SpaceX Says Orbital AI Will Beat Ground Data Centers on Cost. The Math Needs Launch Prices to Fall 98%.

One megawatt of orbital compute needs 5.7 Starmind satellites at roughly 7.5 tons each. At today's commercial Starship price, launch alone costs more than the entire ground data center it would replace.

SpaceX's CFO picked the most boring room in finance to make the most ambitious claim in computing, telling investors at the Goldman Sachs Communacopia conference on September 10 that orbital AI data centers "will eventually be cheaper than running data centers on the ground," citing power shortages, water constraints, and community pushback on Earth. Nobody in that room ran the arithmetic. We did. At the commercial Starship prices SpaceX actually charges today, getting one megawatt of compute into orbit costs about six times more than building the ground data center it would replace, and that is the generous version of the math, the one that counts launch alone and has not yet added the satellite itself.

Start with what is real. SpaceX and NVIDIA announced in August that the first Starmind AI1 satellite will carry a space-optimized version of the Vera Rubin NVL72 rack-scale system: 72 NVIDIA chips, Vera CPUs for agentic orchestration, roughly 175 kilowatts of compute per satellite, the equivalent of one ground-based AI rack. Each satellite generates about 210 kilowatts from a solar array with a 70-meter wingspan, dumps its heat through 1,700 square feet of liquid radiators, and flies a sun-synchronous orbit that keeps it in sunlight 98 percent of the time. NVIDIA's own release confirms the plan: "SpaceXAI's planned first-generation Starmind AI satellite will be based on the optimized NVIDIA Vera Rubin NVL72 rack-scale system." First launch is targeted for the fourth quarter of 2027, with what Elon Musk calls "significant scale" in 2028. The FCC filing goes further, up to one million of these satellites, so sit with the replacement math, because the constellation never stops needing rockets.

One million. UC Berkeley aerospace engineer Christopher Smith ran the treadmill calculation: if each satellite lasts five years, SpaceX must replace 200,000 satellites a year, about 60 per Starship flight, which works out to more than 3,300 launches annually, or roughly one Starship launch every 2.6 hours, every day, forever, just to keep the network alive. SpaceX's own IPO prospectus sketches an even bigger treadmill: deploying 100 gigawatts of orbital compute per year would mean moving roughly one million metric tons to orbit annually, requiring 5,000 to 10,000 Starship launches a year. That is 14 to 27 launches a day. The scale is staggering. Whether the economics close is a separate question, and it is the one nobody has actually computed.

The Comparison Nobody Published

Here is the apples-to-apples frame, with one deliberate simplification: exclude the silicon, because the Vera Rubin chips cost the same whether they sit in Nevada or in low Earth orbit, so the fair contest is everything around them, per megawatt of compute, over ten years.

On the ground, the numbers are well documented. Cushman & Wakefield's 2026 Data Center Development Cost Guide puts all-in greenfield construction at $17.6 million per megawatt in the US and Canada, chips excluded, with power infrastructure the single largest cost category. Add ten years of electricity, one megawatt running 8,760 hours a year at the US commercial average of about seven cents a kilowatt-hour, and you get roughly $6.1 million. Total ground cost per megawatt per decade: about $24 million, the number doing the heavy lifting for every hyperscaler on Earth, and it is public.

In orbit, the inputs are estimates. SpaceX published capabilities, not a bill of materials, so every number below is extrapolated from the announced specs. One megawatt of compute means 5.7 satellites at 175 kilowatts each. Each satellite, per the announced specs, needs a 210-kilowatt solar array (about 1.75 tons at modern roll-out-array densities), 158 square meters of radiators (about 2.4 tons), the space-optimized compute rack (about 1.5 tons, and Musk says it is simpler and lighter than the ground version), plus structure, batteries, propulsion, and communications (about 2 tons). Call it 7.5 tons per satellite, 43 tons per megawatt. Launch price today is the real one: Voyager Technologies' SEC-filed 10-K discloses a $90 million price for a future commercial Starship launch, which works out to roughly $600 per kilogram at full 150-ton loads. The aspirational price is Musk's long-standing $10 million per flight target, about $100 per kilogram, while satellite hardware cost, the murkiest input, gets two cases: $8 million per satellite as a baseline for a space-hardened 70-meter-wingspan machine, $2 million as Musk's "lower cost" mass-production scenario. And the constellation's five-year replacement cycle, per Smith's analysis, means two full deployments per decade.

Ten-year cost per megawatt of compute, silicon excluded, 5-year satellite life
Scenario	Launch ($/kg)	Hardware per sat	Space total	Ground total	Space premium
Today's commercial Starship	$600	$8M	$142M	$24M	5.9x
Aspirational Starship	$100	$8M	$100M	$24M	4.2x
Aspirational + cheap sats	$100	$2M	$31M	$24M	1.3x

The launch-cost story is a distraction. That is the first surprise. Even at Musk's dream $100 per kilogram, with $8 million satellite hardware, orbital compute costs four times ground compute, because the satellite hardware is the load-bearing number, not the rocket. Which produces the second, sharper surprise: with $8 million hardware, orbital AI cannot break even even if launch were free. Two deployments times 5.7 satellites times $8 million is $91 million against a $24 million ground total. Zero-dollar launch still loses. Nobody at the Communacopia said that part out loud.

The $14/kg Number

So give Musk everything he is asking for on the hardware side. Assume satellite mass production drives the per-satellite cost down to $2 million, roughly the Starlink playbook. What launch price then closes the gap? Set two deployments of 5.7 satellites at 7.5 tons each, plus $2 million hardware, equal to $24 million: the answer is about $14 per kilogram. Today's commercial Starship price is roughly $600 per kilogram. The break-even launch cost is one-forty-third of the price SpaceX charges right now, a 98 percent reduction from today's commercial rate, and it requires Musk's $2 million satellites too. There is exactly one scenario in the table where space wins: $100/kg launch, $2 million satellites, and a ten-year satellite life so the hardware is deployed only once ($15.7 million against $24 million). All three at once. None of the three is in the current plan.

That last caveat matters, because the plan is explicit about the replacement cycle. The prospectus math assumes satellites die fast enough to need 5,000 to 10,000 launches a year. Every year you add to satellite life cuts the decade cost nearly in half. Radiation-hardened Vera Rubin chips in a sun-blasted orbit are unproven, and the history of space hardware says the first generation fails early and often. If the life extension happens, the hardware cost assumption gets harder, not easier, because rad-hardened long-life satellites cost more, not less.

The Strongest Case Against This Article

Johnsen was not really arguing about dollars per megawatt, and the strongest version of his claim deserves a fair hearing. Terrestrial compute is hitting walls that money cannot remove: grid interconnection queues stretching five years or more, utilities demanding nine-figure deposits before studies even begin, and powered land in primary US markets hitting $584,000 per megawatt this year, 51 percent above last year, per the same Cushman & Wakefield report. In a sun-synchronous orbit, the power is free, continuous, and unpermitted, the cooling dumps heat straight into the vacuum, and no county board votes on your radiators. The honest comparison is not orbital compute versus a ground data center in Texas; it is orbital compute versus ground compute that never gets built. On that frame, "eventually cheaper" means cheaper than the missing megawatts, which is a claim about constraints, not price sheets.

Then there is the track record of the man making the claim. Musk's timelines are famously fiction and his physics famously real. Starship did not cost $10 million a flight, but it did drag launch economics from $18,500 per kilogram toward triple digits. Starlink's per-satellite cost fell roughly an order of magnitude across generations. If anyone can mass-produce a 70-meter solar-winged AI rack, it is the company that already mass-produces 2.5-ton communications satellites by the hundred. The hardware side of our equation, the number that dooms every scenario above, is also the side where SpaceX has repeatedly embarrassed skeptics. Dismiss the $2 million satellite as fantasy and you are betting against the firm's core competency.

What This Does Not Prove

Honest accounting, because several inputs are estimates wearing lab coats. SpaceX has not published Starmind's mass or cost, so the 7.5-ton and $8 million figures are engineering extrapolations from the announced specs, and the error bars are wide. Smith's own 60-satellites-per-Starship figure implies satellites of only 2 to 2.5 tons, which is dramatically lighter than our estimate and would improve the space math; we kept the heavier number because the 210-kilowatt solar array and 1,700 square feet of radiators have to weigh something, but if SpaceX's packaging is better, the gap narrows. We ignored ground-side operating costs beyond electricity (water, staff, maintenance, which help space) and orbital operating costs entirely (station-keeping propellant, collision avoidance, insurance, which help ground). Ground electricity at seven cents is the national commercial average; AI operators signing long-term PPAs pay less, which widens the gap again. And the five-year life is Smith's assumption, not a spec: longer life helps space, but rad-hardened longevity costs money, which hurts. None of these move the conclusion under today's launch prices. They move it under the aspirational ones, where the honest answer is that the contest is genuinely open.

What You Can Do

If you are evaluating SpaceX's AI ambitions as an investor, stop watching launch price and start watching two numbers: per-satellite hardware cost and actual replacement cadence. The $14-per-kilogram break-even shows launch would need to fall 43x even after hardware gets cheap; the watchable milestones are the first AI1 failure modes and whether generation two gets cheaper or heavier. If you are in energy or infrastructure policy, treat orbital data centers as decade-scale optionality, not a 2027 cost story: the relevant policy question is permitting relief for terrestrial compute, which is what Johnsen was really selling, because if ground megawatts were easy to build, the orbital pitch would not exist. And if you are just following the AI race, the number to remember is 2.6 hours: one Starship launch every 2.6 hours, in perpetuity, is the maintenance treadmill for the million-satellite vision. That cadence is either the greatest industrial machine ever built or the plan's undoing. There is no middle.

The Bottom Line

SpaceX's CFO says orbital AI will eventually be cheaper than ground data centers. The math says "eventually" is doing a heroic amount of work: with realistic satellite hardware, launch would need to cost $14 per kilogram, a 98 percent cut from today's commercial Starship price, and even free launches do not close the gap unless the satellites themselves get four times cheaper. That does not make the project delusional; it makes the interesting part the hardware factory, not the rocket. SpaceX has collapsed satellite costs before. But until a Starmind AI1 costs $2 million and lives ten years, the cheaper-data-center claim is a story about the future, told with the confidence of someone pricing the present.