A space data center is an orbital satellite or satellite cluster carrying high-density Artificial Intelligence (AI) compute hardware powered directly by solar energy in space. Instead of housing racks of servers inside terrestrial warehouses connected to electrical utility grids, an orbital data facility operates above the atmosphere, processing computational tasks in Low Earth Orbit (LEO).
As terrestrial data facilities face electricity grid limits, land constraints, and lengthy utility approval timelines, researchers are evaluating whether deploying compute clusters into orbit can bypass ground-based bottlenecks.
While orbital facilities eliminate the need for heavy battery storage by capturing constant sunlight in specific orbits, the sector remains in early research stages with no commercial operations running in space today.
What a Space Data Center Means
A space data center is an orbital computing platform that runs artificial intelligence workloads on satellites equipped with specialized processor chips. Rather than serving purely as communications relays or Earth-observation platforms, these satellites carry server hardware configured to run machine learning training and inference tasks above the planet.
In an orbital architecture, individual satellites function as modular server nodes. Clusters of these satellites maintain coordinated flight formations and communicate with each other using free-space optical links. These laser cross-links move data between orbital nodes without passing traffic through terrestrial ground stations, creating a high-speed local network in orbit.
The operational concept resembles clusters in LEO satellites, but the primary payload shifts from radio transponders to processing silicon. Ground stations transmit raw training data or compute jobs up to the constellation via high-bandwidth optical uplinks, the orbital cluster executes the computation, and the processed output returns to ground terminals.
Why Build a Space Data Center in Orbit
Placing compute infrastructure in orbit gives processors continuous access to unfiltered sunlight while removing the land and electrical grid bottlenecks that constrain ground facilities. Ground-based artificial intelligence facilities require hundreds of megawatts of electricity and square miles of land, creating intense competition for regional power grids.
In a dawn-dusk sun-synchronous orbit, a satellite travels along the terminator line between day and night, remaining in perpetual sunlight throughout its journey. Because the solar panels never enter Earth’s shadow, they generate electrical power continuously. The absence of an atmosphere makes these orbital panels up to eight times more productive than comparable solar arrays on the surface of the Earth.
This continuous solar illumination removes the requirement for massive battery storage banks that drive up weight and complexity in terrestrial renewable energy systems. Orbital power systems harvest energy constantly, routing direct current straight from the photovoltaic arrays into compute racks. Similar dynamics drive interest in space-based solar power, where power generation occurs outside atmospheric interference.
Ground construction timelines present a separate incentive for orbital deployment. Developing a new utility-scale data facility on Earth often requires years of environmental reviews, local zoning approvals, and negotiations with regional power utilities for grid interconnections. Deploying hardware into space relies on commercial launch schedules and orbital spectrum coordination, avoiding municipal land-use approvals and terrestrial utility queues.
Google’s Project Suncatcher
Google Project Suncatcher is a research moonshot evaluating how to scale machine learning compute in orbit using solar-powered satellite constellations. Google structured the initiative to investigate whether space-based infrastructure can provide scalable capacity for expanding artificial intelligence workloads.
The project relies on Google’s custom Tensor Processing Unit (TPU) chips, which power its commercial machine learning systems on Earth. Standard commercial electronics degrade quickly when exposed to cosmic rays and solar particles in space. To test resilience, Google placed its Trillium-generation TPUs into a particle accelerator, simulating the radiation environment found in low Earth orbit. The Trillium chips survived the simulated radiation bombardment without physical damage, suggesting that commercial-grade accelerator silicon can function in space without custom silicon redesigns.
To test the concept in space, Google partnered with satellite imaging company Planet Labs to build a prototype satellite, which flew on Transporter-18, with two follow-up satellites planned for 2027. These demonstration satellites test solar power capture, onboard TPU operations, and free-space optical communications in an active orbital environment.
Beyond the initial two-satellite test, Google Research published a concept for an 81-satellite cluster arranged in a coordinated 1-kilometer-radius formation. This proposed cluster would share compute tasks across optical links, behaving as a single distributed supercomputer in low Earth orbit. Google notes that if commercial launch costs fall below $200 per kilogram by the mid-2030s, the operating economics of an orbital compute cluster could become comparable to ground facilities.
Transporter-18 Orbital Data Center Tests
Three payloads launched on SpaceX’s Transporter-18 rideshare from Vandenberg to test key building blocks of orbital compute, spanning power generation, laser transmission, and in-space processing. According to Payload’s reporting, the mission carried experimental hardware from Google, Cowboy Space, and Star Catcher. Earlier timelines placed prototype tests further out, but the current picture puts hardware in orbit now. Google’s minimum viable product satellite flew directly on this SpaceX rideshare mission, with two follow-up satellites scheduled for 2027.
The Google payload, developed with Planet Labs for Project Suncatcher, is a refrigerator-sized spacecraft carrying four Trillium Tensor Processing Units (TPUs). The satellite draws roughly one kilowatt of solar power to process Gemini queries in short bursts, though thermal constraints force the processors to shut down every 15 to 20 minutes to prevent overheating. Google stated its two planned 2027 satellites will operate continuously. In an updated paper published in the journal Joule, Google modeled clusters of 81 satellites operating within a one-kilometer radius, each generating 50 to 100 kilowatts. Ground radiation testing showed silent data errors occurred at a rate of one per three million queries, which Google noted is acceptable for inference tasks but problematic for AI model training.
Power infrastructure accounted for the remaining two tests on the flight. Star Catcher launched its Protostar satellite to deploy a cubesat and beam power directly to standard solar panels in orbit, serving prospective clients including Starcloud and Aethero. Cowboy Space, previously named Aetherflux before its May rebrand, launched Reason-1 to test infrared power beaming from orbit to a ground facility. The company aims to deliver 30 to 100 watts to a 10-to-20-meter ground target, calling it the first commercial kilowatt-class laser built to send energy to Earth. Cowboy Space plans to fly H200 GPUs on Reason-2 in the first half of 2027 using the laser for data links, targeting a proprietary rocket launch in December 2028 with an upper stage designed as a one-megawatt data center.
Orbital processing economics still depend on steep launch cost reductions. Google estimated that if launch pricing reaches $200 per kilogram, putting one kilowatt of orbital solar capacity in space would cost roughly $810 per year, matching ground electricity costs for facilities in the United States. Google’s learning-curve model projects that figure could become achievable by the mid-2030s if Starship reaches a flight cadence of roughly 180 launches per year. Cowboy Space’s chief operating officer claimed the core debate has shifted from whether computing moves to space to how it gets deployed, though commercial viability remains tied to unproven launch prices and operational lifetimes.
Starcloud’s Approach
Starcloud is an aerospace startup developing a roadmap for gigawatt-scale orbital compute clusters dedicated to artificial intelligence. While Google approaches the challenge as an internal research moonshot, Starcloud operates as an independent company focused entirely on commercializing orbital computing infrastructure.
Starcloud bases its operational case on three specific advantages:
- Continuous energy and radiative cooling. Abundant solar radiation provides uninterrupted power without battery storage, while radiative heat dissipation lowers thermal operating overhead.
- Unconstrained physical scaling. Orbital deployments can scale toward gigawatt capacities without competing for municipal land parcels or strained regional electricity grids.
- Bypassing terrestrial permitting. Space deployments avoid the multi-year local regulatory hearings and utility interconnection delays common to terrestrial facility development.
Starcloud was founded by Chief Executive Officer (CEO) Philip Johnston, Chief Technology Officer (CTO) Ezra Feilden, and Chief Engineer Adi Oltean. Oltean brings relevant infrastructure experience, having served as a principal engineer at SpaceX and spent 20 years at Microsoft working on Graphics Processing Unit (GPU) clusters.
Starcloud has attracted venture capital backing from major technology investors, including NVIDIA, Sequoia Capital, and Andreessen Horowitz. Starcloud publishes a technical development roadmap outlining four sequential satellite iterations named Starcloud-1 through Starcloud-4. This roadmap serves as an engineering plan toward gigawatt-scale deployment rather than an operational commercial fleet. For investors tracking market developments in orbital hardware, our guide to space technology stocks covers how public markets value established aerospace firms compared to early-stage ventures.
SpaceX Starmind
SpaceX Starmind is a proposed constellation of up to one million satellites meant to work as an orbital data center for artificial intelligence (AI), and it is still at the proposal stage. SpaceX filed an application with the Federal Communications Commission (FCC) in late January 2026, as reported by SpaceNews. Elon Musk confirmed the Starmind name on X. A filing is an application. No FCC decision on it appears in the coverage used here.
Starmind satellites would run AI inference on board and send the results down to Earth. Starlink works differently: it relays internet traffic between points on Earth. With Starmind, the computing would happen in orbit, and the results would travel to the ground.
The reported hardware details so far:
- Power. The AI1 satellite would initially average about 120 kilowatts (kW) for computing, with up to 150 kW at peak, according to secondary coverage of a Space.com report.
- Compute payload. Reports in August 2026 say SpaceX is partnering with Nvidia on the AI1 payload, using Nvidia Rubin graphics processing units (GPUs) and Vera central processing units (CPUs).
- Manufacturing. SpaceX plans to build the satellites at a new Gigasat factory in Bastrop, Texas.
- Launch. Deployment depends on Super Heavy and Starship, which SpaceX hopes will lower the cost of reaching orbit.
The reported timelines vary. One report says two AI1 prototype satellites launch in early 2027, with volume production targeted for the end of 2027. Another says SpaceX aims to launch its first satellites in 2027. Treat both as reported targets. Neither is a confirmed schedule.
Coverage names several open technical questions: radiation hardening, processor details, and thermal and power management. Commercial and regulatory hurdles typically slow deployments of this kind. The coverage used here does not publish the orbit altitude, inter-satellite link speeds, or satellite cluster size.
Space Data Center Engineering Hurdles
Operating computer hardware in low Earth orbit introduces physical challenges that do not exist in climate-controlled ground facilities. Terrestrial data facilities rely on stable gravity, atmospheric convection, and physical technician access, none of which exist in orbit.
Radiation management remains a primary engineering risk. High-energy protons and cosmic rays cause single-event upsets, which flip bits in memory registers and disrupt active computing calculations. Over extended durations, total ionizing radiation degrades semiconductor materials and causes permanent chip failures. While Google demonstrated that Trillium TPUs survived short-term accelerator testing, multi-year reliability under continuous space radiation requires orbital validation.
Thermal dissipation presents an equally demanding engineering obstacle. Terrestrial servers reject thermal energy into ambient air or circulating chilled liquid. In the vacuum of space, conduction and convection do not function because there is no surrounding atmosphere to absorb heat. An orbital compute platform must rely entirely on radiative cooling, transferring heat through circulating fluid loops out to broad radiator panels that radiate infrared energy into space. These thermal radiators add substantial surface area, structural mass, and mechanical complexity to the satellite chassis.
Maintenance constraints further separate orbital facilities from surface infrastructure. When a power supply, memory module, or cooling pump fails in a ground facility, an on-site technician swaps the component within minutes. In orbit, physical servicing is impossible with current operational technology, meaning every satellite must incorporate component redundancy or risk total system failure. These operational challenges parallel the manufacturing hurdles explored in space mining and manufacturing, where machinery must operate autonomously for years without human intervention.
Launch pricing represents the dominant economic variable. Transporting heavy compute hardware, solar arrays, and thermal radiators into orbit requires substantial upfront capital. Even with declining launch rates, lifting thousands of metric tons of computing infrastructure into low Earth orbit remains expensive compared to building steel-frame buildings on Earth.
How Close Is This to Reality
As of 2026, the orbital computing sector remains in a research and prototyping phase, with no commercial space data centers actively operating in orbit. The concept has progressed from theoretical white papers into active laboratory hardware testing, but physical execution in space remains ahead.
The industry roadmap centers on initial orbital demonstrations. The partnership between Google and Planet Labs flew a first experimental satellite on Transporter-18 and plans two follow-ups in 2027, which will provide empirical data on how modern machine learning processors handle real space conditions. Starcloud continues engineering work along its phased roadmap, but has not deployed commercial compute capacity to orbit.
The commercial viability of orbital compute will depend directly on flight data from these upcoming test missions and the trajectory of heavy-lift launch costs. To evaluate the viability of a space data center, track technical milestones from orbital prototype launches and monitor launch cost reductions on public mission manifests.