Welcome to Issue 11, where Orbital data centers get silicon, IP, and a rules warning
Last week the orbital-AI thesis was being tested against launched hardware and hard cost math. This week it grew its connective tissue: the pieces that turn a thesis into infrastructure — a processor, a patent, a coordination framework, and a production platform.
The space-AI buildout is hardening from capability into infrastructure, with the first coordination rules arriving for the orbital compute side.
On the Space4AI side, the orbital data-center thesis picked up its building blocks. Intel unveiled a radiation-hardened AI processor aimed at government spacecraft, while Sophia Space and Caltech secured a U.S. patent on modular orbital data-center architecture. A Secure World Foundation panel then warned that the density these constellations would add to low Earth orbit outruns the current rules of the road.
On the AI4Space side, Earth-observation AI moved toward production and into orbit. Ai2 launched a continent-scale inference platform built on its open OlmoEarth models, and Orienspace flew what is reported as China's first satellite processing imagery onboard.
Across both sides, the theme is the same — the space-AI thesis is being tested against real infrastructure, not slideware.
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
Intel unveils Starfire radiation-hardened space AI processor on 18A process
On July 27, Intel unveiled Starfire, a radiation-hardened processor aimed at U.S. government spacecraft. The chip is built on the Intel 18A process node and combines Panther Lake-derived x86 cores, a neural processing unit, and an Xe GPU.
The three compute blocks are bound together in a Foveros 3D package, and Intel rates the part at up to 75 TOPS of combined AI performance at 35 watts. Intel is targeting initial customer availability by the end of 2026, pending full radiation qualification.
Sophia Space and Caltech awarded U.S. patent for modular orbital data center architecture
On July 30, Sophia Space announced the issuance of U.S. Patent No. 12,679,564, titled "Space-Based Data Centers," jointly owned with the California Institute of Technology. The patent was filed in October 2024 and issued on July 14, 2026, and it covers architectures for large-scale, modular data centers in orbit.
The claimed design centers on a module Sophia Space calls TILE (Thermal Integrated LEO Edge), measuring 1 m × 1 m × 1 cm and housing four enterprise servers. One face captures solar power to run the onboard computers and storage; the opposite face passively radiates waste heat into deep space. The tiles are meant to be assembled into a large-scale computing surface that is constantly powered by the Sun and constantly radiating. The invention emerged from an internal JPL research program, managed by Caltech, that explores high-risk concepts outside JPL's active mission portfolio.
The patent names seven inventors drawn from JPL, Caltech, and Sophia Space. Dr. Alkalai, a longtime JPL Fellow, founded Sophia Space in June 2023 to commercialize the invention as a portfolio company of Mandala Space Ventures, a Pasadena venture studio. Alkalai, now Founder and CTO, said the work built on Caltech's Space Solar Power Project — funded by a 2013 donation from the Donald and Brigitte Bren Foundation — which in 2023 demonstrated wireless transmission of power from space to Earth. "Instead of beaming down energy to Earth from orbit, we decided to consider putting computing in space and beam down data," he said, adding that cooling in space "is a very tough problem to solve."
Alongside the patent, Sophia Space announced a new sponsored research agreement with Caltech under which Prof. Pellegrino's group will develop lightweight, large-scale deployable structures and thermal management techniques for future orbiting data centers. The company reports $22 million in total funding and plans a 2027 on-orbit demonstration aboard an Apex Nova satellite bus.
Secure World Foundation panel warns orbital data center mega-constellations need new LEO coordination rules
On July 29, a Secure World Foundation panel in Washington, D.C. warned that planned mega-constellations of AI compute satellites will need new coordination frameworks to operate safely in low Earth orbit. Government and industry officials on the panel said the current rules of the road do not account for the density these orbital data center plans would add to LEO.
The scale under discussion is large. SpaceX has floated roughly 1 million satellites, Starcloud 88,000, Blue Origin 51,600, and Cowboy Space 20,000. Starcloud CTO Ezra Feilden noted that collision probability scales as the square of the satellite count, and pressed for updated regulatory standards to manage the risk.
AI4Space
Ai2 launches OlmoEarth Platform for planetary-scale geospatial inference
On July 28, Ai2 launched the OlmoEarth Platform, a production inference and fine-tuning service built on its open OlmoEarth family of Earth observation foundation models, which were pretrained on roughly 10 terabytes of multimodal satellite data. The platform runs inference across continent-scale areas in roughly a day, processing dozens of terabytes of imagery at a cost of fractions of a penny per square kilometer. Ai2 positions it for organizations already adapting OlmoEarth for deforestation monitoring, food security, and wildfire risk that lack the engineering teams to manage labeling, fine-tuning, and large-scale inference themselves.
The system splits each job into three hardware-matched stages: CPU-bound data acquisition and preprocessing that fetches, reprojects, and normalizes imagery; GPU-bound inference; and CPU-bound postprocessing that stitches per-window outputs and exports them as Zarr, GeoTIFF, or GeoJSON. The execution layer, called OlmoEarth Run, partitions a region into machine-sized cells fanned out across up to roughly 1,000 worker nodes, then subdivides each into smaller windows processed independently in separate forward passes. Adjacent partitions overlap slightly and are reconciled at assembly so no seam appears in the final raster.
Ai2 recently used the platform to generate a wildfire-risk map covering all of North America. At peak the run used roughly 19,600 CPUs and 994 GPUs in parallel, with network throughput exceeding 168 GB/s. That parallelism reduced an estimated 4,737 hours of serial compute to about 30.5 hours of wall-clock time, a 155× speedup.
To find imagery, the platform relies on public STAC catalogs but maintains its own metadata index, updated via SNS notifications for AWS Open Data scenes or by polling upstream indexes every few minutes, to avoid overwhelming external services such as ESA's or Microsoft Planetary Computer's STAC APIs. The index holds pointers to Sentinel-1, Sentinel-2, Landsat, and NISAR imagery in cloud-optimized formats, allowing windowed reads that retrieve only the bytes needed for a partition rather than whole scenes. Tasks are reentrant and idempotent, so the platform recovers from routine distributed-computing failures through automatic retries, fallback to alternate providers, and a monitoring process that restarts stalled runners.
The platform currently operates on Google Cloud, though Ai2 says OlmoEarth Run requires only virtual machines capable of running a Docker image plus blob storage, and is designed to support other clouds and deployment within a partner's own account. Ai2's stated roadmap includes scheduled and change-triggered model runs, deforestation and flooding alerts, agentic interfaces, incorporation of ERA-5 weather data, and a dedicated embedding model with precomputed global embeddings to replace full forward passes for many tasks.
Orienspace Gravity-1 sea launch deploys Xiguang-2 01, China's first onboard-AI hyperspectral satellite
On July 22, Chinese commercial launch provider Orienspace flew its Gravity-1 solid-propellant rocket from a mobile launch vessel in the East China Sea off Shanghai, delivering nine payloads into low Earth orbit. Among them was Xiguang-2 01, which Marine Insight reports is China's first satellite equipped with onboard AI capable of processing imagery directly in orbit, allowing collected data to be analyzed and distributed in near real time rather than downlinked raw.
Gabriel Honrada, writing for Asia Times, frames the onboard processing as a way to lessen China's dependence on ground-based image-processing facilities and shorten the delay between collection and delivery of insights to tactical units. The article notes that satellite imagery only a few hours old can already be useless against moving targets such as warships, aircraft and mobile missile launchers, so in-orbit detection and preliminary analysis could compress the targeting cycle, though findings would still require cross-checking with UAVs and over-the-horizon radar before a weapons-quality track.
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: