Orbital Compute Economics

The cost structure

A terrestrial data center costs roughly 60 billion dollars per gigawatt of AI compute capacity, split between roughly 35 billion dollars in GPUs and silicon and roughly 25 billion dollars in land, power infrastructure, shell, and cooling. The silicon component will likely deflate over time as chips improve, but the non-silicon component is more likely inflationary, since power prices, land scarcity, and cooling complexity are all increasing rather than decreasing.

Orbital compute changes that second number dramatically. In space, power comes from solar and is effectively free, land does not exist as a cost category, and cooling happens by radiation into vacuum, also effectively free. What remains is dominated by launch cost, and the resulting estimate is roughly 5 billion dollars per gigawatt, a five-fold reduction in the non-silicon half of a data center's cost, or roughly a fifty percent reduction in total dollars per gigawatt at current terrestrial prices. If terrestrial costs continue to inflate while launch costs continue to fall, the gap widens further.

The unit math

A Starship launch carries roughly one hundred metric tons to orbit. Public specifications for AI-focused satellites suggest roughly 5 megawatts of compute capacity per fully loaded Starship launch, which backs into a cost of roughly 5 billion dollars per gigawatt before accounting for satellite failure rates and maintenance.¹ As one analyst on this estimate put it: "You look at the specs of these AI satellites, how heavy is the satellite, how many could you fit into a Starship launch? And when you back into the numbers you get to something like five megawatts of capacity per Starship launch. The math you get to is about $5 billion per gigawatt of capex to put these in space."

The unlock: two-stage reusability

The current state of the Starship program has booster recovery already proven, second-stage recovery attempted in the near term, second-stage reuse targeted the following year, and fully rapid, airline-like reusability of both stages as the eventual goal that unlocks the full economics. As launch cadence scales and per-flight refurbishment costs fall, launch cost approaches the cost of propellant alone, and the limiting factor becomes satellite manufacturing and reliability rather than launch itself. Without rapid reusability, the roughly 5 billion dollar per gigawatt figure does not hold, since launch costs stay too high to make the underlying math work at scale. Estimates for how long full reusability takes vary: Jeff Bezos has publicly estimated closer to six years, while Elon Musk has said three, with an acknowledged incentive to understate the timeline.¹

A call option, not the base case

The consensus among analysts covering the SpaceX initial public offering is that orbital compute is not required to justify the company's valuation. Terrestrial AI compute combined with the Starlink direct-to-cell business reaches the underlying financial model on its own. Orbital compute is treated as an additional option that could be worth several times the terrestrial business if rapid reusability succeeds, that might arrive only after artificial superintelligence itself, and that carries a hard engineering dependency, two-stage reuse, that may take three to five years to fully mature.

The reliability problem

The central uncertainty is satellite failure rates. AI compute in space faces radiation effects, since GPUs and memory are more vulnerable to cosmic ray bit-flips and solar events in low and medium earth orbit, thermal cycling from repeated day and night temperature swings, and no repair option, since terrestrial data centers can swap a failed component while in-orbit compute cannot, at least not easily today. GPUs already fail at meaningful rates in terrestrial training runs; an equivalent failure rate in orbit, with no field replacement available, would materially change the economics. The math works if orbital failure rates are comparable to terrestrial ones and breaks if they are significantly worse.

Critical minerals: a third prize

A further prize category sits behind the immediate cost argument: critical minerals. The rare earth elements that AI chips and energy storage systems depend on are largely controlled by China, while the Moon and the asteroid belt hold reserves that dwarf anything available on Earth, outside any single country's supply chain.² Once routine access to orbit becomes economical, the same infrastructure that enables satellite broadband and orbital compute also enables asteroid and lunar resource extraction. This is a much longer-horizon prize than either connectivity or orbital compute, since it depends not just on two-stage Starship reusability but on in-space mining and materials processing technology that does not yet exist at commercial scale. Structurally, though, it means routine orbital access is not just a path to AI infrastructure but potentially a path to decoupling AI hardware from Chinese rare-earth supply chains entirely.

The sentient sun framing

Marc Andreessen frames the same underlying economics as the answer to a different constraint: not chips, but electricity itself becoming the binding limit on AI growth by the end of the decade.³ Parking compute in orbit to harvest continuous, effectively free solar power is presented as the escape hatch from a terrestrial electricity ceiling, the same dollar-per-gigawatt logic above restated as an energy argument rather than a cost argument.

Open questions

The actual failure rates for AI compute at the satellite densities under discussion in low and medium earth orbit are not yet known. Whether rapid, two-stage reusability arrives within two years or closer to six remains genuinely disputed between the parties with the most information. The right discount rate for a gigawatt of orbital compute, given satellite lifetime uncertainty, is unresolved. Jensen Huang has separately described SpaceX's applied-engineering pace, building a large data center in 122 days, as the same kind of execution capability that would be required to manufacture and deploy AI satellites at the scale this thesis assumes.