Data centers in space: physics, hype, and the $39B question

It's easy to dismiss "data centers in space" as a pitch deck. Harder to dismiss the fact that an NVIDIA H100 has already trained a language model in orbit. So let's take it seriously, and then ask the one question the renderings never answer: where does the heat go?

This is no longer science fiction

On 2 November 2025, Starcloud put the first H100 GPU in space aboard its 60 kg Starcloud-1 satellite; by December it had trained the first LLM in orbit (NanoGPT, on the works of Shakespeare) and was running Google's Gemma model up there too (NVIDIA, CNBC). The company raised a $170M Series A in March 2026 at a $1.1B valuation, and plans Starcloud-2 for October 2026 carrying H100s alongside Blackwell hardware, even hosting the first AWS Outposts in space (TechCrunch, GeekWire).

Google is in too. Project Suncatcher pairs custom Trillium TPU v6e chips with Planet Labs' satellites; Google says it radiation-tested the TPU to survive a five-year low-Earth-orbit mission and demonstrated 1.6 Tbps on a single optical transceiver pair, envisioning kilometre-scale arrays of 81-satellite clusters (Data Center Dynamics). Analysts count eight organisations filing plans or launching hardware within 90 days, and project the orbital data-center market growing from ~$1.77B (2029) toward $39B by 2035 (Introl).

The pitch is genuinely seductive

In orbit you get near-constant, unfiltered sunlight, solar power without night or weather. There's no land to buy, no water to evaporate, no neighbours to object to a substation. For a world where terrestrial AI build-out is increasingly blocked by power and permits, beaming compute above the atmosphere sounds like an escape hatch.

Then you hit the second law of thermodynamics

Here's the catch the artist's renderings leave out: a vacuum is one of the best insulators in the universe. On Earth, you dump waste heat into air and water by convection. In space there is no air, the only way to shed heat is to radiate it, slowly, through enormous panels. A 120 kW Blackwell rack is a furnace; cooling a multi-megawatt cluster in orbit means radiator structures that may dwarf the solar arrays. Add radiation hardening, launch mass, and the small problem that you cannot send a technician to reseat a cable, and the "free cooling" line quietly inverts.

Why is anyone trying, then?

Because the terrestrial constraints are real. The reason serious companies are launching GPUs is that on the ground, power, land and cooling have become the gating factors for AI, not chips alone. Space is an extreme bet that those constraints are easier to beat 500 km up than next to a substation. For specific edge cases, processing satellite imagery where the data already is, it may genuinely win. As a replacement for the hyperscale build-out this decade, the heat and economics make it a long, long shot.