In 2026 orbital data centres stopped being a thought experiment and became a funded industrial programme. Starcloud closed a $250M Series A extension at a $2.3B valuation, roughly double its March 2026 mark, with Nvidia and Cisco Investments joining. SpaceX filed with the FCC for up to one million AI1 “Starmind” satellites targeting 120 GW and between 10 and 100 million frontier-class GPUs. Cowboy Space leased a 291,035-square-foot plant in Kent, Washington, and Google’s Project Suncatcher modelled an 81-satellite cluster with Planet.
None of that softens the physics. With no atmosphere, only radiative cooling remains. SpaceX’s AI1 needs roughly 160 m² of deployable radiators for a 250 kW payload. BCG puts 20-year ownership at $660 to $750M per MW against $230 to $300M terrestrial, a 2.5× to 3× premium that narrows only to about 1.5×. The land rush is real and funded, but heat rejection, the cost premium and unproven chip survival decide what is buildable. This page orientates you and routes you to the article that answers each thread.
In This Series
- What Orbital Data Centres Are and Why 2026 Changed the Calculus: the definition, the architecture, and the funded 2026 evidence.
- The Physics Tax Behind Cooling Orbital Data Centres: why heat rejection in a vacuum is the load-bearing constraint.
- How Eclipse Cycles, Orbit Choice and Radiation Shape Orbital Hardware: what the orbital environment does to power and silicon.
- What Orbital Data Centres Really Cost Compared with Terrestrial: the 20-year cost premium and the path to parity.
- Build Versus Buy for Orbital Compute and How to Vet Vendor Claims: the build/buy/wait/pilot decision and how to vet vendors.
What are orbital data centres, and why is 2026 the inflection point?
An orbital data centre is a dedicated compute facility in orbit, not a communications or imaging satellite: solar arrays generate power, accelerators run AI workloads, radiators reject heat, and optical or RF links move data. 2026 is the inflection point because the category crossed from concept to funded programme, and the evidence behind it is now checkable and worth your attention.
That distinction changes what you are buying. An ordinary satellite relays or observes; an orbital data centre turns the payload into compute throughput, running a four-part solar-to-downlink chain. Starship is why payload-mass assumptions are shifting, and low Earth orbit is the default regime. That shift is unpacked in the orbital data centres explainer.
The land rush sits against two constraints that temper it: the heat-rejection problem and the 20-year cost premium.
Why is cooling the load-bearing constraint on orbital compute?
Space has no atmosphere, so the conduction and convection that cool a terrestrial facility disappear. The only path left is radiation, which scales with the fourth power of temperature under the Stefan-Boltzmann law. That makes radiator area the scarce resource: SpaceX’s AI1 needs roughly 160 m² of radiators for a 250 kW payload, and that area sets the ceiling on everything else.
Radiator area then propagates into spacecraft mass, drag, station-keeping and launch mass, the physics tax on compute density. We unpack why radiative cooling pays a penalty versus convection in full, and it is the origin of the cost premium you pay downstream.
What does the orbital environment do to hardware and power?
Three environmental forces shape every hardware and power decision. Eclipse cycles put LEO satellites in shadow for 30 to 40 minutes of each 90-minute orbit, draining batteries and dictating AI duty cycles. Orbit choice trades eclipse against other hazards. Ionising radiation degrades commercial GPUs through cumulative dose, single-event upsets and latch-up, forcing a rad-hard versus COTS choice with no free answer.
Solar arrays produce nothing in shadow, so bursty, schedulable inference fits better than continuous training. Dawn-dusk sun-synchronous orbit keeps a satellite in near-continuous sunlight, but it is not free. The full picture is in hardware survival in orbit, and it pairs with radiative cooling as the second environmental constraint.
What do orbital data centres really cost against terrestrial infrastructure?
BCG’s modelling puts 20-year ownership at $660 to $750M per MW in orbit versus $230 to $300M terrestrial, a 2.5× to 3× premium narrowing only to about 1.5×. Launch cost per kilogram is the dominant lever, flowing into every gram of structure, radiator and shielding. Parity is plausible around 2040, on falling launch prices and rising terrestrial constraints, independent of vendor promises.
Starship-driven reusability is the clearest credible path that shifts launch cost, which makes the economics a bet on the cost curve. Analyst timelines differ — BCG sees the premium narrowing toward 1.5× by the mid-2030s, while SemiAnalysis models parity nearer 2040 — so treat parity as a scenario, not a forecast. The full cost comparison walks the numbers, and it feeds the build versus buy decisions you make next.
Should you build, buy, wait, or pilot orbital compute?
For most businesses the answer is buy or pilot, not build. Orbital compute suits latency-tolerant, batch or scheduled inference and data already in orbit, rather than interactive workloads. Engage early only where a niche (sovereign compute or in-orbit data) justifies the premium; otherwise wait and watch while terrestrial remains the default. Before any pilot, you should pressure-test vendor economics and radiation claims against evidence.
Due diligence is the counterweight to hype. Pressure-test verifiable per-MW economics, radiator and power budgets, radiation qualification, lifetime and servicing claims, and named pilot references.
The build-versus-buy call turns that into a posture, leaning on the radiation reality and cost parity evidence from the cluster.
Resource Hub: Orbital Data Centres Deep Dives
Here’s the same cluster reorganised by purpose, with a suggested reading order.
The Foundations: What It Is and Why It’s Hard
- What Orbital Data Centres Are and Why 2026 Changed the Calculus: the definition, the solar-to-downlink architecture, and the funded 2026 evidence. (5 min read)
- The Physics Tax Behind Cooling Orbital Data Centres: why radiative cooling in vacuum is the binding limit. (5 min read)
- How Eclipse Cycles, Orbit Choice and Radiation Shape Orbital Hardware: what eclipse, orbit choice and ionising radiation do to your power budget and silicon. (6 min read)
The Economics: What It Actually Costs
- What Orbital Data Centres Really Cost Compared with Terrestrial: the 2.5× to 3× premium, why launch cost dominates, and what must be true for parity around 2040. (6 min read)
The Decision: Build, Buy, or Wait
- Build Versus Buy for Orbital Compute and How to Vet Vendor Claims: the build/buy/wait/pilot calculus and how to pressure-test vendors before a pilot. (6 min read)
Suggested reading order: foundations first (what it is, then the physics, then the environment), then the economics, and finish with the decision.
Frequently Asked Questions
Is it true that space is cold, so cooling orbital data centres should be easy?
No. Space is cold, but a vacuum has no air to carry heat away, so the only cooling path is radiating heat to the black sky. That means large deployable radiators, not a simple fan. SpaceX’s AI1 needs roughly 160 m² of radiators for a 250 kW payload, which makes radiator area the binding limit. Cold surroundings do not help when there is nothing to conduct into.
Can orbital data centres train AI models, or are they only good for inference?
Mostly inference. Eclipse cycles put low Earth orbit satellites in darkness for roughly a third of each orbit, so bursty, schedulable workloads fit better than the continuous power draw of training. A 90-minute orbit carries 30 to 40 minutes of shadow. Training a frontier model in orbit is theoretically possible, but the duty cycle and radiator area make it a poor early use case.
What happens if a GPU fails in orbit? Can it be repaired?
Right now, it is usually lost. Orbital data centres have no routine servicing, so a failed accelerator is typically written off, which is why redundancy and radiation tolerance matter so much. Servicing is being explored but remains unproven at scale. That failure cost is baked into the 20-year ownership numbers, part of why the premium sits at 2.5x to 3x.
How does data get to and from an orbital data centre?
Through optical or RF downlinks, the final stage of the solar-to-downlink chain. Optical links promise higher bandwidth, while RF is more proven. Either way, the ground segment and link budget become part of your architecture, not an afterthought. This is also why workloads with data already in orbit, such as Earth observation processing, can be attractive: you avoid shipping the raw data down.
Do orbital data centres replace terrestrial data centres?
No. They are a niche complement, not a replacement. Terrestrial facilities enjoy cheap power, effectively unlimited radiator area, and easy maintenance, so they will keep handling most workloads. Orbital compute makes sense only where latency tolerance, sovereign isolation, or data already in orbit justify the premium. Think of it as an extension of the compute estate for specific jobs, not a rival to your existing data centre.
What launch price makes orbital compute competitive?
Launch cost per kilogram is the dominant lever, and it flows into every gram of structure, radiator and shielding. BCG’s modelling narrows the 20-year premium from 2.5x to 3x down to about 1.5x as launch prices fall. Full parity is plausible around 2040, but it depends on Starship-driven reusability actually delivering, not on vendor promises.
How are orbital data centres powered during eclipse?
Solar arrays generate power in sunlight, and batteries carry the load through eclipse. In low Earth orbit that shadow lasts 30 to 40 minutes of each 90-minute orbit, roughly a third of the time, so the battery and duty cycle shape what you can run. Dawn-dusk sun-synchronous orbits keep a satellite in near-continuous sunlight, but they constrain where you can fly and what you can point at.
Which AI workloads make sense in orbit today?
Latency-tolerant, batch or scheduled inference, and processing for data already in orbit. Earth observation analytics is the clearest example: analyse imagery where it is captured and downlink only the results. Interactive workloads such as real-time chat or fine-tuning loops are a poor fit because of link latency and the eclipse-driven duty cycle. If your job can wait and is compute-heavy, it is a candidate.
Is radiation a dealbreaker for commercial GPUs in space?
Not a dealbreaker, but a real constraint with no free answer. Ionising radiation degrades commercial chips through cumulative dose and causes single-event upsets, so you choose between radiation-hardened parts and commercial off-the-shelf chips with mitigation. Rad-hard is slower and costlier; commercial is faster and more fragile. The right choice depends on orbit, mission life and how much redundancy you can afford.
What are the biggest red flags in an orbital compute vendor’s pitch?
Vague per-MW economics, missing radiator and power budgets, and radiation claims with no qualification data. Ask for verifiable lifetime and servicing assumptions, named pilot references, and evidence behind any parity promise. If a vendor will not quantify heat rejection or launch mass, the numbers do not add up. Pressure-test claims against the physics and cost evidence before committing to a pilot.
How is an orbital data centre different from a satellite like Starlink?
A communications satellite relays signals and an imaging satellite observes, but an orbital data centre turns its payload into compute throughput. Solar arrays generate power, accelerators run AI workloads, radiators reject heat, and optical or RF links move data. That four-part solar-to-downlink chain is what separates it from a relay or a sensor. Some constellations, like SpaceX’s proposed AI1 Starmind, are designed around compute.
Can I buy orbital compute capacity right now?
Not as a mainstream product. In 2026 the category is a funded industrial programme with pilot projects, not an off-the-shelf service, so early access runs through partnerships and niche arrangements. For most organisations the sensible posture is to pilot where a genuine niche exists, or wait and watch. Engage early only where sovereign compute or in-orbit data justifies the premium.
