Two satellites, not 88,000

The dates are already on paper. 14 August on Musk's 2029 sentence. 21 August on a $250 million raise. 3 February on an 88,000-satellite file. November 2025 on one H100. January, or 2027, on two 8-kilowatt rideshares. Late 2028 on a closed Falcon 9 book.

Two rideshares are the exposed hardware

The object the catalog actually produced is a reservation, not an 88,000-satellite hall. Starcloud-1 launched in November 2025 with an Nvidia H100. About 60 kilograms. One bird. The next object that can change the claim is Starcloud-2: 8 kilowatts, several hundred kilograms, two birds on Falcon 9 rideshares, government inference as the named customer class. SpaceNews, after the 21 August round, put the first of those on a January rideshare. TechCrunch put both in 2027. Either way, the next proof is a radiator and a power budget, not a valuation.

The 88,000-satellite, 20-gigawatt constellation lives in an FCC file dated 3 February 2026. Johnston told SpaceNews the 88,000 depends on Starship flying frequently, that Starlink V3 is the first commercial use of that vehicle, and that customer payloads such as Starcloud sit 18 to 24 months after that, probably 2029 or 2030. That is the same man who told X that 2028 is the mainstream date. Both sentences can be true if mainstream means political permission and 88,000 means a vehicle that does not yet fly paying customers. They cannot both be a 2026 reservation.

Elon Musk, 14 August: orbital compute is the only way to scale AI, probably sometime in 2029, because of power availability and permitting problems on land. Philip Johnston, answering him in the same window: orbital compute needs to be the mainstream by 2028 and not 2029, because public sentiment is moving toward banning even existing data centres on the ground. Then, on 21 August, Johnston posted the raise. Reuters and SpaceNews printed the same numbers the same day. The argument is 2028 versus 2029. The booked hardware is two 8-kilowatt rideshares.

The launch sentence is the one the argument skips. SpaceNews, same week: Falcon 9 rideshare reservations are unavailable beyond late 2028 or early 2029 as SpaceX prepares the Starship transition. TNW's headline was the honest one: Starcloud raised $250 million to buy rockets it cannot yet book. Johnston is opening a 100,000-square-foot plant that can theoretically build 100 satellites a week. A plant that cannot book a fairing is a warehouse.

Johnston has been unusually specific about the hard part. The radiator: around 10 times lower mass per watt of dissipation than the ISS radiator, and around 100 times lower cost per watt. That flies, he said, in eight months. Heat, not FLOPs, is the constraint. A terrestrial H100 is a water problem. An orbital H100 is an area problem. Cryptopolitan, quoting the same week's physics, put a 1-megawatt class at roughly 5,640 square metres of solar and 2,500 square metres of radiator. GPU refresh is one to two years. Satellite life is five to seven. If the chip the plan needs is Nvidia's Space-1 Vera Rubin module, Starcloud itself says that part does not exist yet and hopes to fly it in late 2028. The raise buys radiators, a plant, and launch deposits. It does not buy a chip that has not been manufactured.

SpaceX has told the FCC it wants up to one million compute satellites. Blue Origin's Project Sunrise is about 51,600. Startup Orbital, June 2026, filed for up to 100,000 of 100-kilowatt-class nodes and 10 gigawatts, first purpose-built sat 2028. Google's Suncatcher is two TPU prototypes with Planet by early 2027. Aetherflux rebranded to Cowboy Space Corporation in May, raised about $275 million, and flipped from beaming power at the ground toward in-space compute plus a rocket, first commercial node still Q1 2027 on paper. SpaceX's August earnings language, as SpaceNews reported it, is that Starmind AI satellites next year would be essentially an optimized Vera Rubin NVL72. Overview Energy, April, sold Meta a night-time gigawatt, demo 2028, commercial 2030. Those are opposite logistics claims sharing the noun space.

It is congestion. Citizens for Space-Based Solar Power, 7 August, said the quiet part: efficiency, beam safety, insurance, debris, astronomy. None of those are priced in a $2.3 billion private round.

What a raise cannot do

Cross-asset does not care which bird books a fairing. AI abundance cares only if the middle's hosting cost becomes an orbital cost. It has not, on the invoices those pages already named. Two 8-kilowatt birds are not a token price. The Suncatcher page already named Planet and early 2027. The orbital-compute page already named the raise. This piece is the reservation.

A refutation would look like this: Starcloud-2, or an equivalent 8-kilowatt-class node from anyone, in orbit by the end of 2027, with a named customer running inference, confirms the narrow claim. A 1-megawatt billed cluster serving a named terrestrial AI customer by the end of 2028 would make Musk's 2029 sentence early. Neither has happened. One H100 in a 60-kilogram bird is a demonstration. It is how you get Benchmark and Nvidia into a round. There is no Starcloud ticker.

The honest uncertainty is which object gets a number first. If the two rideshares fly and a named customer runs inference, Starcloud has shown that an 8-kilowatt class can live on a Falcon 9 and talk to the ground. If they slip and the 88,000-satellite file grows, the file moved and the reservation did not. If Suncatcher's two birds fly and Starcloud-2 is still a deposit, the prototype moved and the raise did not. Those three sentences can live in the same year. Two rideshares are a fact about one book. Eighty-eight thousand is a fact about one filing. Neither is a hall that has already trained a model.