The Pitch, Taken Seriously

For most of its life, "data centres in space" was a thought experiment — the kind of thing you sketch on a whiteboard and then quietly erase once someone mentions launch costs. Between roughly 2023 and 2026 it stopped being a sketch. Startups raised real money, a space agency commissioned a feasibility study, a hyperscaler published an architecture, and a satellite operator filed for a constellation. In November 2025 a company called Starcloud put an NVIDIA H100-class GPU into low Earth orbit and said it had trained a language model on it. The idea now has hardware, term sheets, and regulatory filings behind it, which means it deserves to be argued with on its own terms rather than dismissed.

Take the argument seriously, because the people making it are not naive about the difficulty. Starcloud, formerly Lumen Orbit, is the most aggressive: $200 million raised (a $170M Series A led by Benchmark and EQT at a $1.1B valuation), an orbital GPU demonstration flown, and a public roadmap toward multi-gigawatt orbital data centres. Google's Project Suncatcher is the most credible-by-association — a "moonshot" design for constellations of TPU-equipped satellites in a dawn–dusk, sun-synchronous orbit where sunlight is near-continuous. Thales Alenia Space's ASCEND study, funded through the European Commission's Horizon Europe programme, concluded in June 2024 that gigawatt-scale space data centres could eventually be viable, with a staged roadmap through a 10 MW operational facility around 2035. Around them sit Cowboy Space (formerly Aetherflux), whose FCC filings envision tens of thousands of satellites for data-centre service; Lonestar Data Holdings, pursuing lunar data vaults; and network plays from Axiom Space and Blue Origin.

Proposer	What they've done / propose	Scale claimed	The advantage cited
Starcloud (ex-Lumen Orbit)	Raised $200M ($1.1B valuation); flew an H100 GPU in orbit, Nov 2025	40 MW reference design; multi-GW aspiration	Continuous solar; ~95% lower power cost (their claim)
Google — Project Suncatcher	Published a space-based AI infrastructure design	TPU-satellite constellation, dawn–dusk orbit	Near-continuous sunlight; optical inter-satellite mesh
Thales Alenia Space — ASCEND (ESA)	Horizon Europe feasibility study, concluded Jun 2024	10 MW operational ~2035; GW-class long term	Lower-carbon sovereign European compute
Cowboy Space (ex-Aetherflux)	FCC filing for a large data-centre constellation	Tens of thousands of satellites	Assumes Starship-class launch economics
Lonestar Data Holdings	Lunar data-vault demonstrations	Data storage, not compute	Sovereignty, resilience, off-Earth backup

Strip away the branding and the pitch has four pillars. Power: in a sun-synchronous orbit the panels never see night, so the same solar array delivers far more energy than it would on the ground, without the storage overhead that dominates terrestrial renewables. Cooling: space is cold and there is no water bill, so — the argument goes — you can pack servers denser and reject heat for free. Real estate and permitting: no land to buy, no community to placate, no water rights, and none of the multi-year grid-interconnection queue that has become the binding constraint on terrestrial AI build-out. Latency and sovereignty: a constellation can put compute above any jurisdiction and, with laser links, move data around the globe at the speed of light in vacuum. Two of those four — the queue and sovereignty — are genuinely real, and we will grant them fully. The other two are where the physics disagrees.

“Orbital data centres could produce at least ten times lower carbon emissions than terrestrial data centres, even when launch emissions are included, with power expenses potentially reduced by as much as 95 percent.”Philip Johnston, co-founder of Starcloud (as reported by IBM, 2024)

Google Research

That is the case, made as strongly as we can make it for them. It hinges on assumptions — launch price, solar-array mass, satellite lifetime, uptime — and the whole edifice is only as sound as those inputs. So the rest of this Frontier does the thing the pitch decks tend to skip: it takes the proponents' own 40 MW reference design and prices out the physics one system at a time. The heat you cannot convect away. The mass you have to lift. The power and the eclipse tax. The hardware that dies with no one to replace it. How you would even keep it connected. And then the whole bill, against a data centre built on the ground. We wanted the pitch to survive contact with the numbers. It does not.