A million satellites on paper: read the filings, not Musk's 2029 deadline

Elon Musk posted on X on August 14, 2026 that orbital compute will be "the only way to scale" AI "probably sometime in 2029," citing power availability and permitting problems on land ^[1]. Treat the deadline as promotion from a man who sells launch capacity, and treat the paperwork underneath it as the actual signal: regulators are now holding requests for more than a million compute satellites ^[3][5][7][1].

The tally is worth stating plainly. SpaceX has filed with the Federal Communications Commission for an orbital data-center system of up to 1 million satellites ^[3]. Blue Origin filed with the FCC in March for a 51,600-satellite network it calls Project Sunrise, explicitly to move energy- and water-hungry compute off the ground ^[5]. Starcloud has proposed roughly 60,000 ^[7]. That is about 1,111,600 satellites on paper ^[1]. Google is asking for less and has more hardware in motion: Project Suncatcher runs its own TPU chips, with two prototype spacecraft planned for a 2027 launch alongside Planet ^[6]. Starcloud flew the first Nvidia H100 GPU in orbit last November and ran AI workloads on it ^[7].

Filings are cheap, and a request is not a constellation. But four parties with different balance sheets do not queue up for the same escape hatch unless the thing they are escaping is binding. According to the Cryptopolitan account of Musk's argument, AI data centers may take up to 17 percent of US electricity by 2030 ^[8], with new fossil plants under construction to serve that load, older plants worn out, and cooling drawing on already scarce water ^[14]. The pitch for orbit is that sunlight is free to harvest and waste heat can be radiated into vacuum ^[13]. Musk's own framing has moved: he previously treated space as an expansion route rather than a precondition ^[2].

The physics has not moved with it. A 2026 technical study cited in the same report puts a single one-megawatt orbital system at roughly 5,640 square meters of solar panel and 2,500 square meters of radiator, before the spacecraft itself ^[9]. Scale that linearly to a gigawatt-class site and you are deploying about 5.64 square kilometers of panel and 2.5 square kilometers of radiator ^[2]. The same study argues the proposed launch approach would be 3.4 to 13.5 times cheaper than a public Falcon 9 dedicated launch benchmark ^[10], and SpaceX's 2028 activation date rests on Starship hauling 100 metric tons to low-Earth orbit ^[4]. Then there is the depreciation mismatch: AI processors are useful for one to two years, satellites run on a five to seven year cycle ^[11], which means between 2.5 and 7 chip generations pass inside a single satellite's life ^[3] with no truck roll available. Economics, cooling physics and insurability all remain unproven ^[12].

What to watch is narrow and checkable. Whether the FCC treats million-satellite asks as spectrum options or as capacity plans. Whether the Google and Planet prototypes fly in 2027 ^[6]. Whether Starship demonstrates the 100-ton figure the 2028 date depends on ^[4]. And whether any terrestrial operator ever cites orbit as the reason it walked away from a site, which is the only version of this that shows up in a capex line.