The Verdict

Every module in this Frontier has been a single line item. Now put them on one invoice. To run 40 megawatts of computers in orbit you launch a radiator field massing hundreds to over a thousand tonnes, a solar array of half a million square metres, a couple hundred tonnes of batteries to cross the eclipse, and the compute itself — then you operate hardware that fails faster than on the ground and cannot be touched. The proponents' own reference design is 40 megawatts. We priced their design, with their own best-case launch numbers, and it does not beat a building.

What you launch for an orbital data centre

Share of launched mass for a 40 MW facility — the compute is the sliver

Radiators 40Solar array 33Structure & propellant 12Batteries 10Compute 5

That pie is the whole argument in one shape. Ninety-five percent of what you lift to orbit — and pay to lift, and must keep alive — exists only to cool and power the five percent that actually computes. A terrestrial data centre inverts this: the building, the power and the cooling are a managed overhead around a floor full of servers you can walk up to and replace. Orbit turns the overhead into the entire machine, launches it at thousands of dollars a kilogram, and then seals it shut. You are not paying a premium to put computers in space. You are paying to put a power station and a radiator farm in space and letting the computers ride along.

	Orbital (40 MW)	Terrestrial (40 MW)
Power + cooling mass to launch	~1,500–3,000+ tonnes	n/a — built on site
Launch cost (Falcon Heavy, demonstrated)	$2.3–4.5 B	n/a
Launch cost (Starship, aspirational $100/kg)	$150–300 M	n/a
All-in build cost	Above + compute + structure + shielding	$1.2–1.6 B (everything)
Hardware refresh	Effectively impossible	Continuous hot-swap, 1–3 yr cycle
Time to power	Launch campaign + assembly	18–24 mo build, but 4–10+ yr grid queue

There is one more cost the pitch quietly drops, and it is the one that finishes the argument: upkeep. Starcloud's headline is that orbital power costs roughly 95 percent less, and that is true — sunlight is free. But the recurring cost of a data centre is not its power bill; it is replacement. GPUs die in one to three years (Module 5), and on Earth you swap them off a loading dock. In orbit you cannot, so you re-launch them — and the entire power-and-cooling structure they ride on — again and again, for the life of the facility.

Put the two on a like-for-like annual footing, per megawatt of IT, and the shape is clear. Both pay the same brutal hardware-refresh bill (~$8–12.5M per MW-year), because both run the same short-lived silicon. Terrestrial adds a modest power-and-facility charge. Orbit deletes the power charge — the pitch's whole case — and in its place adds the cost of launching every replacement to orbit, plus space insurance (a market growing at ~9% a year, with premiums rising in an increasingly congested orbit), plus station-keeping and a ground segment. The launch-to-replace line alone runs from a few million to tens of millions of dollars per megawatt per year, against the roughly $1.1 million per megawatt-year that going solar actually saved.

Cumulative upkeep, per MW of IT, over 10 years

What it costs to keep 1 MW running — terrestrial vs. orbital

TerrestrialOrbital (Starship prices)Orbital (today's launch)

DC Atlas modelling from the sourced components below. Both include the same GPU-refresh bill; the orbital lines add launch-to-replace, insurance, station-keeping and ground segment, and remove the terrestrial power charge. Even at Starship prices that have not been flown, orbit runs about twice terrestrial; at today's launch costs, ~5×.

Cost component	Terrestrial	Orbital
Power / energy	~$1.1M	~$0 (solar)
Facility: staff, maintenance, water, tax	$0.4–0.6M	—
Hardware refresh (GPUs, 2–3-yr life)	$8–12.5M	$8–12.5M (same silicon)
Launch to replace hardware + power/cooling mass	—	$4–56M (Starship → Falcon Heavy)
Station-keeping, ground segment, space insurance	negligible	$3–8M
Total per MW-year	~$9.5–14M	~$15–77M

Now the strongest case against our own conclusion, stated as fairly as we can. The orbital camp is not wrong about everything — it is right about the one thing that is genuinely broken on the ground. Terrestrial power is slow: grid interconnection queues in the biggest AI markets now stretch four to ten years and beyond, and no amount of capital buys you to the front of the line. An orbital system needs no interconnection agreement at all; it makes its own power the day it reaches orbit. That is a real advantage, and it is the same argument — bypass the queue, generate your own power — that makes on-site nuclear compelling in our first Frontier. The difference is that a reactor on a plot of land in Illinois does not have to be launched at $1,500 a kilogram or cooled in a vacuum.

So name the conditions under which we would be wrong, because an honest debunk has to. If launch costs collapse to Starship's target and hold there; if a radiator breakthrough takes areal density well below 6 kg/m²; and if satellite lifetimes stretch far enough that the refresh problem softens — then the gap in this Frontier narrows from absurd to merely large. We do not believe those three arrive together soon, and even all three would not make orbit cheaper than the ground for mainstream compute; they would make it possible. But if you think all three are coming, you should discount this verdict accordingly. The number to watch is the one in Module 3: dollars per kilogram to orbit, demonstrated, not promised.

And here is where orbit could actually earn its place, which is a real niche rather than a grudging concession. Not a 40 MW hyperscale training campus, but a small, radiation-tolerant, latency-insensitive workload — sovereign or defence data that must sit above any jurisdiction, an off-Earth backup vault, edge inference for other spacecraft, or compute sited where terrestrial power and connectivity simply do not exist at any timeline. In those corners the calculus flips, because the comparison is no longer against a cheap building on the ground but against nothing at all. That is a genuine market, and it is the one Lonestar's lunar vaults and the defence-adjacent proposals are really chasing. It is not the market the pitch is selling, which is the replacement of the data centre — and that market, on today's physics and prices, is closed.

Input	Value used	Basis
Radiator temperature	40–80 °C (313–353 K)	Realistic electronics-cooling range; Stefan–Boltzmann
Radiator emissivity	~0.9	Optimised spacecraft coatings
Radiator areal density	~19 kg/m² (goal <6)	NASA deployable-radiator data (NTRS)
Waste heat	≈ IT load (100%)	Nearly all chip input power leaves as heat
Launch cost (demonstrated)	~$1,500/kg (Falcon Heavy)	SpaceX published capacity/price
Launch cost (aspirational)	$100–200/kg (Starship)	SpaceX target — UNPROVEN, labelled as such
Solar specific power	~170 W/m² (roll-out)	NASA iROSA data
Eclipse fraction	~47% of orbit	~550 km LEO geometry
GPU service life / failure	1–3 yr / ~9% per yr	Terrestrial GPU-fleet studies
Terrestrial baseline	$30–40 M/MW; PUE ~1.1	McKinsey, CBRE/JLL hyperscale data

We wanted this one to be true. An orbital data centre is a genuinely beautiful idea — free sunlight, no water, no land fights, no seven-year wait for a substation — and the people building toward it are serious. But a Frontier is only worth anything if the numbers, not the wishes, get the last word. Take the proponents' own 40 megawatt design, cool it in a vacuum, launch it at the price they hope for, and keep it alive where no one can reach it, and the answer that comes back is the same every way you run it. The frontier that cooks itself is not a place to put a data centre. It is a place to put a small, specialised machine that has nowhere else to go — and, for now, nothing more.