In mid-2025, a crypto-native outlet ran a headline with the gravitational pull of a launch vehicle: SpaceX and Nvidia are building a data center in orbit. The payload was perfect. It combined the world's dominant launch operator, the world's dominant AI chip designer, low Earth orbit, and a promise of AI processing beyond atmospheric limits. There was only one missing component: evidence.
No source. No engineering diagram. No contract signature. No launch date. No named customer. As of this writing, neither company has confirmed any partnership. Public reporting is limited to exploratory discussions about using Starlink laser inter-satellite links to connect a future orbital facility. That is a conversation. The headline says 'building.' The distance between those two words is the entire story.
I have spent a career reading smart-contract disclosures and on-chain flows. The first rule is the same in aerospace: verify state transitions before trusting output. In 2017, I found a reentrancy bug in a LendingBot time-lock contract by reading the withdrawal logic before the mainnet launch; the team patched it and avoided a $2 million drain. The habit stuck. When I see a bold headline, I interrogate it like a transaction. This transaction fails basic validation.
The headline is too good to be true. The physics is worse.
Let's establish the baseline before running the numbers. SpaceX controls the only production reusable launchers, Falcon 9 and Starship, and operates Starlink, a constellation of more than 7,000 low-Earth-orbit satellites. Nvidia controls roughly 90 percent of the AI training market plus the CUDA software lock-in. On paper, the combination reads as a vertical stack: transport on the bottom, communication in the middle, compute on top. That is why the story spread. But orbital infrastructure does not care about PowerPoints.
Start with power. A one-ton satellite with deployed solar arrays can generate roughly 10 to 20 kilowatts. After keeping the bus alive, 5 to 10 kilowatts remain for compute. An Nvidia H100 burns 700 watts. That yields seven to fourteen GPUs per satellite. A single ground-based AI server holds eight H100s. The most optimistic orbital data center is therefore equivalent to one or two ground servers. Microsoft and Meta place orders for tens of thousands of GPUs. The gap is four to five orders of magnitude. Launching a few racks into space cannot close that gap.
Heat is worse. Vacuum is a perfect insulator. Convection does not exist in orbit. Waste heat can be rejected only by radiation, and radiated power follows the Stefan-Boltzmann law: it scales with the fourth power of temperature. To dump ten kilowatts of GPU heat, you need oversized radiator panels or a two-phase ammonia loop that transports heat from the silicon to the radiator. Every kilogram of that plumbing is subtracted directly from payload mass. That is not an engineering bug; it is a physical tax.
Bandwidth is the third wall. Starlink's laser links run at roughly 10 Gbps per link. That is enough for navigation, phone calls, and small inference requests. It is irrelevant for distributed training. Ground data-center fabrics such as NVLink and InfiniBand move hundreds of Gbps per port and Terabits per second in aggregate. An orbital distributed-training run would starve within seconds. The only workable workloads are inference, sensor fusion, and real-time satellite image processing: tasks that can run on a small cluster and send only the answer to Earth.
The unit economics are worse. Assume Starship reaches its target of $100 per kilogram to orbit. A one-ton data-center satellite costs $10 million in launch fees alone. If that satellite carries ten H100-class GPUs, launch cost alone is $1 million per GPU. A ground-deployed H100, including server, cooling, power, and facility, costs $30,000 to $50,000. Even with three years of free solar power in orbit, the space GPU has a total cost of ownership at least ten times higher. The phrase 'zero-carbon orbital compute' is also too good to be true. A Falcon 9 launch emits hundreds of tonnes of CO2; a Starship launch emits thousands. The zero-carbon story disappears before the rocket clears the tower.
The industry timeline confirms the gap. Lumen Orbit, a startup founded in 2024, plans to launch a test satellite with GPUs in 2025. The EU's ASCEND project, led by Thales Alenia Space, completed a feasibility study in 2023 and does not expect an economically viable space data center before 2036. This technology is at proof-of-concept stage, not production. Anyone who describes a construction project today is overstating the state machine.
Now the contrarian layer. If the economics are broken, why would two serious companies explore the idea? The answer is not computation. It is jurisdiction. A satellite in LEO is not physically located on any nation's soil. For a European bank restricted from moving citizen data to US cloud infrastructure, or a defense agency that wants to process raw imagery without routing it through ground networks, an orbital node is a compliance product rather than a compute product. 'In orbit' correlates with 'sovereign,' not 'cheap.' The only customers with enough tolerance for 10x costs are governments and defense contractors. The US Space Force has already listed on-orbit processing as a critical capability.
That logic also exposes the power structure hidden in the headline. SpaceX has the stronger hand. There is exactly one way to put a GPU into orbit at scale today: ride on SpaceX. There is no alternative launch provider with comparable mass-to-orbit. Nvidia, by contrast, can be replaced by AMD or custom ASICs in a data center. So this is not a marriage of equals. It is a platform deal with SpaceX as the landlord and Nvidia as a qualified vendor. The media wants a co-branded joint venture. The balance of power says otherwise. Nvidia is likely running parallel conversations with other satellite platforms and national agencies; a single leak from a crypto outlet is a positioning signal, not a definitive contract.
The deepest blind spot is orbital debris and governance. LEO already contains more than 40,000 tracked objects. A data-center satellite is larger, heavier, and hotter than the average payload. If it collides, the resulting debris threatens the entire Starlink constellation. The legal regime has not caught up. A data payload in orbit has no clear national jurisdiction. If a satellite is hacked, whose court gets the case? The Outer Space Treaty was written for flags and rockets, not for GPUs. Anyone expecting compliance clarity in the next three years is confusing a tweet with a treaty.
The fact that this story first appeared on Crypto Briefing is itself a data point. Crypto media moves narrative faster than primary documents. It reports rumor as project and conversation as construction. That does not make the underlying technology worthless. It makes the specific claim unreliable. The right reaction is not to buy a DePIN token; it is to update a spreadsheet.
Takeaway: ignore the headline and track the milestones. I will believe in an orbital data center when I see one of three things: a test satellite launched, a GPU powered on in orbit, or a named customer paying for processed output. If none of those exists within 18 months, this story was not an engineering project; it was a fundraising event. I have seen this exact sequence before. In 2022, 'decentralized sequencing' for Layer2 was also too good to be true. Two years later, it was still a slide deck. The orbital data center is following the same trajectory: a lot of thrust, no telemetry. If a company ever publishes the first GPU temperature reading from LEO, I will pay attention. Until then, treat the press release as a proof-of-stake. The hardware has not shipped.