So what happens when the two get combined, and the servers themselves move to orbit? For the broader picture of the 2026 shift, see our pillar overview of orbital data centres.
That question stopped being hypothetical in 2026. AI compute demand is colliding with the hard limits of building on the ground: land, power, water and local opposition. Meanwhile, a wave of funding and FCC filings has moved data centres in space from science fiction to a funded spending decision. By the end you’ll be able to define an orbital data centre, trace how it works from solar array to downlink, and say why 2026 became the inflection point. We’ll leave build instructions to the physics and cost articles in this series.
What is an orbital data centre, and how does it differ from an ordinary satellite?
Start with the definition. An orbital data centre is a spacecraft whose primary payload is compute: servers, GPUs or TPUs, plus storage and networking, riding on a standard satellite bus with solar arrays, thermal management, propulsion and communications. Unlike a relay or imaging satellite, it exists to process data in orbit. A communications satellite bounces signals between two points like a bent pipe, and an imaging satellite collects Earth observation data; neither does any processing up there.
That processing payload is the whole distinction. The bus disciplines (power, thermal control, attitude, comms) are shared with every other satellite. What changes is the mission: compute throughput and data processing rather than relay or sensing. Compute-dense modules need higher heat loads and a different power distribution architecture than a modest communications payload.
The default home is Low Earth Orbit (LEO), up to roughly 2,000 km. LEO keeps latency to Earth low and costs less to reach than higher orbits, and some orbits offer near-continuous sunlight. The terminology is still settling; “space-based data centre”, “data centre in space” and “orbital compute” describe the same category, but the definition that matters is the payload. For the wider 2026 picture, see our overview of the space-computing land rush.
How does a space-based data centre actually work, from solar array to downlink?
A space-based data centre is a four-part chain: solar arrays generate the power, the compute payload does the processing, deployable radiators reject the waste heat, and RF downlink or optical inter-satellite links move the data. Each step is simple on its own, but the vacuum complicates every one of them.
Solar is the easy win. Sunlight in orbit delivers about 1,361 watts per square metre, and operators favour dawn-to-dusk, sun-synchronous orbits for near-continuous light. The catch: a LEO satellite spends about a third of each orbit in shadow, so batteries carry the dark minutes. Supporting AI workloads through an eclipse needs battery capacity five to ten times today’s space-grade cells, which is why orbit choice matters. Eclipse cycles and radiation get their own article in this series.
The hard step is cooling. On Earth you move heat into air or water. A vacuum has neither, so conduction and convection are off the table and the only route out is thermal radiation. Run more compute and you launch more radiator: a single H100-class chip needs roughly 1.4 square metres of radiator, and large radiators are only plausible because Starship-class launch has made heavy arrays cheap enough to orbit. That is the load-bearing constraint, and the full detail lives in the physics article.
Data movement is a trade-off. Optical inter-satellite links suit a high-bandwidth satellite-to-satellite mesh, while RF downlink is the slower but dependable pipe back to ground stations. Some systems combine both: Axiom’s nodes take in data over optical links and send results home through relay constellations.
How real are orbital data centres in 2026, and what changed?
In 2026 orbital data centres moved from research papers to funded, filed programmes. Four commitments mark the shift; for how they slot into the wider 2026 land rush, see the pillar overview.
Starcloud is a concrete proof point. In November 2025 it launched a satellite and ran AI workloads in orbit, including training a model. Then in August 2026 it closed a $250 million Series A extension at a $2.3 billion valuation, roughly double its March valuation, with Nvidia and Cisco Investments joining the round.
SpaceX is the scale signal. In early 2026 it filed with the FCC for a constellation called Starmind, whose first satellite, the AI1, is disclosed at 120 kilowatts of average compute payload and 150 kilowatts peak. The filing asks for up to one million satellites, a scale statement of intent. You can read it yourself in the FCC’s filing system under ID SAT-LOA-20260108-00016.
Google anchors the research side: Project Suncatcher, run with Planet Labs, models an 81-satellite cluster flying in formation about 650 km up, with a two-satellite mission planned for early 2027.
Cowboy Space adds the industrial footprint, leasing a 291,000-square-foot Kent, Washington warehouse to build the hardware, turning the paper filings into a factory floor.
None of that means the category is settled. Funded industrial programmes now sit alongside a lot of vendor noise, and the constraint that separates the two is the physics of heat rejection and a cost premium that is still real. If you want to separate the substance from the noise, we have a guide to vetting vendor claims.
So where does that leave orbital data centres?
Strip away the branding and an orbital data centre is simple to describe: a compute payload riding a four-part chain where the cooling step is hard. Starcloud, Starmind, Project Suncatcher and the Cowboy Space factory are commitments to solve the hard steps now, backed by capital and filings that did not exist a year ago.
That changes how you should file this: under funded, constrained infrastructure. The open questions are whether the physics of heat rejection and a long-term cost premium can be beaten. Those are the questions the rest of this series picks up. For now, the summary is that 2026 moved orbital data centres from a white paper to a real budget line, and that is a signal worth your team’s attention. For the full 2026 picture, the pillar ties the programme, the physics and the costs together.
Frequently Asked Questions
Is an orbital data centre just a satellite with a server bolted on?
No. A satellite with a server bolted on would still be a relay or sensing platform. An orbital data centre is defined by its payload: the mission is compute throughput and data processing, not communications or Earth observation. It shares the standard spacecraft bus (power, thermal control, attitude, comms), but that bus exists to serve the servers, GPUs or TPUs rather than a bent pipe.
Why can’t orbital data centres use air conditioning or water cooling like a regular data centre?
Because a vacuum has no air or water to carry heat away. On Earth, conduction and convection move waste heat into air or liquid; in orbit, the only route out is thermal radiation, which is far slower per unit area and offers no free medium to work with. That is why orbital data centres depend on large deployable radiators, and why heat rejection is the load-bearing engineering constraint.
What happens to an orbital data centre during an eclipse?
It keeps running on batteries. LEO satellites spend roughly a third of each orbit in Earth’s shadow, when solar arrays generate nothing, so power budgets have to cover compute, thermal control and communications from stored energy. Eclipse cycles are one reason operators favour dawn-to-dusk or sun-synchronous orbits, which maximise sunlight and keep the power profile predictable.
Can orbital data centres replace terrestrial data centres?
Not on current evidence. The physics of heat rejection and a long-term cost premium mean orbit is a complement, not a replacement, at least for now. The realistic near-term role is niche, latency-tolerant or bandwidth-heavy workloads where launch economics and orbital constraints make sense, rather than hosting the general-purpose compute that dominates terrestrial halls.
Are orbital data centres being used for AI training or only inference?
Both are on the table, and the strongest early proof points lean toward frontier AI. Starcloud ran an Nvidia H100 in orbit, and SpaceX’s Starmind (AI1) filing targets 10M to 100M frontier-class GPUs. Training and inference each carry different trade-offs: training stresses inter-satellite bandwidth, while inference leans on downlink capacity. The four-part chain, not the workload label, sets the ceiling.
Where can I find SpaceX’s FCC filing for the Starmind (AI1) constellation?
Through the FCC’s public filing database, where SpaceX submitted the Starmind (AI1) application. Search for SpaceX and Starmind in the commission’s Space Bureau filings to read the constellation’s stated scale, including up to 1,000,000 satellites and its 120 GW target. The filing itself is the primary source, so treat second-hand summaries with appropriate caution.
Where can I read Google’s Project Suncatcher research paper and design details?
Google published Project Suncatcher as a research paper, which you can find through Google Research’s publication channels and the usual preprint archives. It models an 81-satellite cluster developed with Planet Labs and is the clearest public look at how a research programme reasons about orbital compute architecture, so read it alongside the funding announcements rather than instead of them.
Where can I find funding and market-size data on orbital data centre companies?
Start with primary funding announcements and regulatory filings, then cross-check against reputable market databases. Starcloud’s $250M Series A extension at a $2.3B valuation, with Nvidia and Cisco Investments joining, is a concrete 2026 data point. Vendor press releases tend to inflate totals, so favour disclosed valuations and filings over headline market-size forecasts.
Optical inter-satellite links versus RF downlink: which makes sense for orbital compute?
They do different jobs, so most systems will use both. Optical inter-satellite links (200+ Gbps demonstrated) suit a high-bandwidth satellite-to-satellite mesh, moving data between nodes without touching the ground. RF downlink is slower but robust and well understood, making it the dependable pipe back to ground stations. Think mesh for internal traffic, RF for the final hop.
How much latency does an orbital data centre add to a workload?
It depends on the path, not just the orbit. LEO’s low altitude is chosen precisely to keep round-trip latency to Earth small, but inter-satellite hops and ground-station routing add their own delays. The honest answer is that orbit suits workloads tolerant of tens of milliseconds, not the microsecond-scale coordination some terrestrial clusters are built around.
Do orbital data centres contribute to space debris, and do they have to deorbit?
Operators are expected to plan for disposal, and LEO’s low altitude helps because satellites re-enter naturally within years rather than lingering for centuries. A constellation on the scale of Starmind (AI1) still raises real congestion and collision-tracking concerns, which is one reason regulators scrutinise filings on disposal and collision avoidance before approving that kind of scale.
What is the biggest thing that could stop orbital data centres taking off?
The same two constraints the 2026 land rush is spending against: physics and cost. Rejecting heat in a vacuum is genuinely hard, and orbit carries a long-term cost premium that terrestrial facilities do not. If radiators, radiation-tolerant silicon and launch economics improve faster than those costs bite, the category grows; if not, 2026’s funding becomes the high-water mark.
