In Today’s Issue:

🛰️ What is actually operating in orbit today, against 1.16 million satellites applied for

📡 The bandwidth gap that decides training, and the downlink gap that decides inference

📄 A stock prospectus that argues against itself, and a launch story that drifted

💸 Four different launch prices, and why cheap power is only 7 percent of the bill

⚖️ The export-control gap nobody has written about: an orbit is not a country

🔍 Where each side is weakest, including the sceptics

Dear Readers,

On 24 August, Nvidia and SpaceX said they had designed a space-optimised Vera Rubin NVL72 for SpaceX's first Starmind AI1 satellite, seventy-two GPUs and thirty-six CPUs of the same class that fills frontier data centres. Musk put the first launch in the fourth quarter of 2027. The reaction split immediately into believers and eye-rollers, which is what usually happens when a question gets asked in the wrong shape. Almost nobody asked the more useful one: what is the hardware actually supposed to do up there?

That sounds like a detail. It is the whole argument. Ask the three biggest names in this field and you get three different answers, and two of them come from companies with billions committed.

SpaceX says inference. Its Form S-1/A, the registration statement filed before the largest IPO in history, names inference twice as the orbital workload, once in the business section and once in the glossary (SpaceX Form S-1/A, 06/03/2026). Starcloud says training. The company's whitepaper is titled "Why we should train AI in space." Google says training too, but argues entirely about something else. Project Suncatcher is the largest orbital training research programme in the world, and its case rests almost wholly on the bandwidth between satellites (Agüera y Arcas et al., 11/2025).

And the sceptics reject the premise entirely, at least for customers on the ground. Andrew Cavalier at ABI Research finds training and inference on large language models in space both uneconomic today, and points instead to Earth-observation preprocessing (IEEE Spectrum, 06/11/2026). Bill Ray at Gartner says orbital data centres will not analyse Earth data for Earth applications "for decades, if ever" (The Register, 02/25/2026). Matt Garman, who runs AWS, says the economics do not work and there are not enough rockets (02/2026).

Each of those arguments puts a price on a different cable. A satellite has three of them: the link inside the spacecraft between panels and chips, the link between spacecraft, and the link down to the ground. Each author picks one, prices it, and reaches a verdict about the whole system. Which raises the question this piece is built to answer: is orbit really the future of AI compute, and does the answer hold for training and inference alike?

All the best,

Kim Isenberg

This topic invites category errors, so it is worth being pedantic about four different things: what is operating, what is contracted, what is applied for, and what is on a slide.

Operating, in August 2026, is almost nothing. The entire installed base of dedicated, free-flying orbital compute is roughly forty Nvidia Orin edge processors aboard ten Kepler Communications satellites, two of which carry Axiom Space nodes, plus a single H100 on Starcloud-1, a 60 kg satellite launched in November 2025 (TechCrunch, 04/13/2026; IEEE Spectrum, 11/03/2025). China's Three-Body constellation adds twelve Earth-observation satellites. That is the whole ledger. When Sophia Space tried in April 2026 to bring up an operating system across six GPUs on two spacecraft, it was the first time anyone had attempted in orbit what is a Tuesday afternoon in a terrestrial data centre.

Applied for is a different universe. Filed with the FCC and not granted: SpaceX up to one million satellites, Starcloud 88,000, Blue Origin's Project Sunrise 51,600, Cowboy Space 20,000. That is about 1.16 million spacecraft, against roughly 14,500 active satellites in orbit today. Not one application has been approved. "Accepted for filing" is not approval, and coverage reporting SpaceX as cleared for a million satellites is simply wrong (FCC Public Notice DA 26-113, 02/04/2026).

One useful correction comes from a person with every reason to talk the sector up. Mina Mitry runs what TechCrunch calls the largest compute cluster in orbit and declines the label: Kepler does not see itself as a data centre company, and on the workload question he says the belief is that it is more inference than training (TechCrunch, 04/13/2026).