Welcome to Issue 0 — a pre-release dry-run ahead of the regular weekly cadence.
This week's stories bend toward a single observation: compute is migrating to orbit, from both directions at once.
On the Space4AI side, the orbital-data-center thesis crossed several thresholds in the last few days — a quiet talks report between two of the largest names in tech and space, a unicorn-scale Series B for a vertically integrated newcomer, a venture-backed challenger pitching an inference-tier satellite mesh, and Europe preparing its potential entrance into the Orbital Data Center arena. The commercial side and the policy side moved in step.
On the AI4Space side, the workloads those orbital data centers will eventually run are already meeting the satellites halfway. A next-generation space processor posted a step-change in published performance. An open-source geospatial foundation model became the first of its kind to run in orbit. An in-orbit demonstration moved ML object detection onto the spacecraft itself. And some other stories touching the edge of EO and AI.
Two threads, one direction. Specifics below.
Space4AI
Google and SpaceX reportedly in talks to launch orbital data center test products into LEO
On 2026-05-12, the Wall Street Journal reported that Google and SpaceX are in talks to launch orbital data center test products into low Earth orbit. The two companies are jointly exploring the economics of running AI compute infrastructure in space.
Current cost estimates for orbital compute exceed terrestrial alternatives, according to the report. The pitch centers on two physical advantages of LEO operation: continuous space solar power and passive radiative cooling.
a16z-backed Orbital unveils plan for mesh constellation of GPU satellites for AI inference
On 2026-05-12, IEEE Spectrum reported that Andreessen Horowitz–backed startup Orbital disclosed plans for a mesh constellation of small LEO satellites, each carrying GPU server racks for AI inference workloads. The company intends to finalize satellite designs in 2026 and target a 2027 launch.
Each satellite would draw power from solar panels described as roughly tennis-court-sized and reject heat through radiative cooling panels. Orbital plans to sell capacity to large-model AI inference customers via direct API access rather than co-locating with existing cloud providers.
Cowboy Space (ex-Aetherflux) closes $275M Series B at $2B valuation for orbital DCs and rockets
On 2026-05-08, Cowboy Space Corporation — the company formerly known as Aetherflux, founded by Robinhood co-founder Baiju Bhatt — closed a $275M Series B at a $2B valuation, led by Index Ventures.
The company is pursuing a vertically integrated architecture in which its own rockets' upper stages remain in low Earth orbit and operate as solar-powered, radiatively cooled compute nodes. The stated target workload is AI inference that has been priced out of terrestrial power grids.
European Commission (DG DEFIS/DG RTD) hosts EU Orbital Data Centers workshop in Brussels
The European Commission has been convening an Orbital Data Centers workshop on 13 May in Brussels, co-organised by the Directorate-General for Defence Industry and Space (DG DEFIS) and the Directorate-General Research and Innovation (DG RTD). The workshop was structured to collect stakeholder input on Orbital Data Centers (ODC) from across the European space and research ecosystem.
Image: DG-DEFIS
The session was positioned as an input-gathering exercise rather than a policy announcement, with DG DEFIS and DG RTD jointly running the agenda. The co-organisation pairs the Commission's defence and space industrial portfolio with its research and innovation arm, indicating the file is being scoped across both dimensions in parallel.
Edge Aerospace wins ESA Phase I Space Cloud contract and completes first in-orbit mission
On 5 May 2026, Luxembourg-based Edge Aerospace announced a Phase I contract under the European Space Agency's Space Cloud Advanced Concepts Program to develop an architecture and use-case roadmap for orbital data centers. The same announcement reported completion of the company's first in-orbit mission and a U.S. expansion.
The Phase I scope covers studying the commercial viability of orbital-compute power and identifying pathways for Europe to leverage space-based computing infrastructure. The award places Edge Aerospace inside ESA's early-stage pipeline for orbital data center concepts alongside its operational and commercial expansion steps.
AI4Space
Consumer-derived NPUs drive surge in autonomous onboard satellite processing
On 2026-05-12, SatNews reported that satellites fitted with consumer-derived NPU-based edge processors are now handling terabytes of sensor data on-orbit and downlinking only actionable insights.
The shift moves workloads such as autonomous targeting, automated change detection, and real-time ML inference from the ground segment to the spacecraft itself. According to the report, this reduces ground-station dependency and cuts communication overhead for operators running EO and tasking-heavy missions.
NASA's High Performance Spaceflight Computing (HPSC) processor hits ~500× current rad-hard chip performance in JPL tests
On May 12, 2026, NASA's Jet Propulsion Laboratory reported that the High Performance Spaceflight Computing (HPSC) processor under test is achieving roughly 500× the performance of current radiation-hardened space chips. The palm-sized processor is designed to enable AI-driven autonomous spacecraft operations and deep-space exploration without reliance on Earth commanding.
Existing rad-hard flight processors have constrained what onboard autonomy and machine learning workloads spacecraft can run, forcing dependence on ground-in-the-loop commanding for many decisions. HPSC is positioned by JPL as the compute substrate for autonomous operations beyond that latency horizon, including deep-space missions where round-trip light time makes Earth commanding impractical.
Mitsubishi Heavy Industries' AIRIS space-grade MPU detects ships from orbit on RAPIS-4
On May 11, 2026, Mitsubishi Heavy Industries reported that its onboard AI-based object detector AIRIS, built around a next-generation space-grade MPU, successfully detected ships from imagery captured in orbit aboard Japan's Innovative Satellite Technology Demonstration-4 (RAPIS-4) mission.
The demonstration moves ship detection from ground-based post-processing of downlinked imagery to inference performed on the spacecraft itself using a radiation-tolerant MPU. MHI states that ongoing research targets a ground-to-orbit cycle in which AI models are iteratively improved on the ground and redeployed to the satellite for continuous refinement.
NASA/IBM's Prithvi becomes first geospatial AI foundation model run in orbit
On 7 May 2026, researchers from Adelaide University and SmartSat CRC uploaded and ran NASA and IBM's open-source Prithvi Geospatial AI foundation model on two in-orbit platforms, including one aboard the International Space Station. It is the first geospatial foundation model demonstrated in orbit.
Prithvi was trained on 13 years of Earth observation data and is released as an open-source model by NASA and IBM. The in-orbit demonstration ran the model on hardware already operating on the ISS and a second hosted platform, executing inference onboard rather than on the ground.
Running Prithvi in orbit enables onboard inference for Earth observation tasks, shifting work that previously required downlinking raw imagery to ground segments.
SOCOM taps SkyFi to prototype a sovereign tactical EO imagery platform
On May 11, 2026, U.S. Special Operations Command selected SkyFi to develop an early prototype of a sovereign intelligence platform designed to give troops faster, simpler access to geospatial imagery.
The award positions SkyFi — a commercial Earth-observation marketplace — as the prime on a tactical EO tooling effort rather than a pure data supplier. The stated goal is quicker and easier imagery access for operators in the field, packaged as a sovereign platform for SOCOM.
Till next time,