Mentatcurated
Space high · independent

The data center that left Earth

Starcloud put a single Nvidia H100 in orbit, trained a small language model on it in space, and is now raising money to fly 88,000 more.

"Data centers in space" has been a pitch deck for years. Starcloud is the first to actually fly the hardware: in November 2025 it launched a satellite carrying one of Nvidia's top AI chips, then in orbit ran a chatbot and trained a small model from scratch on the complete works of Shakespeare. The training worked well enough that the model started answering in Shakespearean English.

The first thing ever trained on a GPU in space was a model that learned to speak in Shakespearean English.

On the strength of that, the company is now raising at least $200 million at a roughly $2.2 billion valuation — double what it was worth a month earlier, when its previous round had barely closed. The plan it is funding: a constellation of up to 88,000 satellites, each a small orbital data center, which the US regulator agreed in March to consider. The case is that compute in orbit gets unlimited free sunlight and dumps its heat into the cold of space, while data centers on Earth are stuck waiting years for land, water, and a grid connection.

The catch is arithmetic. Eighty-eight thousand satellites of the size Starcloud plans is on the order of a quarter-million tons of hardware to lift — feasible only if SpaceX's Starship reaches its full payload and launches several times a day, a cadence that has been promised but never flown. And skeptics at a rival space firm calculate that orbital compute still costs about three times more per watt than the terrestrial kind. The signal isn't that the constellation is inevitable; it's that the race is now real enough that SpaceX itself has quietly filed to put up to a million of its own data-center satellites in orbit.

The lenses

Novelty 4
Impact · breadth 3
Impact · depth 3
Actionable 1
Substance 3
Hype 4

The facts

The proofFirst Nvidia H100 flown to orbit (Nov 2025); first to run and train a language model on a GPU in space
The raise~$200M sought at ~$2.2B — double the valuation of a round that closed a month earlier; not yet closed
The planFCC agreed in March 2026 to consider up to 88,000 orbital data-center satellites
The catchHostage to Starship launch cadence never yet demonstrated; rivals peg orbital compute at ~3x the cost per watt
Open spacenews.com →

How this connects

Tap a node to open it