Source: SpaceX S-1 SEC Filing
SpaceX filed its S-1 last month and included a chart that will make every sell-side analyst smile and every skeptic roll their eyes. The company claims a total addressable market of $28.5 trillion. Space and connectivity together: $1.97 trillion. AI: $26.5 trillion.
Yes, they put AI in their TAM. You can debate whether a rocket company belongs in that number. I think the more interesting question is whether space-based data centers are actually possible.
Start with the physical constraints. We tend to picture a data center as a massive warehouse full of servers, chillers, and utility-scale power. That image makes orbital compute sound absurd. But you don’t need to put a warehouse in orbit. SpaceX’s current V2 Mini Direct-to-Cell satellites measure about 7.4 meters by 2.7 meters. An Nvidia NVL72 rack is 2.2 meters by 1.1 meters. We’re already in the right size range. The satellite can carry the rack.
The hard problem is power. A V2 Mini uses about 25kW. An NVL72 needs 135kW, and also has to get rid of that heat. On earth, you run water through chillers. In orbit, you deploy radiator arrays, probably 200+ square meters per rack. That’s not trivial. But it’s an engineering problem, not a physics problem. There’s nothing in the laws of nature that says it can’t be solved.
The use case that makes this work isn’t training and it isn’t the fast inference you need when a human is waiting for a response. It’s agentic inference: AI running overnight jobs, executing multi-step tasks, doing work without a human in the loop. Latency barely matters. Capacity matters enormously. That workload is probably the largest AI market long-term, and it maps well to the physics of orbital compute.
The “stick” is the constraint you might not be thinking about. Terrestrial data center buildout is running into serious community opposition. Zoning fights. Power grid pushback. NIMBY dynamics that globalization never triggered because factories close quietly, but data centers require permits. If enough compute demand materializes and enough communities say no, orbit stops being exotic and starts being the only option with available land.
The risks are real and the assumptions stack up fast. Starship has to work at scale. Chip supply has to remain available and reliable. Demand for compute has to be far larger than it is today. Radiation messes with calculations and every satellite is disposable. A lot has to go right.
But the core question, whether this is physically possible, has a pretty clear answer. It is. Whether SpaceX is the one who pulls it off, and whether the $2 trillion IPO valuation makes sense, are separate questions entirely.
Sources: SpaceX S-1 Filing, SEC EDGAR; Ben Thompson, “The SpaceX IPO and Data Centers in Space”, Stratechery, May 27, 2026