Orbital Data Centers; Musk Promises AI in Space Within Three Years, But the Numbers Don't Add Up

Elon Musk made it clear at the World Economic Forum in Davos, in January:
"The cheapest place to put AI will be space, and that will be true within two years, maybe three at most."
A few weeks later, SpaceX filed a request with the FCC for an orbital constellation of up to 1 million satellites dedicated to AI computation, positioned between 500 and 2,000 kilometers in altitude.
Three days before the stock market listing, Musk also showcased the first specifications of the AI1 satellite, the first-generation orbital data center. The problem is that Musk's timelines rarely hold up to scrutiny.
Complete self-driving by 2017, the first human crew on Mars in 2024, ten thousand humanoid Optimus robots by the end of 2025: promises systematically unmet. For orbital data centers, the numbers also don't add up even in the medium term.
As highlighted by an IEEE Spectrum report, there are currently about 14,500 active satellites orbiting Earth, two-thirds of which already belong to the Starlink constellation. Launching a million satellites would require production capacities and launch cadences that are off-scale compared to what exists today.

The Real Obstacle for AI Data Centers is Heat, Not Space

In the history of astronautics, there have been about 7,000 orbital launches in total: the Starship alone, which can carry up to 60 satellites, would need to conduct 16,666 missions dedicated solely to this purpose. SpaceX ended 2025 with a record of 165 orbital launches: even if that pace were to be multiplied tenfold, it would take a decade just to place the hardware in orbit.
At Starlink's current production rate, about 4,000 satellites per year, and assuming a tenfold increase in industrial capacity, building a million satellites would still take about 25 years.
Complicating matters is thermodynamics. Starcloud, a startup that has already applied to the FCC for a constellation of 88,000 satellites, has sent only one Nvidia H100 GPU into orbit: the onboard radiator proved too weak to operate it at full power.
A single H100 consumes 700 watts and requires 1.4 square meters of radiator to stay at 60°C. A 40-kilowatt rack requires 80 square meters, while a 100-megawatt data center would need 2,500 radiators of that size. There are also those in the astronomical field who fear that a million satellites equipped with enormous radiating panels could compromise sky observation, in addition to increasing the risk of triggering the Kessler syndrome.
According to the source, behind the enthusiasm for AI in space lies mainly economic convenience for Musk, who concurrently controls xAI (which builds the data centers), SpaceX (which launches them), and Tesla (which produces the solar panels).
"It’s almost like he’s paying himself," summarized.
Analysts remain divided on the timeline. Michael Pierce, from Technology Strategy Partners, believes that Starlink's laser link network represents a competitive advantage that is hard to replicate and hypothesizes cost parity with terrestrial data centers within 5-10 years, but only for inference loads: training models require synchronization and low latency that a distributed orbital system can hardly guarantee. Matt Hasan, an independent AI strategy consultant, emphasizes that the AI1 project does not change the underlying logic of linking computation to energy generation, but rather pushes forward the timelines and scale: launch costs, maintenance, hardware replacement, and thermal management remain open issues yet to be resolved.