The Orbit Question: What Space Data Centers Tell Us About Terrestrial Limits

In November 2025, a 60-kilogram satellite carrying an NVIDIA H100 GPU trained a small AI model on Shakespeare while orbiting at 350 kilometers.

The company behind it, Starcloud, has since raised $170 million at a $1.1 billion valuation and filed for an 88,000-satellite constellation, envisioning a five-gigawatt orbital cluster powered by a solar array spanning four square kilometers. Google announced Project Suncatcher — TPU-equipped satellites linked by free-space optics, with two prototype launches targeted for early 2027 and a modeled 81-satellite formation. SpaceX has sought regulatory permission for up to one million compute satellites. Jeff Bezos predicts orbital gigawatt data centers will beat terrestrial costs within 10–20 years.

We take this work seriously — the radiation testing is real, the optical-link demonstrations (1.6 terabits per second in the lab) are real, and the physics rationale is sound: in the right orbit, a solar panel is up to eight times more productive than on Earth.

But we read this evidence differently than the headlines do.

What orbit solves, and what it doesn't

Space data centers are an answer to exactly two constraints: land and power. That the industry's most sophisticated players are willing to contemplate launch costs, radiation hardening, and orbital debris to escape those constraints is the strongest testimony available that terrestrial scaling has hit something real.

What orbit does not solve: repairability (failed hardware cannot be fixed — Starcloud-1's expected operational life is 11 months before it deorbits and burns up), upgradability (frozen technology in a field where chips obsolete in 2–3 years; Google's own paper concedes the answer is launching redundant spares), latency and bandwidth (the realistic model has satellites computing locally and transmitting refined outputs — which is, we note, the local-AI architecture, relocated to the least convenient location imaginable), and environmental accounting (independent experts warn launch emissions and reentry effects could exceed terrestrial impacts).

Our read

The orbital compute race is best understood as a price signal. When rational actors propose lifting silicon into space to escape the cost of land, power, cooling, and water on Earth, they are telling you what those inputs are really worth. There are two demand-side responses to that price: build somewhere stranger, or need less centralized compute in the first place.

The second response is available now, at zero launch cost. Every interactive AI task that runs on a device the user already owns removes load from the same constrained system that orbit is trying to escape — with full repairability, annual upgrade cycles (consumers replace devices anyway), zero marginal land, and zero water. Orbital compute and local compute are not opponents; they are both symptoms of the same diagnosis. We are working on the symptom that ships this decade.