Heat is the binding constraint on orbital compute.

AI ran out of power before it ran out of chips.

The limit used to be silicon. Chip factories could not keep up, and that is being fixed: fabs are being built and supply is catching up. The limit now is electricity, and the grid is not keeping pace.

SemiAnalysis forecast that global data centre IT power demand would grow from 49 GW in 2023 to 96 GW by 2026, and JLL expects global data centre capacity to nearly double to 200 GW by 2030.

New generation is not being built that fast, and money alone doesn’t close the gap quickly. Finding power somewhere else might.

Everywhere on Earth, power has a floor.

You can buy cheap electricity. You cannot buy cheap electricity that is there whenever you need it.

Solar and wind are cheap and getting cheaper: NREL projects $16 to $18/MWh for utility solar and onshore wind by 2050 (ATB 2025, Advanced scenario). But they only produce when the weather cooperates, and a data centre cannot pause overnight. Running around the clock takes firm power, which NREL projects at $59 to $67/MWh by 2035 (geothermal and nuclear, ATB 2025).

Falling storage costs will narrow that gap; how far, and how fast, is the open question. Firmness is exactly what a data centre buys, and on Earth it costs extra.

In orbit, the sun barely sets.

In a dawn-dusk sun-synchronous orbit, a solar array sees sunlight almost continuously, with no weather and only brief eclipse seasons. Sunlight there is also about a third more intense than the standard sunlight used to rate solar panels on the ground, because there is no atmosphere in the way. Orbital compute doesn’t have to buy firmness. It has it structurally, without paying for storage or backup.

This is no longer a thought experiment. Starcloud raised $250M at a $2.3B valuation in August 2026, bringing its Series A to $420M. SpaceX has applied to the FCC for a constellation of up to one million orbital data centre satellites. An EU-funded ASCEND study, led by Thales Alenia Space, found space data centres feasible, but concluded they would need a launcher ten times less emissive over its lifecycle to deliver an emissions benefit.

But free power is not the same as cheap compute.

Orbit has to beat Earth’s floor, not Earth’s average.

In our model, with launch at $300/kg and solid radiators, powering and cooling a gigawatt of compute in orbit costs $5.4B per gigawatt-year. On Earth, the infrastructure and firm power for the same gigawatt cost $1.6B per gigawatt-year. Orbit costs about 3.3 times as much. The chips cost the same in both places, so they are left out; the assumptions are set out in the methodology below.

Four levers dominate the cost.

Launch cost, solar efficiency, compute density and heat rejection. Radiation tolerance, lifetime, maintenance, networking and launch availability matter too, but these four dominate the answer.

Three of them improve every year.

Rockets are getting cheaper. Falcon 9 costs about $3,400/kg at list price today, and for Starship, SpaceX targets below $100/kg with full reuse.

Solar panels are getting lighter, and chips are doing more work per watt. These three ride the same industrial learning curves that made AI possible on Earth, and they are the trends that convinced serious people to start building.

The fourth improves slowly, and has a ceiling.

A chip in orbit turns electricity into heat exactly as it does on Earth. On Earth you blow air over it or run water through it. In space there is nothing to carry the heat away, so the only way to shed it is to radiate it, from panels that grow in step with every watt of compute.

How much heat a surface can radiate is bounded by εσT⁴ (the physics is on our technology page). Emissivity is already close to its limit, and temperature is capped by what the electronics can tolerate. Engineering has made panels lighter per square metre and will keep doing so, but it cannot reduce the area needed per watt at a given temperature. Every watt added is a watt that has to be radiated.

The radiator is most of the bill.

In our model, at Starship’s target launch price, heat rejection is nearly three quarters of the cost of power and cooling in orbit. Not the rockets, and not the solar arrays. Per year, the radiator alone costs about twice as much as running a gigawatt data centre on Earth.

Launch price$2,700/kg$300/kg (model base case)$100/kgFree launch

Data table

Launch price	Earth	Orbit: launch	Orbit: solar	Orbit: heat rejection	Orbit total	Heat rejection share
$2,700/kg	$1.63B	$2.84B	$1.21B	$7.25B	$11.30B	64%
$300/kg (model base case)	$1.63B	$0.32B	$1.21B	$3.84B	$5.37B	72%
$100/kg	$1.63B	$0.10B	$1.21B	$3.56B	$4.88B	73%
Free launch	$1.63B	$0.00B	$1.21B	$3.41B	$4.63B	74%

Fig. 1: Cost of power and cooling for a gigawatt of compute, per year, on Earth and in orbit, in our model (Pluvio cost model v3, 1 October 2026). Billions of dollars per GW-year. Compute hardware is excluded because it costs the same in both places. Heat rejection is radiator hardware plus the launch of its mass. Model output, not measured data.

In our model, heat rejection is nearly two thirds of the cost at today’s launch prices, nearly three quarters at Starship’s target price, and about three quarters even if launch were free. Most of the radiator’s cost is the hardware itself, so cheaper launch barely touches it. Every dollar taken out of launch raises the share of the remaining bill that is radiator.

What Pluvio is building.

Pluvio is developing a liquid droplet radiator. Instead of radiating heat from solid panels, it sends the hot coolant across open space as a sheet of fine droplets, then recaptures them. Its mass still grows with power, but, if it works, it needs far less mass per kilowatt.

In our modelling it is 52% to 74% lighter per kilowatt than a two-sided solid radiator, under assumptions that favour the panel. Because the saving is per kilowatt, it grows with power: the higher a spacecraft’s power, the more mass a droplet radiator takes off it.

In our model the radiator is the largest term in the orbital bill, so a lighter radiator moves the answer more than any other single lever. How far it moves depends on a result we have not yet shown in hardware. Our test programme is designed to find out, in order: form and recapture a droplet sheet on the bench, repeat it in vacuum, test it in microgravity, then fly it.

Methodology and assumptions

Every figure on this page is either sourced (listed below) or an output of Pluvio’s cost model, labelled “in our model”.

Figures are from Pluvio's orbital versus terrestrial cost model (v3, 1 October 2026) and compare the cost of power and cooling per gigawatt of compute. Compute hardware is assumed to cost the same in both places and is excluded. Orbital cost is the cost of launch, solar arrays and radiators, spread over a 10 year life at 85% utilisation. Terrestrial cost is data centre infrastructure, excluding chips ($21B per GW over 20 years), plus firm power at $60/MWh with a PUE of 1.1. Orbital inputs: launch at $300/kg, solar arrays at 261 W/kg and $8/W, a two-sided solid radiator at 4 kg/m² and 325 K. Adding loop, deployment and micrometeoroid allowances gives 9.7 kg/kW, with radiator hardware costed from a private industry research model. Costs are undiscounted. Values change with launch price; see the chart.

Every comparison is deliberately unfair in the incumbent architecture’s favour. The full model is available on request.

Ask for the model

Sources

SemiAnalysis forecast that global data centre IT power demand would grow from 49 GW in 2023 to 96 GW by 2026 newsletter.semianalysis.com/p/ai-datacenter-energy-dilemma-race
JLL expects global data centre capacity to nearly double to 200 GW by 2030 www.jll.com/en-us/newsroom/global-data-center-sector-to-nearly-double-to-200gw-amid-ai-infrastructure-boom
$16 to $18/MWh for utility solar and onshore wind by 2050 (ATB 2025, Advanced scenario) atb.nlr.gov
$59 to $67/MWh by 2035 (geothermal and nuclear, ATB 2025) atb.nlr.gov
In a dawn-dusk sun-synchronous orbit, a solar array sees sunlight almost continuously, with no weather and only brief eclipse seasons. www.cosmos.esa.int/web/proba-2/about-proba2
about a third more intense than the standard sunlight used to rate solar panels on the ground doi.org/10.1029/2010GL045777
Starcloud raised $250M at a $2.3B valuation in August 2026, bringing its Series A to $420M. techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/
SpaceX has applied to the FCC for a constellation of up to one million orbital data centre satellites. docs.fcc.gov/public/attachments/DA-26-113A1.txt
EU-funded ASCEND study, led by Thales Alenia Space, found space data centres feasible, but concluded they would need a launcher ten times less emissive over its lifecycle to deliver an emissions benefit www.thalesaleniaspace.com/en/press-releases/thales-alenia-space-reveals-results-ascend-feasibility-study-space-data-centers-0
about $3,400/kg at list price www.spacex.com/assets/media/Capabilities&Services.pdf
below $100/kg with full reuse x.com/elonmusk/status/2082612362226639328