On January 30, SpaceX asked the Federal Communications Commission for authority to launch up to one million satellites. In the same application, it asked to be excused from the rules that would require it to launch any of them on a schedule.1
The waivers are specific. Section 25.164 gives the holder of a non-geostationary license six years from grant to put half the constellation in orbit and nine years to finish it.2 Section 25.165 backs that deadline with a surety bond that is forfeited on a miss.3
SpaceX asked to be released from both.1
Nine months later, Elon Musk put a date on it anyway. Orbital compute deployed per month "probably exceeds terrestrial around the end of 2028 and then far exceeds it in 2029," he wrote on October 30, and terrestrial compute "will be peanuts compared to orbital compute."4
The forecast has a date. The filing has none.
The physics was never the problem
The argument about orbital data centers has been a physics argument. Sunlight in the right orbit is free and nearly continuous, which the grid is not. Against that, heat leaves only by radiation; radiation degrades the hardware, and the debris field grows every year.
Free Solar, Fatal Math ran the second half in June and got it wrong.5 It modeled a gigawatt-class facility as a single structure with a half-square-kilometer cross-section, found roughly fifty impacts from centimeter-class debris over five years, and treated that as fatal. The Poisson arithmetic holds. The monolith does not. SpaceX is proposing a fleet. At AI1's 120 kW apiece, that is roughly 8,300 satellites per gigawatt, and fifty hits spread across 8,300 spacecraft is a replacement line item.
The correction moves the question off the physics entirely. The thermal design closes on its own numbers. The debris exposure is survivable. What decides it is cost per unit of delivered computation. SpaceX has filed the power-to-mass ratio, the orbit, the depreciation lives, and the launch vehicle. It has not filed that.
The physics survives. The debris math survives correction. The arithmetic has not been filed.
Launch against the power bill
The S-1 states the engineering assumption without hedging: satellites "carrying over 100 kilowatts of compute power per metric ton," and a long-run ambition of "deploying up to 100 gigawatts of power to orbit annually," which the filing itself converts into "approximately one million metric tons to orbit annually."6
One hundred kilowatts per ton is ten kilograms of spacecraft for every kilowatt of compute. Spread the launch cost of those ten kilograms over a five-year life and the annual launch cost per kilowatt is twice the price per kilogram.
| Launch price per kilogram | Annualized launch cost per kW |
|---|---|
| $7,000 — advertised rideshare rate to Sun-synchronous orbit | $14,000 |
| $5,700 — the $74 million list price over the 13 tonnes a Falcon 9 averages | $11,400 |
| $1,150 — the $15 million estimated internal cost of a reused Falcon 9, same 13 tonnes | $2,300 |
| $500 | $1,000 |
| $200 — Google's mid-2030s projection | $400 |
Sun-synchronous is the orbit the S-1 names for this constellation, so the advertised rideshare rate is the scope-matched one.7 The middle rows divide the dedicated Falcon 9 list price and the estimated cost of flying a reused one, by the 13-tonne average mission the S-1 reports.89 The last is Google's figure for Project Suncatcher, which Google frames as a projection under a sustained learning rate rather than a price anyone pays.10
Important
A terrestrial facility buying power at $0.08/kWh around the clock at a power usage effectiveness of 1.2 spends $841 per kilowatt-year. At the advertised rideshare rate, launch alone runs 16.6 times that electricity bill. At SpaceX's own estimated internal cost, it still runs 2.7 times. Launch has to fall below about $420 a kilogram just to undercut that power bill, before a dollar is spent on the spacecraft.
An orbital facility also never pours a foundation. JLL forecasts a 2026 global average of $11.3 million per megawatt for data-center shell and core, excluding the tech fit-out, which it puts as high as $25 million per megawatt for AI infrastructure.11
The question is how long to amortize it, and SpaceX's own books answer that. The S-1 depreciates data center infrastructure over 20 to 25 years.12
| Terrestrial expense | Per kW-year |
|---|---|
| Electricity at PUE 1.2 | $841 |
| Shell and core over 20 years | $565 |
| Combined benchmark | $1,406 |
That benchmark is a floor. It omits land, financing, maintenance and the interconnection queue that is the real constraint on the ground, all of which would raise it and widen the budget orbit has to work inside. Amortizing the building over ten years instead of twenty raises it to $1,971. It is generous to orbit in exactly one place: it assumes a kilowatt in orbit does as much useful work as a kilowatt in Virginia.
Against $1,406, launch must come in under about $703 a kilogram. At Google's projected $200, launch consumes $400 and leaves $1,006 per kilowatt-year, or roughly $5,000 per kilowatt across five years.
For a 120 kW satellite that is about $600,000 per spacecraft, or $940,000 on the ten-year reading of the building. Both sides exclude the accelerators, so the comparison holds. That residual is the entire budget for the bus, the solar arrays, the radiators, the coolant loops, the laser terminals, radiation hardening, ground segment, operations and loss.
Whether SpaceX can build a 70-meter compute satellite for the price of a house is a real question with a real answer. The S-1 does not give it.
The radiator is not where it breaks
Presenting the AI1 design in June, Musk gave the power budget as "150kW peak, 120kW sustained," and director of satellite engineering Ian Dahl said that experience with xAI showed "we can support 120kW of average compute." The design assumes solar arrays at roughly 250 watts per square meter and radiators at about 1,400 watts per square meter, with 110 square meters of deployable liquid radiator and redundant pumping loops.13
Those last two numbers settle their own ambiguity.
A 110-square-meter panel at 1,400 W/m² rejects 154 kW, which is the peak load. So 1,400 is the density of the panel, not of one radiating face, and a free-flying deployable panel sheds heat from both faces. Each face carries 700 W/m², which at an emissivity of 0.9 against an ideal cold sky needs a surface around 69°C. Ordinary spacecraft radiator territory.
Important
Read as one-sided, the same figure would demand about 134°C, which is where the design starts to look impossible — and nothing SpaceX has published says which reading is intended.
The engineering problems sit around that number rather than inside it. The coolant loop has to move heat out of the accelerators and deliver it to a panel above 69°C, so the chips run hotter still. Earth fills much of the sky at a few hundred kilometers and radiates back into both faces, raising the effective sink temperature. And the radiator, the pumps and the deployment hardware all come out of the same ten kilograms per kilowatt.
The GAO's April assessment is blunt that "cooling solutions at this scale are also unproven" and that large orbital data centers would need "solar arrays that are larger than any launched and assembled in space as of April 2026."14 Its own illustrative satellite pairs a 10,000-square-foot array generating "up to a few hundred kilowatts" with up to 5,000 square feet of radiator, roughly twice the radiator area per kilowatt that AI1's published numbers imply.14
The number to audit is mass, and it is the one number missing.
On SpaceX's own assumptions, a gigawatt of continuous power needs about 4 square kilometers of solar array and 0.71 square kilometers of radiator. At ten kilograms per kilowatt, that gigawatt weighs 10,000 tonnes — 2.5 kilograms for every square meter of array, with radiators, structure, chips, coolant, batteries, propulsion and laser terminals all paid out of the same 2.5 kilograms.
SpaceX has not published a mass for AI1. Without it, the 100 kW-per-ton assumption cannot be checked against the one satellite that exists.
Fifty hits, eight thousand satellites
Modularity changes where the debris risk lands. A gigawatt-class cross-section expects about fifty centimeter-class impacts over five years, and at 120 kW per satellite, a gigawatt is about 8,300 spacecraft. If each impact disables exactly one, the direct loss is 6 MW over five years, 0.6 percent of the fleet's capacity.5
What modularity does not do is make the debris go away.
A centimeter fragment disables a satellite by breaking it, and a broken satellite sheds fragments into the shell where its 8,299 siblings fly. SpaceX's FCC application asks for shells from 500 to 2,000 kilometers, which runs straight through the band where dead hardware and old fragment clouds already pile up.1 ESA's models put about 1.5 million objects between 1 and 10 centimeters in orbit against roughly 47,000 that are tracked and cataloged.15 The exposure is correlated by construction: every satellite in a shell shares it.
The operational load is already visible. In 2025, the roughly 9,600 Starlink satellites performed "over 1,000 automated collision avoidance maneuvers per day."16 That is about 38 maneuvers per satellite per year. The orbital data center application asks for up to a million satellites.
In June, Free Solar, Fatal Math closed on the insurance market: underwriters were being asked to cover AI hardware in orbit and the conversation was still about "whether the risk can be modeled, rather than what the premium should be."17 The S-1 answers that. SpaceX does "not maintain as much insurance coverage as many other companies do, and in some cases... not maintain any at all, including with respect to our in-orbit satellites, which we currently do not insure and do not expect to insure in the future."18
The unpriceable risk does not block the project. It moves onto SpaceX's balance sheet, where no underwriter has to write a number, and no reader gets to see one.
The rack and the building retire together
A terrestrial data center depreciates the building over decades and the servers inside it over a few years. When the servers are done, you pull them, and the building keeps earning. SpaceX's own schedule puts data center infrastructure at 20 to 25 years and servers and networking equipment at 5 to 6.12
In orbit, they are one asset. AI1's 150 kW peak is roughly one Nvidia rack, and the first-generation design is a space-optimized Vera Rubin NVL72.19 The satellite is the rack. It is also the building, the switchgear, the cooling plant and the floor.
The S-1 names the consequence as a risk factor: "the useful life of our satellites is inherently shorter than that of the information technology systems and infrastructure they host," which "may truncate the effective lifespan of those underlying information technology systems and infrastructure."12 The same schedule puts satellites at 3 to 5 years — broadband satellites at five, first-generation mobile satellites at three.12
SpaceX plans the other squeeze too, "shifting older hardware to lower intensity workloads as performance characteristics evolve," and concedes that retirements "may occur sooner than our estimates for the useful lives of our satellites."12 So the spacecraft can outlive the economic value of its accelerators, or the accelerators can outlive the spacecraft. Both squeeze the same residual.
A constellation that has to rebuild its revenue-generating base every five years has been the structural problem with this business since part three.20 The hardware inside runs the same race, depreciating against the next generation faster than against the calendar.21
| Economic life at $200/kg | Annual launch cost per kW | Left of the $1,406 benchmark |
|---|---|---|
| 2 years | $1,000 | $406 |
| 3 years | $667 | $739 |
| 5 years | $400 | $1,006 |
Three years is the bottom of SpaceX's own range, and the build cost amortizes on the same shortened schedule. The metric that settles this is discounted revenue per kilowatt across the life of the spacecraft, and nobody has published it.
A year of launches every month
Musk's crossover is a monthly rate, which sounds modest until it is converted into tonnage. JLL projects global data center capacity rising from 103 GW to 200 GW by 2030, nearly 100 GW of additions, roughly 20 GW a year, including conventional non-AI facilities.22 Matching that in orbit at SpaceX's power-to-mass ratio looks like this.
| Orbital capacity per year | Mass to orbit | 100-tonne Starship flights |
|---|---|---|
| 1 GW | 10,000 t | 100 |
| 10 GW | 100,000 t | 1,000 |
| 20 GW | 200,000 t | 2,000 |
| 100 GW | 1,000,000 t | 10,000 |
One gigawatt a year is 10,000 tonnes, more than half of everything currently in Earth orbit, launched every year.15
Twenty gigawatts a year is 200,000 tonnes, about 90 times the 2,213 tonnes SpaceX launched in 2025.8 The monthly framing implies roughly 1.67 GW a month by late 2028, which is 16,700 tonnes and 167 fully loaded Starship flights every month. SpaceX flew 170 missions in all of 2025.8
A year of launches, every month, on a vehicle that has not yet flown a 100-tonne payload.8
Important
The filing is more conservative than the forecast. The S-1 expects to "begin deploying our orbital AI compute satellites as early as 2028," says the remaining challenge "is one of scale," and states that about 10 gigawatts a year "would still enable a commercially attractive AI compute business."6 A tenth of the headline rate, starting the year the crossover is supposed to happen.
The objection that lands
Asking the FCC to waive a deployment milestone is ordinary. The rule was written for constellations of hundreds; nobody could launch 500,000 satellites in six years, and any applicant at this scale would request the same relief. Granting it says nothing about whether SpaceX believes its own forecast. That is the point rather than a rebuttal of it: the forecast and the obligation are different objects, and only one of them carries a penalty.
The harder objection is that this entire analysis prices a kilowatt while SpaceX is selling tokens. Accelerators in a vacuum with a dedicated radiator face a different thermal limit than accelerators in a hot aisle; a dawn-dusk Sun-synchronous orbit gives near-constant power without the battery mass; 23,000 inter-satellite lasers are a network a terrestrial operator has to lease.6 If a kilowatt in orbit delivers more useful computation than a kilowatt in Virginia, every table above understates the case, and the benchmark should move in SpaceX's favor.
Which is why the cost per delivered token is the only number that settles it, and why its absence is the finding rather than a gap in the reporting.
Google reached the same structure independently and the same ceiling. Its own economic claim for Project Suncatcher is that at $200 a kilogram the cost of launching and operating a space-based data center "could become roughly comparable to the reported energy costs of an equivalent terrestrial data center on a per-kilowatt/year basis."10 Parity with the electricity line alone, from the other serious party, with the building, the bus and the radiators still to be paid for.
Note
Three of the four pieces in this series have come down against the orbital case, and a fourth that found the case closing would cost the arc its spine. The correction above runs the other way: the debris argument that gave part two its title does not survive a modular fleet.5
The number
SpaceX built its listing on the economics of cheap tokens, and the market paid a trillion dollars for optionality on them.23 The S-1 argues that "lower cost per token enables more frequent model training, larger and more sophisticated models, longer chains of processing," and that whoever gets there first wins the race.6
Then it filed kilowatts per ton.
The first number is the one that decides whether orbit beats the ground: launch, spacecraft, thermal systems, networking, depreciation, debris exposure and replacement, divided by useful tokens delivered.
In June, the people who price catastrophe for a living said the market was still working out whether the risk could be modeled at all. The company then told the SEC it does not insure its satellites and does not intend to. It asked the Commission to be released from the only deadline anyone could have held it to. Then it forecast the crossover anyway, with a year attached.
Everything in that forecast is in the filings except the arithmetic.
A note on the math: annualized launch cost is (10 kg/kW x price per kg) / 5 years. Terrestrial electricity is 8,760 h x $0.08/kWh x 1.2 PUE. Radiator temperatures use the Stefan-Boltzmann law at emissivity 0.9 against an ideal 0 K sink, so real temperatures in low orbit, with Earth infrared and albedo loading, run higher. Debris figures reuse the flux and cross-section inputs from Free Solar, Fatal Math (about 2x10^-5 impacts/m²/yr for objects 1 cm and larger, about 10^-7 for 10 cm and larger, 0.5 km²) and assume one satellite lost per impact with no cascade, an intentionally optimistic bound.
This is analysis and commentary, not legal or investment advice.