buddhist concept of a storehouse consciousness.
The distinguishing engineering move is rocket-satellite integration: the second stage of Nayuta's partially reusable Xuanniao-R rocket does not become debris after launch — it becomes the computing spacecraft itself. Each converted stage would carry roughly one hundred square metres of radiators and four solar arrays, each 3.5 metres wide and 30 metres long, generating about 80 kilowatts for the compute on board; the first stage remains reusable during these launches, leaving up to 17,000 kilograms of mass budget for the second stage's conversion. Recovery relies on Aerodynamic Deceleration–Horizontal Landing, which slows the returning stage with atmospheric drag rather than Falcon-9-style landing burns, keeping the fuel penalty for reuse below 3% — an approach the company estimates yields roughly 30% more payload, about four to six tons per heavy-lift launch. If that architecture sounds familiar to Bits & Orbits readers, it should: it is essentially the pitch Cowboy Space (ex-Aetherflux) raised its $275 million Series B on — the company's own rocket upper stages staying in low Earth orbit as solar-powered, radiatively cooled compute nodes — which we covered in our Pre-Release Issue 0.
The timeline starts small: two technology-demonstration satellites on the first orbital Xuanniao-R flight in the first half of 2027, verifying the overall design, inter-satellite links, and on-orbit data processing; at least one second-stage computing spacecraft serving customers the year after; and hundreds of spacecraft per year from the 2030s if the concept holds. Founder and CEO Li Rui frames the constellation as a way to grow launch cadence on the expectation that AI adoption keeps compute demand climbing. The caveats are real, though: no filings for the thousands of planned spacecraft have reached national or international regulators, and Nayuta has not said how the converted stages will hold attitude or deorbit at end of life. Alaya is also not China's first orbital-compute play — Zhejiang Lab's Three-Body Computing Constellation put its first twelve satellites up in May 2025, targeting a network of thousands with a combined 1,000 peta-operations per second. We covered that launch in Issue 7, where the constellation appears under commercial partner ADA Space's name with a planned build-out of 2,800 satellites.
AI4Space
ESA Hera completes deep-space software upgrade ahead of November Dimorphos rendezvous
On July 8, ESA's Hera spacecraft completed a more-than-two-week software upgrade from 140 million kilometres away, executing two sequential deep-space reboots — one per redundant processor stream — that verified its autonomous navigation stack ahead of the November 2026 rendezvous with Dimorphos, the 160-metre moonlet NASA's DART struck in September 2022. With an eight-minute one-way signal delay and bandwidth at roughly 0.004% of a typical home internet connection, the upload step alone took about three hours. The upgrade also commissions Hera's remaining instruments and activates the inter-satellite links it will use to talk to the Milani and Juventas CubeSats it deploys in early 2027.
The core of the update is the Fault Detection, Isolation and Recovery (FDIR) stack, a data-fusion algorithm that fuses inputs from Hera's cameras, laser altimeter, inertial measurement unit, and star trackers in real time, cross-checking each sensor against the others and flagging discrepancies as faults. During close flybys within one kilometre of Dimorphos, a sixteen-minute round-trip command delay makes real-time ground control impossible, so FDIR lets Hera reconfigure sensors or actuators, and execute a collision-avoidance manoeuvre, on its own authority — autonomy ESA compares to a self-driving car and has identified as foundational for future deep-space missions, including Mars Sample Return and autonomous orbital servicing vehicles.
NASA R5-S9 CubeSat launches on Transporter-17 to demo onboard edge computing
On July 7, NASA's R5-S9 CubeSat was launched aboard a SpaceX Falcon 9, flying as part of the Transporter-17 rideshare mission. The spacecraft — full name Realizing Rapid, Reduced-cost high-Risk Research project Spacecraft 9 — was developed by NASA's Johnson Space Center in Houston. R5-S9 is the latest in NASA's R5 series, which builds spacecraft from commercial off-the-shelf components and designs custom systems only when necessary, improving each vehicle incrementally from its predecessors. The team took roughly four months from design to delivery on this bus. Houston-based SEOPS manifested the CubeSat under a task order on NASA's VADR (Venture-class Acquisition of Dedicated and Rideshare) contract.
The mission's core payload, built in partnership with Sandia National Laboratories, demonstrates edge computing for autonomous onboard observation of phenomena on Earth and in space. R5-S9 also carries a low-cost optical communication system developed by The Aerospace Corporation and supported through NASA's Center Innovation Fund. The satellite hosts multiple technology prototypes for in-orbit validation.
Space Force awards Slingshot Aerospace $69M for TALOS, an AI adversary for satellite-defense training
On July 15, the U.S. Space Force awarded Slingshot Aerospace a $69.2 million, 4.5-year contract — a Small Business Innovation Research Phase 3 award supporting the service's Operational Test and Training Infrastructure (OTTI) program, which aims to give Space Force units more realistic tools for testing systems and preparing personnel for increasingly complex operations in space. Slingshot, based in Windsor, Colorado, made its name in satellite tracking and orbital data analytics, and has been applying those capabilities to training tools for a contested space environment.
The AI in "AI-enabled training" has a name: TALOS, short for Thinking Agent for Logical Operations and Strategy — an AI agent that plays the adversary. Rather than following a predetermined script, TALOS simulates how another spacecraft could maneuver, respond to an operator's actions, or interfere with a mission, which makes exercises less predictable and forces trainees to adjust their decisions as a scenario develops. "TALOS models realistic spacecraft behaviors, generates strategic response options, processes vast amounts of complex data, and supports mission rehearsal across continuously evolving space scenarios," the company said, promising operators "high-fidelity, AI-enabled environments where they can rehearse protect-and-defend scenarios, evaluate courses of action, and sharpen decision-making under realistic operational conditions." The technology was first developed under an earlier SBIR contract.
Esri launches geospatial foundation models (Global Location Encoder, Vision-Language Model) in ArcGIS
At its 2026 User Conference, Esri introduced a new generation of geospatial foundation models across ArcGIS, detailed in a July 12 blog post by Rohit Singh, Vinay Viswambharan, and Priyanka Tuteja. Singh directs Esri's R&D Center in New Delhi; Viswambharan is a principal product manager on the imagery team; Tuteja is a principal product engineer on the GeoAI team. The rollout spans three areas: location encoder models, a Geospatial Vision Language Model, and remote sensing foundation models.
The Global Location Encoder (Sentinel-2) learns from globally available Sentinel-2 imagery to produce embeddings that describe the natural and built environment without manually engineered variables, and ships through ArcGIS Living Atlas as a Deep Learning Package for similarity search, clustering, change detection, and predictive modeling. A companion Geodemographic Foundation Model is trained on thousands of authoritative variables — including U.S. Census, American Community Survey, housing, and environmental data — and its USA Geodemographic Embeddings are released as a beta feature layer. Esri frames embeddings as first-class GIS data that can be stored, analyzed, and combined with traditional spatial information.
GeoVLM connects satellite imagery with natural-language prompts, supporting object detection, pixel classification and segmentation, image captioning, object counting, visual question answering, and region classification from a single model. It was trained on millions of image-text pairs across multiple regions using both Esri-generated and open datasets, and is distributed as a Deep Learning Package so organizations can keep imagery within their own infrastructure.
The remote sensing foundation models are trained directly on satellite and aerial imagery from sensors such as Sentinel, Landsat, and NAIP, and Esri has integrated ArcGIS with several open-source models — TerraMind, Prithvi EO 2.0, Clay, DOFA, and DINO — for embedding generation and fine-tuning via ArcGIS Pro and the arcgis.learn API. Esri credited AWS for the cloud infrastructure used to train the models.
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: