Satellite InferenceAI Inference in OrbitGet in touch 
AI INFERENCE IN ORBIT
Satellite Inference is building the software and flight hardware to run AI models in orbit. We start with data that is already in space, and build toward solar-powered compute as launch and spacecraft costs fall.
Targets and model results, not completed milestones. See the program and its gates
01 / WHY NOW
AI hardware already runs in space. The open questions are reliability, the cost of each step and which customers come first.
Data centers used about 415 TWh of electricity in 2024. The International Energy Agency's base case reaches around 945 TWh by 2030, and new AI capacity increasingly waits for generation and grid connections.
Source: International Energy Agency (IEA), Energy and AI (April 2025). Accessed October 2026.Embedded GPUs already process data on satellites. In November 2025 a data-center-class GPU reached orbit, and by December it was running AI models. Several companies now plan commercial orbital compute.
Sources: Starcloud-1 mission updates (November and December 2025); SpaceX and Blue Origin FCC filings (2026). Accessed October 2026.Launch price to the operating orbit, spacecraft cost per watt and mass per kilowatt. Orbit moves closer to ground cost only as all three fall together, so each of our steps has to be useful on its own.
Satellite Inference engineering and cost model v3.1.02 / THE ORBIT
A dawn–dusk sun-synchronous orbit follows Earth's terminator, so a solar array faces the Sun almost all year. Eclipses are seasonal and short: in our model, under 20 minutes, for about three months a year.
That is roughly 1.4 to 1.5 times the sunlight of the low orbits most satellites use today, which is why our future nodes are designed around it.
Fraction of a year a satellite spends outside Earth's shadow. Screening model (2027 solar geometry, cylindrical Earth shadow); not mission analysis. The SI-P1 pathfinder flies in its host's orbit.
Sun-synchronous, ascending node at 18:00. Our model orbit for SI Nodes. Year runs the seasons; drag to turn the view. Earth imagery: NASA.
Spacecraft safety, power and communications define the limits for the compute payload.
Software that runs approved AI workloads on spacecraft accelerators within the host's power, thermal and contact limits.
A compact hosted payload built around Thor-class embedded compute.
Dedicated solar compute spacecraft, aggregated into clusters as earlier missions establish the case.
Functional overview. No flight hardware is depicted.
03 / WHAT WE BUILD FIRST
SI-P1 is a hosted demonstration of AI inference on NVIDIA Jetson Thor-class compute, with an Orin fallback. Before it flies, Orbit-0 targets running our runtime on a GPU that is already in orbit.
04 / THE ECONOMICS, PLAINLY
Our cost model prices the three numbers that decide the cost of computing in orbit: launch to the operating orbit, the satellite's cost per watt and its mass per kilowatt. It then compares a GPU-hour in orbit with the same GPUs in a US data centre.
Launch is falling fastest. Google's Project Suncatcher study projects launch prices below $200/kg around 2035; at that price, launching a satellite costs roughly what powering an equivalent data centre on the ground does, per kilowatt-year. Our model agrees on launch, then adds the satellite and the chips, which that comparison leaves out: an orbital GPU-hour still costs about 1.5 times a terrestrial one.
Speed comes first: data that is already in space becomes results on board and reaches users sooner, with less downlink. Capacity comes next. On the ground, power has become the bottleneck: the IEA estimates that around one in five planned data-centre projects could be delayed by grid constraints, and new transmission lines take four to eight years in advanced economies. In orbit, each node brings its own power; the limit there is launch.
Each larger step, Node-1 and then clusters, needs its own funding and goes ahead only when launch and spacecraft prices are contracted rather than forecast.
All-in cost: GPUs, power, cooling, launch to a dawn–dusk orbit and the satellite, annualised at 10% over five years; 1.2 kW per GPU, 70% utilisation and 15% spare GPUs in orbit. Sources: Satellite Inference cost model v3.1; Google Research, Project Suncatcher (November 2025) and arXiv:2511.19468; SpaceX rideshare list price, $7,000/kg (2026); US EIA Short-Term Energy Outlook (September 2026), industrial electricity 9.0¢/kWh; IEA, Energy and AI (April 2025), grid constraints.
Follow the costs year by year to 2045 Size a node: power and heat Fly a cluster: formation and control05 / FOR EARTH-OBSERVATION OPERATORS
We are looking for design partners with a specific task, accessible data and a reason to process it onboard. Together, we would compare the result with your current ground workflow.
Discuss a design-partner evaluationAgree the model, data rights, quality threshold and delivery deadline that matter to your mission.
Measure useful results, resources and integration requirements on representative hardware.
Work with the host operator on an approved deployment, then evaluate the results together.
The benefit depends on the workload and available communication windows. Design partners and host agreements are being sought.
06 / THE DEVELOPMENT ROADMAP
Target windows, conditional on technical results, host access, customer demand and new financing at each stage.
Develop the software, pursue Orbit-0 on an existing orbital GPU and prepare flight-ready pathfinder hardware.
Gate: host access, qualification evidence and flight fundingDemonstrate inference on the hosted payload and evaluate the mission with customers.
Gate: host manifest and a separately funded flightDevelop the first 10–20 kW SI Node around a purchased spacecraft bus and a defined customer workload. Node-1 flies alone; the 100 kW node modeled on our power, heat and economics pages is the building block of later clusters.
Gate: pathfinder evidence and dedicated project capitalTarget an initial MW-class cluster in 2031–32, with a longer-term ambition of 100 MW-class capacity by 2035.
Gate: launch capacity, economics and customer commitmentsLaunch-agnostic by design: a 20 kW node is planned at under a tonne, within rideshare and medium-lift capacity. Cluster economics improve sharply with heavy-lift prices, and only the largest clusters need heavy-lift vehicles.
See the program and its gates07 / THE TEAM
Orbital compute is, at scale, a power-infrastructure business. The founder leads the company, financing and customers; the fractional CTO brings twenty-five years of space electronics; a full-time technical cofounder will lead the build.
FOUNDER & CEOLeads strategy, financing, partnerships and customer development, and coordinates the SI-P1 engineering program. Hands-on experience in data centres, including server and network infrastructure.
Contact the founder
FRACTIONAL CTOTwenty-five years in space electronics. Founder and CEO of Spacechips (UK); previously headed mixed-signal space electronics at Airbus Defence & Space for twelve years. PhD, University of Edinburgh. Leads SI-P1 architecture, radiation risk, qualification planning and technical review.
LinkedIn profileYou have taken a spacecraft payload or avionics system through integration, qualification and flight, and want to build solar-powered compute in orbit full-time.
Read the roleThe fractional CTO role is agreed; the appointment takes effect when Satellite Inference Inc. is formed and the agreement is signed. Advisors, customers and partners are named only once agreements are signed.
More about the teamJOIN THE FIRST MISSION
We are developing SI Runtime, an Orbit-0 experiment and a flight-ready SI-P1. We are looking for host operators, design partners and a technical cofounder.