At $3k/kg, orbit costs 2.5× the ground. It breaks even at $54/kg.
Breakeven launch cost
$54/kg
vs Falcon 9 today at $2,720/kg, Starship target near $10–50/kg
$/GPU-hour vs launch cost ($/kg, log)
you are at $3k/kg
Orbit Earth your launch cost breakeven
What a gigawatt actually costs
Build the same compute on the ground and in orbit, then see which pays back, and when.
scale
Earth
$37B
GPUs / IT $25BLand + building $2.0BPower delivery $6.0BCooling plant $4.0B
+ grid power $911M/yr, every year
Orbit at $3k/kg · 1.11× radiation
$109B
GPUs / IT $28BBus hardware $8.0BLaunch (t0) $73B
power $0/yr · re-launch compute $7.3B/yr
t0 build cost for equal useful compute. Earth runs 7,143 racks; orbit flies 7,937 (1.11× for radiation redundancy). Earth trades a cheaper build for a recurring power bill; orbit front-loads the launch to erase it.
Payback period at $3k/kg
unlikely
at this launch price orbit's re-launch toll exceeds Earth's whole power bill, so payback looks unlikely for now. As launch costs keep falling (see the trajectory below), cut radiator mass or extend compute life to bring it forward.
orbit cumulative Earth cumulative
When does the rocket catch up?
The launch price isn't a fixed slider, it's falling ~9%/yr over the long run, ~22–28%/yr in the reusability era. Project it forward and find the year it drops below your breakeven (the green line, driven by every other assumption on this page).
decline rate
Parity with your breakeven
~2038
Launch crosses your breakeven of $54/kg in ~2038, about 12 yrs out at −22%/yr.
Reaches the cited $50/kg
~2039
the round-number threshold the industry quotes, independent of this model
actual projected your breakeven $50/kg
The Grid Wall: why build it anyway
The honest case for orbit isn't a lower $/GPU-hour, it's that on Earth you increasingly can't build at all. NIMBY zoning, water and noise fights, and 7-year grid-interconnect queues mean compute is becoming speed-constrained, not chip-constrained. GPUs bought to sit in a queue depreciate before they earn.
N. Virginia/PJM up to ~7 yr; Dublin moratorium to 2028; new transmission lines 10–15 yr. Orbit deploys in ~12 months.
Orbit time-to-compute
~1yr
Earth time-to-compute
4yr
GPU value eroded in the queue
76%
At 30%/yr depreciation, a $25B GPU order loses $19B of useful value while it waits 4 yr for power. That write-down, not the launch bill, is what orbit's instant deployment buys back.
Tune the scenario
Two levers move almost everything: how cheap launch gets, and how heavy cooling stays. Start from a preset, then push the dials and watch the verdict above move.
Jump to a launch-cost scenario
Orbit type
Sun-synchronous dawn-dusk orbit stays in continuous sunlight, the steelman for space, no battery mass (Google Suncatcher's choice).
Data-gravity overlay
Core economics are workload-agnostic (launch vs radiator). Turn on a workload to add the laser-downlink penalty, illustrative and editable in Assumptions.
Launch cost: the master variable
Falcon 9 today ≈ $2,720/kg; Starship targets $10–50/kg. Drag down to find the breakeven.
Radiator mass: the real constraint
ISS heritage ≈ 291; advanced integrated panels ≈ 10; idealized droplet radiators ≈ 0.7. In a vacuum you can only shed heat by radiating it.
What you actually launch: 3,376 kg per rack
GPU rack
1,360 kg · 40%
Solar array
280 kg · 8%
Battery
0 kg · 0%
Radiator
1,400 kg · 41%
Structure
336 kg · 10%
At today's radiator setting the heat-rejection hardware is 1,400 kg, 41% of launch mass, often more than the rack itself. Launching the whole stack costs $9.18M per rack at $3k/kg. The obsolescence split matters: the 1,360 kg of compute is re-launched every 4 yr, but the 2,016 kg bus (solar, radiator, structure) is launched once per 12 yr.
Why cooling, not power, is the wall
In a vacuum there is no air or water to carry heat away — you can only radiate it. How much radiator area that takes is set by two design choices, and SpaceX's AI1 pushes both to the limit: how hot you run the skin, and whether you fly edge-on to the Sun.
On Earth a chip sheds heat into air and water (convection + conduction). In orbit the only exit is thermal radiation — here for a panel radiating from both faces:
P / A = 2 · ε · σ · T⁴
Rejection scales with the fourth power of skin temperature. Run the panel cool and let it face the Sun and, after the ~1,361 W/m² solar wash plus Earth's albedo and ~280 K IR, it nets only ~350 W/m². Run it hot — AI1's ~65 °C skin — and fly it edge-on so each face sees deep space, not the Sun, and the two faces together reject ~1,400 W/m² (2 · 0.95 · σ · 338 K⁴ ≈ 1,408).
That ~4× swing is the whole ballgame. A 140 kW rack sized the naive way needs:
140 kW ÷ 350 W/m² ≈ 400 m²
But AI1 rejects its 150 kW peak in just 110 m² (≈1,360 W/m² net) — about 3.6× smaller than the naive sizing. Credit where it's due: that hot, double-sided, edge-on radiator is the core move that turns “too big to launch” into a real satellite. The price is thermal margin — silicon near ~90 °C, a hair under its ~105 °C throttle — so it leans on high-reliability pumps with little headroom.
Even at 110 m², the panel is still a meaningful solar sail at low altitude — drag that demands station-keeping propellant — which is why a higher, edge-on dawn-dusk orbit (where it flies edge-on anyway) is the sane default. Note this is radiator area; the launch mass per kW (the slider above) is a separate lever, and the one that actually drives the economics.
Read
What AI1 actually shows. SpaceX's announced bird is, almost to the kilowatt, one rack per satellite (~120 kW average, 150 kW peak). The headline isn't the solar wing — it's the radiator. At a conservative ~350 W/m² you'd size the heat-rejection panel for a 150 kW load at ~430 m²; AI1 does it in 110 m². The trick is pure Stefan-Boltzmann: run the skin hot (~65 °C) and fly it edge-on so both faces see deep space instead of the Sun, and rejection jumps ~4× to ~1,400 W/m². The cost is thermal margin — silicon near 90 °C, just under its throttle — but it is what turns “too big to launch” into a real satellite. That is the engineering to respect.
Cheap rockets are necessary but not sufficient. Drop launch cost to Starship's target and orbit still has to launch, and then never service, a radiator that, with today's tech, can outweigh the computer it cools. The orbital thesis really rests on two bets stacked on top of cheap launch: a 10–30× lighter radiator and hardware that survives years without a technician. Defaults assume the kindest case: a dawn-dusk sun-synchronous orbit with no batteries; switch to a 30° LEO and battery mass plus an oversized array make it worse. The obsolescence trap is the quiet drag: GPUs age out every few years and must be re-launched, even as the bus lives on. Where it gets interesting is power-constrained grids: push Earth electricity up (or assume you simply can't get the megawatts on the ground) and free 24/7 solar starts to pay for the launch.
At facility scale the deal is a swap, not a saving: orbit converts a recurring grid bill (opex) into an upfront launch (capex), and pays it back only if the avoided power outruns the re-launch toll of replacing GPUs every few years. So the payback flips on launch price: cheap enough and orbit wins from day one; expensive enough and, at today's prices, the power bill is unlikely to repay it. The quiet, under-priced advantage isn't dollars, it's time-to-power: a multi-year grid-interconnect queue on the ground versus sunlight that's available the day you arrive.
Three physics taxes this model now lets you turn on, and two it leaves as caveats. Radiation: cosmic rays flip bits and corrupt training gradients, so orbit spends compute on redundancy; dial the overhead from ECC-plus-checkpoint (~10%) to full triple-modular redundancy (67%) and watch the breakeven move. Interconnect: frontier training is synchronous and bandwidth-bound; inter-satellite lasers can't carry it, so space is structurally an inference and edge tier, not a training one. Heat as area: the radiator isn't just mass, it's ~300 m² per 120 kW rack, a sail that drags in low orbits and needs station-keeping propellant (our dawn-dusk SSO default sits high enough to soften this; drop to a low LEO and it bites). Two we don't price: the re-entry externality (alumina from burning up refreshed hardware is a live ozone-and-regulatory risk) and the real value driversbeyond cost: grid-bypass, in-orbit edge inference that downlinks only insights, and physical/sovereign security. Net: orbit looks like a specialized premium tier, not a replacement for hyperscale.
What this models & sources
Launch: Falcon 9 ≈ $2,720/kg to LEO (NASA cost survey); Starship 150 t reusable, near-term $90–100/kg, $10–50/kg long-term target range. Google's Project Suncatcher names $200/kg as the inflection; Citi GPS projects $100/kg by 2040.
Launch-cost trajectory (constant ~2025 USD): Space Shuttle ≈ $54,500/kg → Falcon 9 ≈ $2,720 → Falcon Heavy ≈ $1,400 → reusable Falcon 9 ≈ $1,500 (2024) → Starship early single-use ≈ $1,000–1,200, partial reuse ≈ $80–94, high reuse $13–32, marginal floor ≈ $10. Implied decline: ≈ −9%/yr over 1981–2018, ≈ −22% to −28%/yr in the reusability era. Projection compounds the chosen rate from a 2026 anchor; the $50/kg and breakeven crossings are model outputs, not forecasts.
Grid Wall: N. Virginia/PJM interconnect queues up to ~7 yr; Dublin (EirGrid) data-center connection moratorium to 2028; new US transmission lines take 10–15 yr to permit/build. Frontier AI GPUs lose ~50% of value per 18–24 mo, so hardware waiting in a queue depreciates before first compute. That is the time-value/regulatory arbitrage that orbit's ~12-month deployment buys back.
Anchors: SpaceX's Jan 2026 FCC filing for a ~1-million-satellite orbital data-center system; the EU/Thales ASCEND feasibility study; Starcloud (ex-Lumen Orbit) and Google Project Suncatcher.
SpaceX AI1 (announced 2026; physics verified, roadmap claims not): ~120 kW average / 150 kW peak payload per satellite; a 110 m² double-sided liquid radiator run at a ~65 °C skin (≈90 °C silicon junction) flown edge-on to the Sun; ~70 m solar wing; ~600 km LEO; optical inter-satellite mesh. The radiator math is self-consistent (2·0.95·σ·338 K⁴ ≈ 1,408 W/m² → 150 kW ÷ ~1,360 net ≈ 110 m²); the constellation scale, chip/JV and per-link bandwidth figures are roadmap statements, not verified specs, and are treated as such.
Power & storage: solar constant ≈ 1,361 W/m²; space solar 0.59–7.7 kg/kW; dawn-dusk SSO needs no batteries, a ~30° LEO is eclipsed ~39% (battery ≈ P·eclipse / (specific-energy·DoD), with the solar array oversized to recharge). US industrial power ≈ $0.08/kWh; PUE 1.08–1.58.
Cooling: radiator mass 0.7 kg/kW (idealized droplet) to ~291 kg/kW (ISS heritage), ~10 kg/kW for advanced integrated panels. The dominant, contested lever.
Lifecycle (obsolescence trap): high-utilization AI GPUs last 1–3 yr and must be re-launched (no in-orbit servicing); the long-lived bus (solar/radiator/structure) amortizes over a separate, slower cycle. Meta reported 30.1% GPU failures in a large training run.
Data gravity & interconnect: training demands heavy space-to-ground + inter-satellite laser bandwidth; inference is light. Modelled as an optional, illustrative $/GPU-hr overlay (off by default). The deeper limit is structural: frontier synchronous training needs TB/s GPU-to-GPU fabric (NVLink/InfiniBand); inter-satellite optical links (~10–100 Gbps, with pointing jitter) can't sustain it across a constellation, biasing orbit toward inference/edge.
Radiation reliability: sub-5nm GPUs in LEO suffer single-event upsets / silent data corruption and latch-ups. Mitigation (ECC + checkpoint/roll-back, spares, up to triple-modular redundancy) is modelled as a % of useful compute lost, so orbit flies that many extra racks for equal output. Default 10% (kind case); TMR ≈ 67%.
Thermal as area & drag (caveat, not priced): radiative rejection is a design choice, not a constant. A cool, Sun-facing panel nets ~350 W/m²; a hot (~65 °C skin), edge-on, double-sided panel reaches ~1,400 W/m² (2·0.95·σ·338 K⁴). SpaceX AI1 uses the latter to cool 150 kW in 110 m²; the naive ~350 W/m² sizing would demand ~3.9× the area. Large panels still add atmospheric drag → station-keeping propellant (altitude-exponential), so a higher, edge-on dawn-dusk orbit softens it. Radiator mass per kW is what the economics price; this area figure drives only the drag caveat.
Re-entry externality (caveat, not priced): refreshing compute means de-orbiting hardware; vaporized aluminium deposits alumina that can catalyse ozone loss, an emerging environmental/regulatory risk to mega-constellation refresh, not a unit-cost input.
Value drivers beyond cost: grid-bypass / time-to-power; in-orbit edge inference (downlink only insights, cutting bandwidth ~1000×); and physical/jurisdictional security (cf. the EU ASCEND motivation). These, not raw $/GPU-hr, are why orbital compute is being pursued.
Facility TCO (1 GW mode): a 1 GW AI build runs ~$40–50B all-in (Stargate ≈ $500B / 10 GW). Earth physical plant (land/building, power delivery with substation and transformers, and cooling) is broken out at ~$10–12/W; the rest is IT. Space prices bus hardware (solar/radiator/structure) ex-launch plus the launch itself; power opex is ~$0. All $/W and $/kg inputs are editable. Payback compares cumulative spend: orbit's extra upfront vs Earth's recurring grid bill, net of orbit's GPU re-launch toll.
Caveat: defaults are illustrative, sourced order-of-magnitude figures, not disclosed economics. Ground-station capex and bus refresh within the horizon are not separately priced; the obsolescence trap assumes no in-orbit servicing. Every input is editable; the formulas are the deliverable.