SpaceX's orbital data center runs on free sunlight. "Free" is doing a lot of work.

The hardest problem in building a data center in 2026 is not the chips. It is a transformer with a multi-year lead time, an interconnection queue holding more proposed capacity than the entire existing United States grid, and a permitting process that treats a gigawatt like an act of parliament. The chips arrive in eighteen months. The electricity arrives in five years, if the county says yes. That gap — between how fast you can buy compute and how slowly you can power it — is now the binding constraint on the AI build-out, and everyone who builds data centers knows it.

SpaceX's answer, laid out in a half-hour video three days before the company's stock market debut, is to leave. The satellite is called AI1, and it is the first detailed look at what the company means when it talks about data centers in orbit: a 70-meter spacecraft — wider, tip to tip, than a 747's wingspan — that is essentially a rack of AI compute bolted to a power plant, with no building, no transformer, and no queue. Elon Musk says designing it was easier than designing a Starlink satellite. He may well be right, and that is not where the hard part lives. The hard part lives where it always lives in energy: on the invoice.

What SpaceX actually showed

Strip the renderings and the AI1 disclosure is refreshingly specific, which deserves credit — most space-compute pitches to date have been a slide with the sun on it. The published design:

• A compute payload that sustains about 120 kilowatts on average and peaks at 150 kilowatts — roughly one top-end AI rack, in current terrestrial terms
• A solar array sized to about 150 kilowatts, built on technology from the Starlink V3 program and delivering a claimed 250 watts per square meter
• Around 110 square meters of deployable liquid radiators, with redundant pumping loops and micrometeoroid shielding, to reject the heat
• An interchangeable compute module, so the chips inside are not married to one vendor
• Laser links borrowed from Starlink for moving data, and a claimed 70 kilowatts of compute per ton of spacecraft
• A planned factory — Gigasat, in Bastrop, Texas, on more than a thousand acres — with production targeted to begin around the end of 2027

Two prototype satellites are reported to be slated for launch in early 2027, and the company has talked about reaching a gigawatt of orbital compute capacity per year of production shortly after, scaling by an order of magnitude from there — with Musk gesturing, further out, at "terawatts." Those last figures should be read the way all pre-production capacity figures should be read: as targets, announced the week of an IPO. The 30-to-50-satellites-per-Starship-launch figure circulating in the company's materials belongs in the same bucket, since it assumes a launch vehicle that is still scattering engine parts over the Gulf on its way home.

The case for orbit is an energy case

It is worth being honest about why this idea refuses to die, because the physics genuinely favors it in one specific way. A solar panel on Earth is a part-time worker. Clouds, night, latitude, and dust hold a good terrestrial site to a capacity factor of 20 to 25 percent, and the grid has to paper over the other 75 percent with storage or gas. The same panel in the right sun-synchronous orbit works essentially every hour of every day — no night, no weather, no seasons that matter. Watt for watt of installed panel, orbit delivers four to five times the kilowatt-hours, and it delivers them flat, which is exactly the shape of load an AI cluster wants.

Add what orbit removes: no land acquisition, no interconnection application, no substation, no transformer order sitting in a five-year backlog, no water permit, no neighbors. The entire bureaucratic apparatus that currently sets the pace of the AI build-out simply does not exist at 550 kilometers. When the pitch is framed that way — and SpaceX frames it exactly that way — it sounds less like science fiction and more like regulatory arbitrage with solar panels. Which is what it is.

On Earth the sun sets and the transformer takes five years. In orbit the sun is free and always on. Every other line on the invoice gets bigger.

Cooling is the tax

Now the lines that get bigger. Start with heat, because computing is the business of turning electricity into heat at industrial scale. On Earth, the atmosphere is a free cooling appliance: blow air, evaporate water, dump the heat into the sky. In vacuum there is no air to blow. Radiation is the only exit, which is why a 150-kilowatt satellite needs on the order of 110 square meters of liquid-loop radiator — a deployable structure nearly as consequential as the solar wings, with pumps that must run for the life of the spacecraft and plumbing that must shrug off micrometeoroids. SpaceX's design takes the problem seriously, and the redundant loops say its engineers know what a single stuck pump would mean: a rack you can never send a technician to, cooking itself at 550 kilometers.

That is the second tax: serviceability is zero. A terrestrial data center replaces failed boards every week of its life. AI1 gets no replacements, no upgrades, no swapped optics — every failure is permanent and every design margin flies as extra mass. The interchangeable compute module is a smart concession to this on the production line, letting SpaceX slot in whatever accelerator wins the next two years. It does nothing for a satellite already in orbit. And orbit adds a hazard the ground never sees: radiation, which corrupts exactly the kind of dense, cutting-edge silicon AI compute depends on. The company has said little about how much shielding or error-correction overhead AI1 carries; that silence is one of the more expensive unknowns in the design.

The arithmetic of scale

Here is the calculation the renderings do not show. One AI1 sustains about 120 kilowatts of compute — one rack, give or take. A single large terrestrial AI campus now under construction draws a gigawatt and up. To match it, you need on the order of eight thousand AI1-class satellites. That is not a constellation; that is a second Starlink — it took SpaceX six years of the most aggressive launch cadence in history to put roughly that many satellites in orbit. At the optimistic 30 to 50 satellites per Starship flight, a gigawatt of orbital compute is something like 170 to 280 dedicated launches, every one of them a vehicle that has not yet finished development. The 'terawatts' line, priced in these units, is millions of satellites. The word costs nothing to say.

Then the meter starts running on refresh. AI hardware depreciates the way fish does, not the way buildings do — terrestrial operators argue with their accountants over whether a GPU is worth anything after five years. A data center swaps its silicon and keeps its shell, its land, its grid connection. An orbital data center IS its silicon. When the chips age out, the whole spacecraft ages out: deorbit, rebuild, relaunch. Starlink already lives this churn, replacing its fleet roughly every five years, and SpaceX is the one company on Earth with proven muscle for it. But it means the launch bill is not a construction cost. It is an operating cost, forever.

What a kilowatt-hour costs up there

So price it, roughly, the way a grid planner would. At the claimed 70 kilowatts per ton, an AI1 masses around two tons. On today's launch economics — Falcon-class costs in the low thousands of dollars per kilogram — putting one up runs several million dollars before you have built the spacecraft, bought its compute module, or paid to replace it in five years. Spread a satellite's roughly five million kilowatt-hours of five-year output across plausible all-in costs and the orbital electron lands several times the price of the terrestrial one — against a US industrial grid price around eight cents, or against the solar-plus-storage that now clears tenders across India, the Gulf, and increasingly Africa at prices the hyperscalers would sign tomorrow if the interconnection existed.

Every assumption in that envelope bends in SpaceX's favor if Starship works as advertised. At the aspirational hundred-some dollars per kilogram that full reuse is supposed to deliver, the launch line shrinks by an order of magnitude and the orbital kilowatt-hour starts conversing with the terrestrial one — especially the terrestrial one that cannot actually be bought, because it is stuck in a queue behind 2,300 gigawatts of other applications. That is the honest version of the bet: AI1 does not need to beat the grid's price. It needs to beat the grid's wait, by enough, for customers whose alternative cost is measured in unlaunched products. There are worse bets. But note what the whole structure rests on — a launch vehicle whose cost curve is still a projection, sold by the company that builds it.

Who actually pays

The timing of the reveal was not subtle. The video landed on a Monday; the company starts trading on the Nasdaq this Friday at a hoped-for $1.75 trillion valuation, raising some $75 billion — with an unusually large slice, as much as 30 percent, reportedly reserved for retail investors. The use-of-proceeds language points squarely at Starship and the compute build-out. Which means the people funding the experiment of whether a data center belongs in orbit will include, to an unprecedented degree, households buying through brokerage apps. They should at least see the spec sheet priced before they do.

And the idea is suddenly fashionable beyond Hawthorne: Nvidia-backed Starcloud has flown demonstration hardware, and Google has reportedly explored its own orbital-compute concept under the name Suncatcher. When several balance sheets chase the same physics, the physics is usually real. So is the pattern from energy history: the grid wall that orbit so elegantly bypasses is, at bottom, a queue — transformers, permits, substations, paperwork. Queues are miserable, and they are also fixable, and an industry that can find $75 billion for satellites can find it for transformer factories. The most dangerous competitor to a data center in space is a boring reform to interconnection on the ground.

SpaceX has done the unglamorous parts right: a specific design, a named factory, a real thermal architecture, a credible heritage in mass-producing satellites that die young. That is more than this idea has ever had behind it. What it does not have yet is the only number that decides: the cost of a kilowatt-hour of compute, delivered to a paying customer, including the launch, the refresh, and the deorbit. If that number ever beats the grid — or merely beats the queue — the build-out leaves the ground for good. Until then, AI1 is the most rigorous expensive answer yet offered to a problem the ground could still solve cheaply, and the company selling it is, conveniently, the one that owns the rockets.