Data centres in orbit: the plug is free and the heat is the bill
SpaceX has told American regulators it wants to put AI data centres into orbit. The Federal Communications Commission has accepted its application for up to one million satellites for review.
The pitch is power. In the right orbit the sun almost never sets, and there is no queue for the grid. But every watt a chip uses comes back out as heat, and in a vacuum there is no air or water to carry it away. Heat can leave only by glowing off a radiator, a panel that faces cold space.
So orbit is a contest of weight, because radiators and solar panels must be launched and launch is priced by the kilogram. My claim: by the end of 2028, nobody will report a compute satellite in service at SpaceX’s own target for power per tonne launched, and no rocket company will publish a price low enough to make the sums easy.
In orbit the plug is free, and the heat is the bill.
Explain it like I don't work in finance
Picture a laptop sealed inside a vacuum flask on a sunny windowsill. The sun charges it for nothing, but the flask that keeps tea hot also traps the laptop's heat.
The only way out is a large black fin poking out of the flask, glowing the heat away. A hotter fin can be smaller. And every fin goes up by courier, charged by the kilogram.
The orbit these plans favour is called dawn-dusk sun-synchronous. The satellite rides the line between day and night, so its panels face the sun almost all the time. SpaceX’s finance chief says a solar cell there gets about five times the energy it would on the ground.
The radiator is the hard part. The heat a panel sheds rises with its temperature to the fourth power, so a slightly hotter panel sheds far more. Starcloud, a start-up, works out about 633 watts per square metre for a panel at 20°C with the sun on one face. SpaceX assumes 1,400, which on my arithmetic means a panel near 67°C.
That is the same lever as on the ground. Warm coolant lets a hall stop using water; in orbit, warm coolant shrinks the radiator.
Weight is where the sums turn. SpaceX targets 100 kilowatts of computing for every tonne it launches. Radiators as heavy for their size as the International Space Station’s, even at SpaceX’s 1,400, would use that whole budget alone, before a single chip or solar panel.
The regulator noticed. In April the Securities and Exchange Commission asked SpaceX to disclose the total mass of its system. The final prospectus said the mass of each satellite was still being refined.
Then there is the fare. Google, which is designing its own orbital system, estimates today’s launch price at about $3,600 a kilogram. At that price its own sums put power in orbit at several times what an American hall pays for electricity.
Google cites $200 a kilogram as the point where launch could stop being the limit, and sees prices reaching it around 2035 if rockets keep getting cheaper at their past pace.
Orbit can still work. The free plug is simply paid for in kilograms, and no compute satellite in service has yet been weighed.
For investors
The plug in orbit is the sun, and it has no queue
The designs converge on one orbit, because it nearly removes the night. Google’s Suncatcher paper (arXiv, 22 November 2025) puts its satellites in “dawn-dusk, sun-synchronous low-Earth orbit” to “enable near-continuous power generation”. Starcloud’s FCC application of 4 February 2026 asks for up to 88,000 satellites “with a nominal 06:00 crossing time”, the same dawn-dusk line.
SpaceX’s application is far larger. The FCC’s Space Bureau accepted it for filing on 4 February 2026 (DA 26-113): up to one million satellites, between 500 and 2,000 km, in 30 degree and sun-synchronous orbits. Its finance chief, Bret Johnsen, said in an interview SpaceX filed with the SEC on 9 June 2026 that cells there get “roughly five times, if not more, energy per cell than they get terrestrially”.
The pull is the plug on the ground. Starcloud’s filing says the buildout of terrestrial energy in the western world “has not kept pace with the demand for AI”. SpaceX’s own risk factors tie the case to the same thing. If “advances in nuclear energy” or similar “alleviate infrastructure constraints”, the viability of its orbital compute “may be materially diminished”.
That is the American plug queue seen from above. Spectrum is not the gate either. Both ask for Ka-band radio only “on a non-interference, unprotected basis”, and both move their data mainly by laser.
Heat leaves only as light, so the radiator’s temperature sets its size
Almost every watt a chip draws becomes heat, and in a vacuum nothing carries it off by touch. SpaceX’s final prospectus (12 June 2026) says so plainly: “thermal control must be accomplished through radiation rather than convection and conduction.” The law that governs radiation makes the heat shed rise with absolute temperature to the fourth power.
Starcloud’s white paper (September 2024) works the sum. A black plate at 20°C with an emissivity of 0.92, radiating from both faces, emits about 770 watts per square metre. After it absorbs some sunlight and the Earth’s own glow, it nets 633 W/m².
SpaceX assumes far more. In the 9 June transcript, Elon Musk puts SpaceX’s figure at about 1,400 W/m², from radiators that radiate on both sides and sit “knife edge to the sun”. On my arithmetic, a shaded two-sided panel of the same emissivity reaches 1,400 W/m² only near 67°C.
A radiator sheds heat with the fourth power of its temperature
Heat radiated per square metre of a two-sided panel in shade, watts, by panel temperature
Going from 20°C to 67°C raises absolute temperature by about 16%, but sheds about 1.8 times the heat per square metre. That is the lever Heat and water found on the ground, pulled for a different prize. There, a warm loop at ASHRAE class W45 lets a dry cooler replace a cooling tower. In orbit, the warmer the coolant the chips accept, the smaller the radiator.
The International Space Station sits at the cold end. Its ammonia supply is set at 37°F, about 3°C, because it cools a crew. On my arithmetic its six main radiator units reject about 147 W per square metre of panel as designed.
Every kilogram is a launch bill, so the measure is kilowatts per tonne
Because mass is the cost, SpaceX states its goal as power per mass. Its prospectus says deploying 100 gigawatts a year “via satellites carrying over 100 kilowatts of compute power per metric ton” needs “approximately one million metric tons to orbit annually”. That is 10 kg launched for each kW of computing, the measure the rest of this piece uses.
The first design is smaller. In the 9 June transcript, Musk described a draft version-one satellite of 150 kW peak and 120 kW sustained, about one GB300 rack. The solar assumption is 250 W/m².
SpaceX’s web page for its first AI satellite, AI1, now gives a design “vehicle efficiency” of “75 kW / ton”. It does not say whether that counts compute or solar power, or tons launched. Read as compute per tonne launched, it is a quarter short of the target.
On my arithmetic, 120 kW needs about 86 m² of radiator at 1,400 W/m², and about 480 m² of solar array.
At 10 kg per kW, the whole satellite could weigh no more than about 1.2 tonnes. Radiator, panels, chips and frame must all fit inside it.
Johnsen put the first generation at “About 5 megawatts per launch”, or “roughly 200 launches for every gigawatt”. At the 100 tonnes a launch SpaceX targets over time, that is 50 kW per tonne, half the goal. SpaceX has not said what the first launches will carry.
Google’s best proxy is Starlink. Its paper puts a Starlink V2 mini at 575 kg and, on its own photometric estimate, about 28 kW of solar power. That is about 20.5 kg per kW on my arithmetic, twice SpaceX’s target for a satellite that would also carry the chips.
At today’s launch price, the free plug costs more than ground power
Google turns kilograms into money with a “launched power price”, the launch cost of each kW spread over the satellite’s life. It uses a current price of $3,600/kg, “based on Falcon 9 (reusable configuration)”. Over a five-year life, a Starlink-class satellite comes to $14,700 per kW per year.
Its comparison for American AI halls is power at $0.06 to $0.25 per kWh and a PUE of 1.09 to 1.4. PUE, power usage effectiveness, is total power entering a hall divided by the power reaching its computers. That gives $570 to $3,000 per kW per year.
At today's launch price, power in orbit costs several times a ground hall's power bill
Launch cost per kilowatt per year for a Starlink-class satellite, against the power bill of a US AI hall, US dollars
At $200/kg the same satellite falls to $810, inside the ground range. Google notes that $200/kg “is often cited, by SpaceX and others”, as the point where launch stops being the limiting cost. It projects prices could fall below it by about 2035 if SpaceX’s learning rate holds, which it says would take about 180 Starship launches a year.
The case is very sensitive to that price. Starcloud’s white paper assumes about $30/kg in the long term, roughly a 120th of Google’s current figure. I found no launch provider’s published price per kilogram for a whole rocket to low Earth orbit; the $3,600 is Google’s estimate.
The regulator asked for the mass, and the prospectus does not give it
The SEC’s staff went straight to the lever. Its comment letter of 24 April 2026 on SpaceX’s draft registration asked for the “power-to-mass ratios” behind the 100 gigawatt goal. It asked SpaceX to discuss “the total estimated solar collection area, total system mass, and on-orbit assembly requirements”. It also asked for “the material differences between connectivity satellites and AI compute satellites”.
The final prospectus (424B4, 12 June 2026) answered on design, not on mass. AI compute satellites need “substantially larger radiators for thermal management”. They need about “five times the solar array output compared to V3 satellite designs”. On the number asked for, it says: “The precise solar collection area, total system mass per satellite, and on-orbit assembly requirements associated with this goal continue to be refined.”
The next quarterly report did not fill the gap. SpaceX’s 10-Q for the quarter to 30 September 2026, filed on 22 October, does not mention orbital compute at all.
Heritage hardware shows why the missing number matters. NASA’s lunar habitat study (IEEE Aerospace, 2022) sizes radiators at the station’s areal density, 14 kg per m². At Starcloud’s 633 W/m², that is about 22 kg of radiator for each kW. At SpaceX’s 1,400 W/m², it is exactly 10.
At the space station's weight, the radiator alone uses SpaceX's whole budget
Kilograms launched per kilowatt
So a radiator of the station’s weight, run as hot as SpaceX assumes, would use the whole 100 kW per tonne budget on its own. The station’s own wings are far heavier: six units of 1,122.64 kg each for 70 kW, about 96 kg per kW. The target needs radiators several times lighter per square metre than the station’s, or run hotter than 67°C, or both. No filing I read says which.
What could make orbit cheaper than I think
The case against is strong and specific. The station’s radiators were designed decades ago for a crew, with a cold loop, eight panels to a unit and rotating joints. A compute radiator need not carry all of that, so 14 kg per m² is a heritage ceiling, not a law.
An independent preprint goes further (Gaalema, Indyk and Staley, arXiv, 9 April 2026; not peer reviewed). It argues that panels combining solar cells, chips and radiator could reach “about 500 W/kg”, or 2 kg per kW. Its 16 MW, 150 tonne satellite would just beat SpaceX’s target, at about 107 kW per tonne, though it assumes “custom solutions for all components”.
The chips may cope with space too. Google put its Trillium TPU in a proton beam: the memory showed irregularities only after 2 krad(Si), almost three times the 750 rad(Si) a shielded five-year mission expects. And SpaceX calls its 1,400 W/m² a very achievable goal.
These are limits on my reading, not yet evidence against it. The preprint is a design, Google’s result is a ground test and SpaceX’s figure is an assumption. The claim asks for a satellite in service at the target, or a published price, and on 2 November 2026 neither exists.
So the plug in orbit is nearly free, as promised. The flask still needs its fin, and every fin goes up by the kilogram. Until someone weighs a working one, the heat is the bill.
In orbit the building wears out before the chip
On the ground, the hall and its grid gear outlast three or four chip generations. In orbit the hardware gives out first. SpaceX’s risk factors say “the useful life of our satellites is inherently shorter than that of the information technology systems and infrastructure they host”, so it “must periodically launch replacement satellites”. Google spreads Starlink’s launch cost over a five-year life.
So the launch bill becomes rent. The radiator, the solar wings and the frame go up again with every replacement, not once. SpaceX says its first Starmind compute satellites fly in 2027. One flown then would, on that life, wear out around 2032. Google’s own curve does not reach $200 a kilogram until about 2035.
On Google's numbers, a compute satellite launched in 2027 wears out before launch reaches $200 a kilogram
Dated plans for orbital compute, a five-year satellite life, and Google's projection for launch prices
So on Google’s curve, the first generation is launched and relaunched before the price that makes the sums work arrives. Planet is contracted to deploy two Suncatcher prototypes, “targeting launch in 2027”, by its 8-K of 10 December 2025. Gwynne Shotwell, SpaceX’s president, said on its call of 20 October that it will “begin flying Mobile V2 and Starmind satellites in 2027”.
Each will show, for the first time, what a kilowatt of orbital compute weighs in service. Until one does, every gigawatt promised in orbit is a radiator nobody has launched.
What would prove me wrong
Two observations break the claim. An operator discloses, in a filing, a release or a regulator’s record, an AI compute satellite in service at 100 kW or more of compute power per tonne launched, by 31 December 2028. Or a launch provider publishes a price to low Earth orbit of $200 per kilogram or less by then.
SpaceX’s 10-Qs and 10-K, Planet’s and Google’s prototypes, and any Starcloud release will show the first. A published price sheet or contract would show the second. Better radiators in a paper do not break it; one weighed in service does.
Research notes
SpaceX has an initiation note resting on this claim: SpaceX. The ground-side mechanisms it leans on are in Heat and water and America rations the plug. Whether any of this is already in share prices is a question for the notes, not for the journal.
The claims on record, word for word
The test this piece stands behind, quoted word for word from the claims page:
In orbit the plug is free and the heat is the bill (2 November 2026)
Claim: “In orbit the power is nearly free and the heat is the bill: by 31 December 2028 no operator will report an AI compute satellite in service at SpaceX’s own target of 100 kilowatts of compute power per metric ton launched, and no launch provider will publish a price to low Earth orbit at or below Google’s $200 per kilogram threshold.”
Breaks if: “By 31 December 2028, SpaceX, Starcloud, Google or any other operator discloses in a filing, release or regulator’s record an orbital AI compute satellite in service at 100 kilowatts or more of compute power per metric ton of launched mass, or a launch provider publishes a list or contract price to low Earth orbit of $200 per kilogram or less.”
This piece absorbs no earlier piece, so it carries no older claims.
The figures behind the charts
The radiator physics:
| Measure | Figure | Source |
|---|---|---|
| Plate at 20°C, emissivity 0.92, two faces | 770.48 W/m² emitted; 633.08 W/m² net after 122.94 W/m² of sunlight (absorptivity 0.09) and 14.46 W/m² from the Earth | Starcloud (then Lumen Orbit) white paper, September 2024 |
| SpaceX’s design assumptions | Solar 250 W/m²; radiators “about 1,400” W/m², double-sided, “knife edge to the sun”; 150 kW peak, 120 kW sustained | Interview transcript, FWP filed 9 June 2026 |
| Panel temperature for 1,400 W/m² | About 340 K, 67°C (68°C allowing for the Earth’s glow) | My arithmetic: two faces, emissivity 0.92 |
| ISS main radiators | 70 kW for two loops; six units, each 23.3 m by 3.4 m and 1,122.64 kg; ammonia supply 37°F | NASA ISS Active Thermal Control System overview, undated |
| ISS, derived | 6,736 kg for 70 kW, about 96 kg per kW; 475 m², about 147 W/m² and 14.2 kg/m² | My arithmetic |
| ISS areal density used by NASA | 14 kg/m² | Schunk, Babiak and others, IEEE Aerospace 2022 (NTRS 20210026557) |
The curve in the first chart is 2 × 0.92 × 5.67 × 10⁻⁸ × T⁴, with T in kelvin: 581 W/m² at 0°C, 771 at 20°C, 1,069 at 45°C, 1,400 at 67.2°C and 2,023 at 100°C. The ISS point is total capacity over planform area at design, not a measured flux.
Mass per kilowatt:
| Measure | Figure | Source |
|---|---|---|
| SpaceX target | “over 100 kilowatts of compute power per metric ton”; 100 GW a year needs “approximately one million metric tons to orbit annually” | Prospectus (424B4), 12 June 2026, about p. 153; S-1/A No. 2, 3 June 2026 |
| SpaceX first generation | “About 5 megawatts per launch”; “roughly 200 launches for every gigawatt” | Johnsen, transcript filed 9 June 2026 |
| SpaceX first flights | “we’ll begin flying Mobile V2 and Starmind satellites in 2027” | Shotwell, third-quarter call, 20 October 2026 (ir.spacex.com transcript) |
| AI1 design figure | “75 kW / ton” vehicle efficiency; basis not stated | SpaceX AI1 web page, read 2 November 2026 |
| Starship capacity | “100 Metric tons per launch over time” | Roadshow slides, FWP filed 4 June 2026 |
| Starlink V2 mini | 575 kg; about 28 kW, Google’s estimate from about 105 m² of array | Google, Suncatcher paper, 22 November 2025 |
| Radiator at 14 kg/m² | 22.1 kg per kW at 633 W/m²; 10.0 at 1,400 W/m² | My arithmetic |
The 50 kW per tonne for the first generation is my arithmetic, 5 MW over 100 tonnes; if the first launches carry less mass, the figure is higher. A 120 kW satellite at 1,400 W/m² needs about 86 m² of radiator, about 1.2 tonnes at station density, and about 480 m² of solar at 250 W/m².
Launch price:
| Measure | Figure | Source |
|---|---|---|
| Current price used | $3,600/kg, Falcon 9 reusable, Google’s estimate | Google, Suncatcher paper |
| Threshold | $200/kg, “often cited, by SpaceX and others” | same |
| Launched power price, Starlink V2 mini, five years | $14,700 per kW per year at $3,600/kg; $810 at $200/kg | same, Table 1 |
| Other satellites at $3,600/kg | Starlink V1 $26,600; OneWeb $135,800; Iridium $124,600 per kW per year | same, Table 1 |
| US AI hall power | $0.06 to $0.25 per kWh, PUE 1.09 to 1.4: $570 to $3,000 per kW per year | same |
| Projection | Below $200/kg by about 2035 at a learning rate near 20%, about 180 Starship launches a year | same |
| Starcloud’s assumption | About $5 million a launch long term, about $30/kg | Starcloud white paper, September 2024 |
The SEC record:
| Document | What it says | Date |
|---|---|---|
| Staff comment 26 | Asks for “power-to-mass ratios” and “total estimated solar collection area, total system mass, and on-orbit assembly requirements” | 24 April 2026 |
| Staff comment 31 | Asks for “the material differences between connectivity satellites and AI compute satellites” | 24 April 2026 |
| Staff comment 3 | Asks SpaceX to “clarify when you may potentially begin deploying AI compute satellites” | 19 May 2026 |
| Prospectus | Mass per satellite “continue[s] to be refined”; “substantially larger radiators”; deployment “as early as 2028” | 424B4, 12 June 2026 |
| 10-Q | No mention of orbital compute | Quarter to 30 September 2026, filed 22 October 2026 |
Who is exposed if the heat is the bill in orbit
This maps who operates in each layer. It is not a recommendation, and naming a company is not a view on its shares. Figures in the piece are disclosures and published estimates used as evidence for the mechanism.
- The orbital compute builders
- SpaceX builds rockets and satellites and is designing an AI compute satellite. Starcloud builds compute satellites.
- The chips
- Google designs the TPU chips its Suncatcher prototypes would carry. Nvidia makes the GB300 and Rubin chips in SpaceX's reference design.
- The satellite maker
- Planet Labs builds satellites, including the two Suncatcher prototypes.
- On the other side
- Anyone counting orbital gigawatts announced as compute to come, and designs that rest on radiators lighter than any yet flown at this scale.
Sources: Google, "Towards a future space-based, highly scalable AI infrastructure system design" (arXiv 2511.19468, 22 November 2025); SpaceX's SEC record: roadshow free writing prospectus (4 June 2026), interview transcripts (9 June 2026), S-1/A No. 2 (3 June 2026), final prospectus (12 June 2026), SEC staff comment letters (24 April and 19 May 2026) and 10-Q for the quarter to 30 September 2026 (22 October 2026); SpaceX's third-quarter call transcript (20 October 2026) and its AI1 web page (read 2 November 2026); FCC public notice DA 26-113 (4 February 2026) and Starcloud's FCC application narrative (4 February 2026); Starcloud (Lumen Orbit) white paper, version 1.03 (September 2024); NASA's ISS Active Thermal Control System overview and Schunk and others, IEEE Aerospace 2022; Planet Labs' 8-K (10 December 2025); Gaalema, Indyk and Staley (arXiv 2604.07760, 9 April 2026). Figures as stated on the dates cited. Personal research, not investment advice.