in short
Falcon 9 currently launches payloads to low Earth orbit for around $2,600–$3,000 per kilogram — down from roughly $54,000 per kilogram in the Space Shuttle era, a genuine 90%-plus collapse. The entire case for orbital data centers depends on that number falling by another order of magnitude, to somewhere near $100 per kilogram on SpaceX's Starship. That is a design target. It has never been achieved.
The technical case for orbital compute gets most of the public attention — free solar power, no atmosphere, a cold vacuum for cooling. The economic case gets far less scrutiny than it deserves, and the scrutiny that does exist, from engineers and analysts working the numbers directly rather than repeating the pitch, is considerably less settled than the momentum around the idea would suggest.
The Physics Problem Nobody Disputes
Heat has to go somewhere. On Earth, air and water carry it away from hardware efficiently and for free. In orbit, in a vacuum, neither convection nor conduction exists — heat can only leave through radiation, a slower and far more mass-intensive process. The International Space Station’s own thermal control system provides the only large-scale, real-world data available: its radiators reject roughly 126 kilowatts of heat from a combined radiator mass with a specific heat rejection of about 13.5 watts per kilogram (independent technical analysis, Angadh Nanjangud). That figure is the actual constraint an orbital data center has to design around, and it means every watt of usable compute has to be paired with real, launched mass dedicated purely to getting rid of heat — mass that a terrestrial data center simply doesn’t need to carry at all.
Applying that constraint to a realistic AI compute payload produces a sobering result. One detailed independent techno-economic analysis found that support infrastructure — radiators, structure, power systems — outweighs the servers themselves by nearly four times, meaning a deployment that a company might claim needs a single launch actually requires on the order of twenty-two, pushing costs from roughly $5 million to somewhere between $110 million and $2.2 billion depending on which launch-cost assumption is used (Angadh Nanjangud). The analysis is careful to note this doesn’t mean orbital data centers have zero value — but it does mean thermal management, not launch access or solar power, is the variable that actually dominates the unit economics, and it’s the variable getting the least public attention.
The Launch Cost Assumption Everything Else Depends On
Every optimistic orbital compute projection rests on the same load-bearing assumption: that Starship gets launch costs down to roughly $100 per kilogram, alongside solar panel efficiency above 40% and satellite hardware surviving a full five-year operational lifespan in LEO’s radiation environment (Introl). None of those three conditions has been demonstrated yet. Today’s actual number, on a reused Falcon 9, is closer to $2,600–$3,000 per kilogram for a dedicated launch — genuinely remarkable progress from the Shuttle era, but roughly twenty-six to thirty times higher than the number the orbital compute thesis needs to work (Orbital Radar).
One founder building directly in this space was unusually candid about the gap: current launch costs run around $7,000 per kilogram on a shared Falcon 9 ride, against Musk’s own stated target of roughly $10 per kilogram — “a big, couple orders of magnitude” gap that has to close before the economics work at all (The Register). An independent modeling tool built by an engineer at Varda Space Industries reached a comparable base-case conclusion using different assumptions: orbital compute currently costs roughly three times more per watt than ground-based equivalents. As that engineer put it, the physics doesn’t immediately rule it out, but the economics are unforgiving (Introl).
The European Space Policy Institute has been similarly blunt, describing the dependence on a sub-$10-million-per-flight Starship price as “unrealistic in the near-term” — a direct challenge to the specific number multiple orbital compute business cases are quietly built on (Introl).
What the Skeptical Financial Analysis Actually Shows
For scale, it helps to know what a gigawatt of AI infrastructure actually costs on the ground: former Google CEO Eric Schmidt has put the figure at roughly $50 billion per gigawatt of hardware, software, and data centers combined — meaning the 100 gigawatts SpaceX’s S-1 targets annually would, on Earth-equivalent terms, represent on the order of $5 trillion in infrastructure value (The Singju Post, transcript). That framing cuts both ways: it shows why the orbital opportunity, if it works, is enormous — and why a $170 billion valuation for SpaceX’s entire orbital AI segment looks conservative relative to the stated goal, or the goal looks wildly disconnected from anything achievable in the near term, depending on which side of the launch-cost assumption you believe.
One nuance worth holding alongside the skepticism above: orbital compute is not purely theoretical today. Google’s Project Suncatcher already runs TPU clusters live on Planet Labs satellites, a real commercial deployment generating revenue now (PodcastAlpha, summarizing Moonshots with Peter Diamandis EP #266). That doesn’t resolve the thermal and launch-cost problems detailed above — Suncatcher’s workloads are almost certainly far smaller and less power-dense than a full AI training cluster — but it does mean the honest position is “unsolved at the scale being promised,” not “unsolved, period.”
The clearest illustration of how thin the current investment case is comes from outside analysis of SpaceX’s own numbers. One detailed valuation breakdown put the core launch and Starlink businesses at roughly $611 billion using demonstrated, current performance — and valued the entire orbital AI compute opportunity separately at only about $170 billion, a figure reportedly described by Morningstar as closer to the value of a call option on future commercialization than a business with proven unit economics (independent analysis, citing Morningstar). A call option is a legitimate way to value real optionality. It is also, definitionally, an acknowledgment that the underlying economics are not yet proven — a meaningfully different claim than the one implied by most public orbital-compute messaging.
There’s a physical constraint underneath the financial one that deserves its own mention: the satellites large enough to carry meaningful AI compute payloads exceed what Falcon 9’s payload fairing can carry at all, meaning the entire near-term timeline depends on Starship succeeding operationally, not just economically — and Starship’s first payload-deployment test only occurred in May 2026, after only a handful of prior test flights (independent analysis).
The Challenge: This Piece Is Not Trying to Resolve the Question
Intellectual honesty requires stating plainly what this analysis does and doesn’t show. None of this proves orbital compute can’t eventually work. Launch costs have already fallen by more than 90% once, within a single engineer’s career, through reusability — a trajectory nobody would have confidently predicted in advance either. It’s entirely possible thermal engineering, solar efficiency, and launch costs all improve enough, on overlapping timelines, for the unit economics to close. What the evidence supports today is narrower and more useful than either extreme: the case for orbital compute is not yet an economic case. It is a set of specific, named, quantified conditions — a $100/kg Starship, 40%+ solar efficiency, five-year radiation survival, thermal mass ratios far better than the ISS currently achieves — that have to move together before it becomes one.
Conclusion
The orbital compute thesis is a real possibility. It is not yet a real economics. Anyone allocating capital against it today is not investing in a proven infrastructure category — they’re pricing a call option on three separate engineering problems resolving in the same direction, on a timeline nobody serious is currently willing to guarantee.
source
covers
ISS thermal data; radiator mass constraints; launch multiplier analysis
Three-assumption dependency; Varda engineer’s cost model; ESPI critique
Historical and current $/kg figures
Founder-level candor on the cost gap
SpaceX valuation breakdown; “call option” framing; Starship fairing constraint
