Relocate compute into orbit, powered by continuous sunlight and cooled by radiators

The proposal

Fly accelerators on satellites in sun-synchronous or dawn-dusk orbit, drawing near-continuous solar power without atmospheric losses, rejecting heat through radiators, and linking satellites by free-space optical fabric. The community argument is absolute: no host community, no municipal water, no grid interconnection queue, no noise.

What has actually been demonstrated

Starcloud-1 launched on 2 November 2025 on a Falcon 9 into a 325 km orbit: a 60 kg satellite on an Astro Digital bus carrying a single Nvidia H100, the first data-centre-class GPU in orbit, with a design life of about 11 months and controlled deorbit (Gunter's Space Page, Data Center Frontier). Company-sourced reporting says it ran Gemma inference and trained a nanoGPT-scale model.

That demonstrates one GPU surviving and operating in low Earth orbit. It does not demonstrate multi-kilowatt thermal rejection, inter-satellite fabric, multi-year radiation survival, or any economics.

Elsewhere: Google's Project Suncatcher plans 81 satellites at about 650 km using Trillium TPUs with two Planet Labs prototypes by early 2027, and nothing has flown (Google). China launched the first 12 satellites of a computing constellation in May 2025 at up to 744 TOPS each with 100 Gbps laser links, holding ITU approval for 2,800 (China-in-Space). Starcloud raised $170 million at a $1.1 billion valuation in March 2026 and describes a long-term 5 GW deployment with a 4 square kilometre solar array (DCD).

The binding physical constraint

In vacuum the only heat path is radiation, so a data centre in orbit is mostly a radiator. Published estimates per megawatt: about 1,600 m² (World Economic Forum), about 1,200 m² (EE Times), and 700 to 1,800 m² (RCR Wireless). Scaling ISS-era thermal hardware implies roughly 100 tonnes of radiator per megawatt against about 10 tonnes of compute. A gigawatt implies 0.7 to 2 km² of radiator, and coating degradation from UV and atomic oxygen increases required area over the mission.

Trade-offs and limitations

• Cost. IEEE Spectrum concludes that even assuming Starship at $44/kg and terrestrial power at $0.20/kWh, launching and running a GPU in orbit for a year costs at least an order of magnitude more than the same work on the ground (IEEE Spectrum). Current launch is about $2,700–$3,000/kg, implying roughly $40 billion in freight alone for a gigawatt.
• The most favourable rigorous analysis still requires heroics. Forethought finds parity at about $100/kg against roughly $1,500/kg today — a fifteen-fold reduction contingent on Starship achieving Falcon-9-class reusability, plus regulatory approval for a thirty-five-fold launch-rate increase — and prices unrepairable hardware at about 9% annual compute loss. It concludes orbital compute is unlikely to be a meaningful share of capacity before 2030 (Forethought).
• The vendor is a sceptic. Nvidia's CEO on the 26 February 2026 earnings call: