Every headline quotes a different premium for orbital compute. BCG puts the 20-year gap at 2.5× to 3×, SemiAnalysis says more than 4×, and Wood Mackenzie prices a 1 GW orbital build at more than three times the terrestrial equivalent. Different scopes and metrics, one lever: launch cost per kilogram.
Orbital compute is not cheaper today, whatever the 2026 space-computing land rush suggests. A 20-year total cost of ownership (TCO) lands around $660 to $750 million per MW in orbit against $230 to $300 million per MW on the ground, a 2.5× to 3× premium. What follows is why that number is what it is, and when the gap might close.
Why are launch costs the single biggest driver of orbital data centre economics?
Launch cost per kilogram is the biggest driver because it taxes every mass-scaled subsystem. The satellite bus, radiators, solar arrays and shielding all scale with mass, so cheaper launch lowers every upstream line at once. SemiAnalysis breaks out a 30.5 kW B300 cluster where launch is about $1.6M of a ~$3.1M data-centre capital bill; across the full TCO launch is about one-fifth, GPUs roughly half.
Radiators and solar arrays consume 65 to 70% of a satellite’s total mass, and a 100 kW satellite needs a 400 m² radiator. That is the physics tax: no air in space to carry heat away, so cooling hardware grows with compute, and every extra kilogram is another you pay to launch.
The economics debate turns on launch cost. Falcon 9 charges roughly $1,400 to $1,800/kg. Forethought’s Starship model runs from about $475/kg expendable to $93/kg for early full reuse and $19/kg on a bullish case. Space-based solar reaches energy parity near $250/kg, orbital data centres near $100/kg, and at roughly $50/kg orbital power beats every terrestrial energy source.
Model TCO as a continuous function of $/kg. When the launch line moves, the premium and parity date move with it, and the architecture being costed rides the same curve. That curve is what sets the headline premium.
Orbital versus terrestrial data centres: which is cheaper, and how much more expensive per MW?
Terrestrial data centres remain cheaper, for now. BCG’s 20-year ownership figures put the orbital premium at 2.5× to 3×. The catch: you pay that premium for orbit-advantaged compute (sovereign AI, latency-tolerant inference, workloads acting on space-generated data), and it buys you nothing on general-purpose capacity.
Know what the comparison is measured against. SemiAnalysis decomposes the terrestrial baseline into a four-layer supply stack: grid-connected capacity around $12 to $15M/MW, converted capacity and powered land around $10 to $15M/MW, behind-the-meter generation at $15 to $20M/MW, and industrial production beyond $20/MW, with a universal semiconductor layer on top. Orbital costs get compared to whichever layer you pick, one reason the headlines disagree.
The other reason is the metric. SemiAnalysis puts orbital levelised cost of compute (LCOC) at $10.91/hr/GPU versus $2.49/hr/GPU terrestrial, or $0.73 versus $0.17 per PFLOP-hour. LCOC always sits above raw TCO because it grosses up for radiation-driven availability losses and redundant GPUs.
That gross-up explains the highest quoted premiums. SemiAnalysis attributes it to a five-year orbital hardware life against fifteen years terrestrial, which pushes levelised data-centre capex to 17 to 18× terrestrial. Financing compounds it: the same model starts orbital deployments at a 15% cost of capital, declining to 10.3% over a decade, against a flat 10.3% on the ground. Those same cost lines set the parity date.
Space versus Earth data centres: when do their costs actually reach parity?
The parity window is around 2040, but it is a conditional scenario. SemiAnalysis’s base case has orbital levelised costs starting at roughly four times terrestrial in 2026 and only dropping below terrestrial cost around 2040. Its aggressive case pulls that into the early 2030s, when space could be only about 30% more expensive; the sceptical camp says never.
Four things have to land together: launch at ~$100/kg or below on full Starship reusability, terrestrial electricity and land costs still climbing, more efficient radiators, and more compute per kilogram launched. BCG’s realistic trajectory narrows the premium to about 1.5× over the next decade, and even an aggressive full Starship case lands at 1.1× to 1.2×, still above terrestrial.
The accelerants sit on the ground. Grid-interconnection queues in PJM now run about seven years, and behind-the-meter power no longer undercuts grid supply by much. Those constraints, plus the sovereign-AI demand surge, make orbital overflow capacity interesting for your roadmap.
The delayers are launch and hardware. Starship has flown 11 times, the ship has not yet demonstrated re-flight, and FAA approvals cap it at 145 launches a year, well below the cadence a buildout needs, a ceiling that would have to rise roughly 35-fold. Cheap ground power or failure-rate surprises push parity past 2040 or remove it entirely. That changes your build-versus-buy decision, hinging on how radiation shapes the silicon. Track the lever.
How do I build a defensible 20-year TCO comparison between space and terrestrial compute?
A defensible comparison lives on its scope. Hold a 20-year horizon, the same workload and SLA, and the same cost-of-capital treatment on both sides, normalised on a levelised metric such as $/hr/GPU or $/PFLOP-hour. Then expose your sensitivities.
Start with the four that matter most: launch cost per kilogram, satellite mass, electricity price, and hardware lifetime. Add the failure rate: BCG’s sensitivity runs swing the premium from 1.8× at a 30% failure rate to 1.5× at 10% under the same launch assumptions. Published rates run anywhere from 5% to 30%, and a 9% annual chip bleed, drawn from Meta’s observed failure rate, imposes a 38% non-chip premium over a five-year lifecycle. Model it as a range.
Then price the layers most models drop: financing and WACC, insurance, deorbit, and licensing. In-orbit servicing is still nascent, so redundancy replaces maintenance, and GAO notes data centres may be decommissioned more often than other satellites.
For inputs, BCG supplies the per-MW premium, SemiAnalysis the LCOC and compute economics, and GAO an independent cross-check. Test any vendor TCO claim against those sources and report ranges rather than point estimates. If a number will not name its launch and failure assumptions, treat it as marketing.
The premium is a launch-cost tax that propagates through radiator mass, solar-array mass and shielding, and that one lever explains both the 2.5× to 3× gap today and the conditions under which it could close. Parity is a conditional scenario: it arrives only if launch falls below ~$100/kg while ground energy and land costs keep rising, and holding any one of those fixed pushes ~2040 out or removes it entirely. Which leaves a discipline. A defensible comparison lives in its scope and in which inputs you trust: BCG, SemiAnalysis, GAO. The next time a headline quotes ‘more than 4×’, ask which launch-cost and failure-rate assumptions it hides. When to engage a provider follows from that answer, and so does the full cost picture.
Frequently Asked Questions
Does cheaper Starship launch alone make orbital data centres cheaper than terrestrial ones?
No. Cheaper launch is necessary but not sufficient. Even at around $100/kg, orbital compute reaches parity only if terrestrial energy and land costs keep climbing, radiators get more efficient, and hardware failures stay low. Launch is the dominant lever, but hold the other assumptions fixed and the premium narrows rather than disappears.
Why do orbital data centres have a shorter useful life than terrestrial ones?
Orbital hardware lives about five years against roughly fifteen on the ground. The gap comes from radiation degradation of silicon, a harsh thermal environment across eclipse cycles, and the near-impossibility of repairing failed components. That shorter life forces the whole capex bill to be amortised over a third of the period, which alone pushes the levelised cost of compute higher.
What is chip bleed, and how much does radiation-driven failure add to orbital costs?
Chip bleed is the steady loss of working silicon to radiation, and it is usually modelled at about 9% of chips per year. That annual bleed imposes roughly a 38% non-chip premium over a five-year lifecycle once you add redundancy to mask the losses. Because published failure rates range from 5% to 30%, model it as a range rather than a single value.
Can orbital data centres be serviced or upgraded once they are in orbit?
Not easily, and that is a cost line most models understate. In-orbit servicing is still nascent, so failed chips and degraded solar arrays are often replaced by redundancy rather than repair. Until servicing matures, treat orbital hardware as effectively disposable, and price the resulting redundancy and replacement costs into the twenty-year total rather than assuming a terrestrial-style maintenance regime.
Why do orbital data centres need such large radiators, and what does cooling add to the bill?
Space has no air to carry heat away, so an orbital data centre must reject it by radiation, and radiator area scales with the mass you launch. That physics tax works through both the radiator hardware and its launch cost, which is why cooling is one of the largest penalties. Improving radiator efficiency directly lowers the parity threshold.
How does the cost of capital change the orbital-versus-terrestrial comparison?
It widens the gap. Orbital deployments currently carry a higher weighted average cost of capital, around 15% declining toward roughly 10.3%, because the technology is immature and unproven at scale. Terrestrial facilities borrow against decades of operating history and cheaper financing. Over a twenty-year horizon that financing premium compounds, so it belongs in the model alongside hardware and energy costs.
What workloads actually justify paying the orbital premium?
Orbit-advantaged compute, not general-purpose capacity. That means sovereign AI deployments that need jurisdiction-free processing, latency-tolerant inference and training, and workloads acting on data already generated in space, such as Earth observation. If a workload can run cheaply on the ground, the premium buys nothing, so the orbital case rests on strategic or physical advantage rather than raw $/PFLOP-hour.
What happens to the cost comparison if launch falls to $100 per kilogram?
The premium narrows sharply but does not vanish. At around $100/kg, which assumes full Starship reusability, the launch line stops dominating and the remaining costs, radiators, hardware life, financing and failure rates, decide the outcome. That is why analysts frame the roughly 2040 parity date as conditional: cheap launch gets you close, but the other inputs have to move too.
Is space-based solar power the same thing as an orbital data centre?
No, though they share hardware and economics. Space-based solar power generates electricity in orbit and beams it to Earth, and it reaches energy parity at roughly $250/kg. An orbital data centre keeps that power on board to run compute, so it needs a lower launch cost, around $100/kg, before parity. They are related but distinct propositions with different thresholds.
How much power, land and water does the terrestrial data centre baseline actually use?
It depends which layer you measure. The four-layer supply stack runs from grid-connected capacity, through converted capacity and powered land, to behind-the-meter generation and industrial capacity, each carrying different power, land and cooling costs. Those physical constraints, especially multi-year grid-interconnection queues that can stretch to around seven years in PJM, are exactly what make terrestrial costs rise and orbital parity move closer.
Does an orbital data centre still need batteries or storage for eclipse periods?
Yes, and it is a mass penalty you cannot ignore. A satellite passes through eclipse on each orbit, so it needs enough energy storage, or enough spare solar capacity, to keep computing through the dark. That storage adds mass, which adds launch cost, so eclipse cycling feeds straight back into the launch-cost lever that dominates the whole model.
