A first-principles underwriting of a single 1 GW orbital data center, against the same gigawatt built on the ground. Mass, launches, time and money, with the best case argued as strongly as the numbers allow, and then the reasons it won't happen on schedule.
Companion to "Should you buy SpaceX?" (spacex-underwriting.pages.dev). Same chips on both sides of every comparison. Arrow keys to navigate.
The hardware for one orbital gigawatt masses 22,500 tonnes in the best case, roughly 50 International Space Stations, and it has to be replaced on a four-to-five-year cycle. The claim that space becomes the cheapest place for AI compute within two to three years requires five industrial curves to each move about an order of magnitude in the same window. None has a history of doing that.
1 GW of IT load, the size of a single hyperscale campus. Everything scales linearly from here, and the scaling claims (100 GW a year) are checked against it on slide 21.
A swarm, not a structure: 265 self-deploying satellites in dawn-dusk sun-synchronous orbit, one per Starship, flying formation with optical links. This is what every serious proposal assumes, because it removes in-space assembly. Any other orbit costs about 1.6× on everything before batteries.
Best case: every subsystem at its published research target, launch at Musk's internal cost target, a third of all Starship flights dedicated.
Flown: hardware that has been in orbit or ground-tested at scale.
First the arithmetic (slides 4 to 11), then why the best case is a conjunction (slide 12), then the reference class of what actually happens to schedules (slide 13).
| Solar constant, AM0 | 1,361 W/m² |
| III-V cell efficiency, beginning of life | 30% |
| Areal power, normal incidence | 408 W/m² |
| Degradation to end of life ×1.15, conditioning ×1.1, pointing ×1.05 estimate | ×1.35 |
| Array power per GW of load, BOL | 1.35 GW |
| Array area per GW of load | 3.3 km² |
Three Starlink constellations of solar area, or about a thousand ISS arrays, and this is before the radiators. In dawn-dusk orbit the arrays sit edge-on to the velocity vector, which is the only reason drag stays manageable.
| Specific power, whole wing | W/kg | Status |
|---|---|---|
| ISS solar wings, 1998 | ~30 | flown |
| JPL APSA program goal, set 1985 | 130 → 300 | 300 unmet |
| Redwire iROSA blanket | >100 | flown 2021 |
| UltraFlex, BOL | >150 | flown |
| NASA benchmark, 30–50 kW blankets | ~150 | ground test |
| SBIR target, array level | 500 | target |
| Spectrolab 75 µm bare cell | 1,000 | not an array |
| 1.35 GW of array at | Mass | Starship @85 t |
|---|---|---|
| 150 W/kg, best case | 9,000 t | 105 |
| 100 W/kg, flown | 13,500 t | 160 |
| 500 W/kg, target only | 2,700 t | 32 |
Every figure in the left table is for a 20 to 50 kW wing. A gigawatt array is 30,000 of them plus the truss and harness that tie them together, and no one has a demonstrated W/kg for that.
| Heat to reject (1 GW IT + conditioning losses) | 1.1 GW |
| Radiator surface temperature, best case | 60°C (333 K) |
| Emission per face, εσT⁴ with ε = 0.9 | 627 W/m² |
| Net, two faces, after Earth infrared estimate | ~1,000 W/m² |
| Radiator area per GW | 1.1 km² |
| Areal mass | kg/m² | Status |
|---|---|---|
| ISS radiators, running near 0°C estimate | ~15 | flown |
| Composite panel with heat pipes estimate | ~6 | flown, small |
| Lightweight target | ~3 | target |
| Best case: 3 kg/m² + 35% for pumps and fluid loops | 4,450 t |
| Flown: 6 kg/m² + 35% | 8,900 t |
Running the radiator hot is what keeps this to half the array mass: emission scales with T⁴. It also requires the whole chip-to-fin temperature drop to fit inside 30 K, which is a two-phase pumped loop nobody has flown at a fraction of this scale.
A GB200 NVL72 rack is roughly 1,360 kg for ~125 kW, about 10.5 kg/kW as built for a building. Strip the cabinet and the fans, keep the boards and cold plates: 4 kg/kW best case, 8 kg/kW flown-equivalent. estimate
Suncatcher-style: minimal aluminum, accept the upsets, let ECC and checkpointing absorb them. +20% of compute mass. estimate
Two stages from array bus to chip. 4 kW/kg best case, 2 kW/kg flown. estimate
LEO plasma limits practical array voltage to a few hundred volts (the ISS runs 160 V with plasma contactors). At 300 V a 3.8 MW node carries ~13,000 A in copper. estimate
12% of the above best case, 15% flown, in line with large spacecraft bus fractions. estimate
Raise from a 400 km drop-off, five years of drag make-up with arrays edge-on, and an 85 m/s disposal burn at end of life. Hall thrusters at 1,600 s. Small, because the geometry is the design. estimate
Not in any column: segmenting the radiator loops so a single debris strike doesn't drain a node (slide 15), and spares.
| Subsystem, 1 GW of load | Best case | Flown |
|---|---|---|
| Solar arrays, 1.35 GW BOL | 9,000 t | 13,500 t |
| Radiators and loops | 4,450 t | 8,900 t |
| Compute | 4,000 t | 8,000 t |
| Shielding | 800 t | 1,600 t |
| Power conditioning | 500 t | 1,000 t |
| Harness | 800 t | 1,200 t |
| Structure, ADCS, avionics | 2,400 t | 5,000 t |
| Propulsion and propellant | 600 t | 800 t |
| Total | ~22,500 t | ~40,000 t |
| Specific mass | 22.5 kg/kW | 40 kg/kW |
SpaceX's FCC filing: one million tonnes launched per year yields 100 GW. That is 10 kg/kW, below the best case here by more than half. It needs 500 W/kg arrays and 2 kg/m² radiators, plus 2 kg/kW compute, all at once. None has flown.
An 85-tonne Starship load at 22.5 kg/kW is a 3.8 MW satellite: 12,500 m² of array (three ISS's worth), 4,200 m² of radiator, 100-metre-class wingspans. 265 of them in formation. Total mass about 50 ISS.
Each mark is one Starship at 85 t to dawn-dusk orbit. Arrays alone, 105 flights. The whole ISS, five flights. Remaining best-case hardware to 265. Flown-hardware case continues to 470. Falcon 9 equivalents: 1,320 and 2,350 flights at 17 t.
| Starship flights | 2023 | 2024 | 2025 | 2026 to Aug |
|---|---|---|---|---|
| Flown | 2 | 4 | 5 | 2 |
| FAA-licensed cap | 5 | 5 | 25 | 25 |
| Delivered to orbit | 0 | 0 | 0 | 0 |
Version 3, the 100-tonne vehicle, first flew on 22 May 2026. Flight 13 deployed 20 Starlink V3 satellites on a suborbital arc in July. Flight 14, due at the end of August, is the first attempt to put V3 satellites into orbit and catch the ship. The manifest is described as full through 2028 and the program has slipped three times this year.
| Flights per year estimate | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 |
|---|---|---|---|---|---|---|---|---|
| Falcon 9 analog (6 yrs to 100/yr) | 10 | 15 | 25 | 40 | 65 | 100 | 140 | 170 |
| Doubling every year | 10 | 20 | 40 | 80 | 160 | 320 | 640 | n/a |
2027 is taken as the first year of real cargo flights. Starlink V3, Artemis tanker flights, Starshield and paying customers all compete for the same vehicle; 30% to one data center is generous.
| Year the hardware finishes arriving | 265 flights | 470 flights |
|---|---|---|
| Falcon 9 analog, 30% share | 2036 | 2039 |
| Doubling, 30% share | 2033 | 2034 |
| Doubling, every flight dedicated | 2031 | 2032 |
The 2033 case requires about 640 Starship flights in that year, more than double every orbital launch worldwide in 2025, a third of them carrying nothing but this.
Starlink's line, about 2,500 satellites a year at ~100 m² each, makes roughly 0.25 km² of space array per year and is the largest ever built. 3.3 km² is 13 years of it, or a 5× expansion to meet a 2032 launch window. estimate
The cheap path is silicon in orbit: new modules pitched at under $5/W with six-month delivery, against ~$40/W for the III-V cells every flown array uses.
1.1 km² of pumped-loop radiator. There is no production line for deployable radiators at any scale that rounds to this. The largest flown units are ISS ORUs of about 75 m² each; this needs 15,000 of them.
The largest self-deploying array ever flown is iROSA at ~20 kW. One node here is 3.8 MW, 190 of those. The largest formation proposal on record is Google's ~80-satellite Suncatcher cluster; this is 265 hectare-class bodies with 100-metre wingspans.
None of these is impossible. Each is a scale-up of 10× to 200× of something that exists at demonstration size, and all three must land in the same window as the launch curve.
| Curve | Today | Best case needs | Move | What history says |
|---|---|---|---|---|
| Starship cargo flights per year | 2 so far in 2026; 5 in 2025 against a cap of 25 | 300+ | 100× | Falcon 9 took six years from a reliable vehicle to 100 per year |
| Array W/kg, whole wing | 100–150 | 150–500 | 1.5–5× | One doubling since 1985; JPL's 300 W/kg goal still unmet after 41 years |
| Radiator kg/m² at 60°C | ~15 on the ISS near 0°C | 3 | 5× | No pumped-loop radiator flown at that mass or temperature |
| Space array production, km²/yr | ~0.25 | 1–2 | 4–8× | Current line is already the largest in history |
| Largest self-deploying spacecraft | iROSA, 20 kW | 3.8 MW × 265 in formation | 190× | Nothing hectare-scale has deployed itself |
Give each curve a 50% chance of landing on schedule and the joint probability is 0.5⁵ = 3%. Give each the 20–30% that aerospace programs actually post and it is 0.3⁵ = 0.2%. The best case is a one-in-40 to one-in-400 outcome, and it is the case the two-to-three-year claim depends on.
The curves are not independent in the helpful direction. A slow Starship ramp starves the array factory of demand, which delays the W/kg learning curve, which raises the launch count. Correlation runs toward late.
| Program | Promised | Actual | Slip |
|---|---|---|---|
| Starship orbital flight | 2020 | Apr 2023 | +3 yrs |
| Starship flights in 2025 | 25 licensed | 5 flown | 20% |
| Starlink V3 launches on Starship | First half 2026 | Aug 2026 at the earliest | +6 mo, counting |
| Starship lunar landing (HLS) | 2024 | Now 2028 | +4 yrs |
| Falcon Heavy | 2013 | February 2018 | +5 yrs |
| Crew Dragon first crew | 2015 | May 2020 | +3 yrs |
| James Webb Space Telescope | 2007 | Dec 2021, $4.5B over | +14 yrs |
| SLS and Exploration Ground Systems | 2017 | November 2022 | +4 yrs each |
| NASA major-project portfolio, 18 projects | baselines | 14 yrs cumulative | ~9 mo each |
| JPL 300 W/kg array goal | early 1990s | Unmet, 2026 | +35 yrs |
| Stargate Abilene buildings 3 and 4 | March 2026 | Behind, Jul 2026 | slipping |
| US data center pipeline, 2026 | 16 GW announced | 30–50% delayed/cancelled | 9 cancelled |
This is the reference class, and it includes the ground. Terrestrial gigawatt campuses slip on transformers and turbines. An orbital one carries every terrestrial risk plus a launch vehicle that has not yet delivered a payload.
Apply the median slip in this table, two to three years, to the Falcon 9-analog cadence with flown hardware and the first orbital gigawatt lands in 2041 to 2042. Apply it to the best-case hardware and it is 2038 to 2039. Musk's "two to three years" is not on this page.
Triple-junction cells went from ~28% to ~32% over two decades: about 0.2 points a year. The wing-level W/kg metric is worse because it is booms and harness, and it has doubled once since 1985.
CdTe: lab cell in the early 1990s, gigawatt-scale manufacturing around 2008. PERC: invented 1989, dominant architecture 2018. Lab-to-volume has run 15 to 25 years every time, and orbit adds a qualification step (radiation and thermal cycling) that ground PV never took.
A process node is 3 to 5 years from risk production to volume; Arizona N4 was about 4.5 years from announcement to first wafers. Performance per watt doubles every 2.5 to 3 years and it is a wash here, because it improves the terrestrial campus identically. What moves the launch count is kilograms per kilowatt of rack, which is sheet metal and copper, and that barely moves between nodes.
Launch cadence is the only variable in this system with a demonstrated 2×-per-year history, and it is the only one the optimists are actually betting on. Everything else has decades of evidence that it moves at single-digit percent a year.
| Object population, 2026 | Count | Tracked? |
|---|---|---|
| Larger than 10 cm | 43,000+ | yes |
| 1 to 10 cm ("lethal non-trackable") | ~1.2 million | no |
| 1 mm to 1 cm | ~140 million | no |
At ~550 km the flux of >1 cm objects is on the order of 1e-6 to 1e-5 per m² per year. Across 4.4 km² of array and radiator that is roughly 4 to 44 strikes a year from objects nothing can see coming, and thousands from millimetre-class debris. estimate
A hole in a solar blanket costs a few cells. A hole in a pumped radiator loop drains it, which is how Soyuz MS-22 lost its coolant to a millimetre-class strike in 2022. So the 1.1 km² of radiator has to be hundreds of independent circuits, and that mass is in no column of the ledger.
The traffic side compounds it. Starlink's 10,000 satellites made over 355,000 avoidance manoeuvres in the year to May 2026, more than 40 per satellite, in the same altitude band. Each node here has the cross-section of a small town.
Epoch's terrestrial model annualises server cost at $5B a year for the same gigawatt. Orbit pays that too, because the chips are the same, and adds the launches. At Musk's $10M internal target the refresh launches are a rounding error. At today's Falcon 9 price of $74M they are $1.1B to $1.5B a year. The refresh economics are entirely a launch-price bet.
| Epoch AI, 1 GW hyperscaler campus, up-front | $M |
|---|---|
| Servers | 21,188 |
| Facility | 11,433 |
| Network infrastructure | 4,925 |
| Land | 172 |
| Utility works | 164 |
| Total | 37,883 |
| Non-chip subtotal | 16,700 |
| Annual opex (energy $0.6B at PUE 1.14) | 900 / yr |
| Time to 1 GW | |
|---|---|
| Amazon New Carlisle, Indiana: ground-break to >1 GW (Mar 2026) | ~24 months |
| Stargate Abilene: start Jun 2024, first 200 MW live Sep 2025, ~600 MW Jul 2026 | ~2 yrs to half |
| Colossus, Memphis: warehouse to 100,000 GPUs | 122 days |
Other estimates bracket Epoch: Bernstein $35B, Nvidia $50–60B, Orennia $60B for a gas-powered campus. The spread is chips and power strategy, not concrete.
| Constraint | Today |
|---|---|
| Utility interconnection, Northern Virginia / Phoenix / Dallas | 4 to 7 years |
| Large power transformers | 2 to 4 years |
| ERCOT large-load queue | 410 GW, 87% data centers |
| Behind-the-meter gas, equipment to commissioning | 12 to 18 months |
| Behind-the-meter capex premium over a utility tie | $2–4M per MW |
| Abilene on-site generation | 29 turbines, ~1 GW |
| 2026 US pipeline delayed or cancelled | 30–50% of 16 GW |
This is the honest version of the space pitch: Earth's constraints are political and orbit's are engineering. SpaceX is better at engineering than at zoning, and that is a real argument.
It is not a 2030 argument. The slow path on the ground is 4 to 7 years and the fast path is under two. The best day in orbit is seven years out, and the reference class says twelve to fifteen. A gigawatt of gas turbines costs $2–4B and shows up in 18 months; a gigawatt of Starship flights is 265 launches on a vehicle with none delivered.
| Line item | Orbit, best case | Orbit, flown | Ground |
|---|---|---|---|
| Launch (265 / 470 flights) | 2.7 at $10M · 6.6 at $25M | 11.8 at $25M · 47 at $100M | 0 |
| Arrays, 1.35 GW BOL | 13.5 at $10/W | 40 at $30/W | PV + storage 10–13, or gas 2–4 |
| Radiators, 1.1 km² estimate | 5 | 15 | in facility |
| Bus, structure, propulsion, harness estimate | 4.3 at $1k/kg | 24 at $3k/kg | in facility |
| Compute integration premium estimate | 4 | 8 | 0 |
| Facility, network, land, utility works | 0 | 0 | 16.7 |
| Non-chip total | 30–34 | 100–135 | 19–30 |
| Chips, both sides | 21 | 21 | 21 |
| Electricity per year | 0 | 0 | 0.6 |
| Refresh launches per year | 15–20 | 15–20 | 0 |
At best-case assumptions orbit costs 1 to 2× the ground and arrives five to ten years later. With flown hardware it is 4 to 7×. Free power is worth $0.6B a year against that. The $10/W array line is the silicon-in-space pitch, doubled for the wing; the III-V cells every flown array uses would put arrays alone above $130B.
Google's own conditional projection, orbit cheaper than the ground by 2035 if Starship reuse and cadence both land at the projected cost, sits inside this range. Musk's "within two to three years" does not.
The comparison that matters is not orbit against the grid queue. It is orbit against whatever the ground does next, and the ground has already answered the queue with turbines. By the time the first orbital gigawatt exists, the terrestrial gigawatt it competes with will be on its third generation of chips.
"The lowest cost way to generate AI compute will be in space within two to three years."
Orbital data centers may not prove viable.
One million satellites, 500 to 2,000 km, "first step towards becoming a Kardashev II-level civilization."
| Who | Claim or filing |
|---|---|
| SpaceX / xAI | 1,000,000 satellites; 1M tonnes per year → 100 GW; pilot nodes on Starlink V3 in 2026 |
| Starcloud | Trained a model in orbit on an H100, Dec 2025; FCC filing for 88,000 satellites, ~20 GW; Starcloud-2 with Blackwell, Oct 2026 |
| Suncatcher: two prototype satellites with Planet, early 2027; orbit cheaper than ground by 2035, conditional on Starship | |
| Aetherflux | $2B valuation, first commercial node Q1 2027 |
| Orbital (startup) | FCC filing June 2026 for 100,000 satellites, ~10 GW |
| Axiom | First free-flying orbital data center nodes, Jan 2026, 2.5 Gbps optical links |
| Jeff Bezos | Gigawatt data centers in space in "10+ years" |
Filed capacity across three companies: about 130 GW. Flown capacity across all of them: a few racks.
The strongest case for orbit was never economic. It sorts into three, and the deck's own numbers size each one.
Process sensor data where it is collected, so raw feeds never come down. Already flying: Axiom's free-flying nodes in January, Starcloud's H100 in December. Scale: megawatts. Nothing here argues against it.
Compute outside any jurisdiction, beyond seizure and beyond any switch. That is Lonestar's pitch; the honest name for it is compute no one but the operator can inspect or turn off. A feature to the operator, a governance problem for everyone else. The value is per rack, not per gigawatt.
This one turns around. A gigawatt swarm at 550 km is the least survivable compute anyone could build: one intercept or one fragmentation in the shell and debris finishes the job. Fengyun-1C's 3,000 tracked fragments from 2007 are still up there. Orbit is the hardest compute to regulate and the easiest to destroy.
How it gets built anyway. None of the three needs a gigawatt, so a gigawatt in orbit will not come from strategy. It can still come from a decision: an FCC milestone waiver (already requested), an executive designation, a procurement line, or Defense Production Act priority. Each moves the cost from the builder to the public; none moves the calendar. The beneficiary is positioned: SpaceX and xAI merged in February, the million-satellite application is pending with its waiver, and IPO proceeds were earmarked for orbital data centers. Whim can move the money. It cannot move 265 launches or the 2033 floor.
What it needs to be true: about 10 kg/kW at node level with a radiator design nobody has published, and a Starship cadence in the hundreds by 2030. Any one of those disclosed with numbers would move this analysis.
| A gigawatt-class orbital cluster before 2033 | under 10% |
| Before 2040 | ~1 in 3 |
| Orbit ever reaches single-digit percent of global data center load | ~45% |
The last number is the honest one. Terrestrial cost curves keep moving while orbit catches up, and the ground gets to use every improvement that orbit does.
| When | What | Watch for |
|---|---|---|
| End Aug 2026 | Starship Flight 14 | First V3 satellites into orbit and a ship catch. Until then every cadence curve is a drawing. |
| Oct 2026 | Starcloud-2 | Blackwell at 100× the first satellite's power. Radiator kg/m² and temperature, if disclosed. |
| Late 2026 | SpaceX pilot nodes on Starlink V3 | Any per-node kg/kW figure. Under 15 would be new. |
| Q1 2027 | Aetherflux first node | Power level and mass, not the press release. |
| Early 2027 | Suncatcher prototype pair | Cluster operation over optical links and real thermal data. |
| 2027 | FCC action on the 1M-satellite filing | Whether the six-year milestone waiver is granted. |
| 31 Dec 2027 | Count Block 3 cargo flights | 25 or more: doubling curve, dates move in. Under 10: Falcon 9 curve, first gigawatt ~2040. |
A flown array wing above 200 W/kg, and a flown pumped-loop radiator under 4 kg/m² at over 50°C. Together they cut the launch count by a third. Neither has appeared in any filing.
The single most useful signal is the 2027 flight count. It does more work than every announcement on slide 21 combined, and it is the number SpaceX has missed by the widest margin so far.
Epoch AI, "Total cost of ownership of a one-gigawatt AI data center" and CSV, May 2026 · Epoch AI, "OpenAI Stargate: where the US sites stand", Apr 2026 · Distilled / Cleanview, Stargate vs New Carlisle, Jun 2026 · Silicon Report, Stargate Abilene profile, Jul 2026 · Sightline Climate via Bloomberg on interconnection, May 2026 · iMasons report via constructionowners.com, Jul 2026 · EnkiAI on ERCOT queue, Jun 2026 · SAVRN operator brief on behind-the-meter gas, Jun 2026 · Investing.com on Bernstein and Nvidia $/GW, Nov 2025 · Orennia, "What it costs to build a 1 GW data center", May 2026 · FCC DA 26-113, SpaceX Orbital Data Center application, Feb 2026 · Introl, filing analysis and Musk quote, Feb 2026 · TNW on SpaceX S-1 risk factors, May 2026 · DCD, SpaceX million-satellite filing and Google projection, 2026 · Luminix, "Data centers in space: feasibility and economics", Jul 2026 · Spaceflight Now, Starship V3 first flight, May 2026 · SpacePolicyOnline, Flights 13 and 14 and HLS, Aug 2026 · Wikipedia, "SpaceX Starship" and Starlink launch list, May 2026 · nexi.fund and Orbital Radar, launch cost comparisons, 2026 · New Space Economy on Musk's $/kg claims, Jun 2026 · SpaceOrbitals on Starship V3 costs, 2026 · Joule / ScienceDirect on high-specific-power arrays, Nov 2025 · NASA NTRS 20150006022 · JPL APSA via science.gov · SBIR.gov 388460 · Spectrolab WCPEC-4, 2006 · arXiv 2310.19430 · pv Europe on silicon space modules, Aug 2026 · eoPortal "Orbital Debris", 2026 · ESA Space Environment Report 2026 via FODNews · Space.com and Starpath on Starlink manoeuvres, Jul 2026 · The Register on Starlink V3 timing, Jul 2025 · GAO-26-108556 and GAO-23-106021 on NASA major projects.
Solar margin factor ×1.35 · Starlink and ISS array areas · radiator net flux at 60°C · radiator kg/m² for each column · 35% loop allowance · stripped compute kg/kW · 20% shielding · power conditioning kW/kg · harness mass · structure fractions · propulsion budget · Starship 85 t to dawn-dusk orbit · both cadence ramps and the 30% share · debris strike rates · refresh mass and degradation · PV-plus-storage cost · array $/W at wing level · radiator and bus $/kg · integration premium · world upmass ~3,000 t/yr · the probabilities on slides 12 and 23.
Every one of these is on the slide where it is used, tagged. If you think one is wrong, change it and rerun the ledger on slide 8; nothing else in the deck is hidden behind it.