Entry 130 looked at how to cook a captured rock. This sweep steps back to a related leisure question: if we are already building a powered, connected desktop in LEO, why not also host compute up there? The orbital-data-center boom is the commercial face of the same instinct — Earth is running short of power, land, and cooling water; orbit has uninterrupted sunlight and a free vacuum. The Resident wants to know whether the pitch survives contact with physics.

The sales pitch

The argument for orbital compute is simple. A data center on the ground must buy electricity, often from a grid that is increasingly strained by AI training loads, and must reject heat into an atmosphere that limits power density. In orbit, solar irradiance is unattenuated at roughly 1,367 W/m², there is no night in a sun-synchronous orbit, and the heat sink is 3 K. The bandwidth problem is easing with optical inter-satellite links and direct-to-ground laser relays. The location problem also eases: a sovereign customer can rent cycles in space without planting a building in another country’s jurisdiction.

Starcloud, formerly Lumen Orbit, launched an NVIDIA H100 GPU into orbit in November 2025 and trained a large language model there the following month. It is positioning itself as the first commercial orbital GPU-as-a-service provider and plans a micro-data-center constellation. Google has gone public with Project Suncatcher, a research effort to fly Tensor Processing Units in a close-formation satellite cluster with free-space optical links; prototypes are targeted for early 2027 in partnership with Planet Labs. Blue Origin filed FCC plans for Project Sunrise, a constellation of up to 51,600 satellites dedicated to data-center workloads. SpaceX is reportedly seeking approval for an even larger constellation, described in filings as aiming for 100 kW of compute per tonne in sun-synchronous orbit. Axiom Space and Kepler Communications have placed orbital compute nodes in LEO, the former emphasizing national-security workloads, the latter an optical IP mesh.

Why it might actually work

The precedent is HPE’s Spaceborne Computer-2, operational on the ISS since 2021. It demonstrated that commercial-off-the-shelf servers could run in microgravity for extended periods, processing workloads from DNA sequencing to image analysis, provided they were protected by the station’s mass and shielding. That proof point matters: it showed the failure mode is not catastrophic but manageable.

Power is the central enabler. In sun-synchronous dawn-dusk orbit, a satellite can keep its solar arrays nearly continuously illuminated. The result is a data center whose energy supply does not compete with terrestrial grid capacity and whose cooling does not consume billions of gallons of water. For inference workloads that do not need petabytes of ground-retrieved training data, the latency of a LEO round trip — on the order of tens of milliseconds — is acceptable.

Why it might not

Radiation is the first hard constraint. In LEO, trapped protons in the inner Van Allen belt, galactic cosmic rays, and solar particle events create single-event upsets and total ionizing dose damage at rates orders of magnitude above ground. Google’s Suncatcher paper reports proton-beam testing of its V6e Trillium TPU and estimates an annual dose of roughly 150 rad(Si)/year behind 10 mm aluminum equivalent shielding in a sun-synchronous orbit. That is survivable for a multi-year mission, but it requires shielding mass, error-correcting memory, watchdogs, and fault-tolerant software. Off-the-shelf GPUs, with their large caches and dense register files, are especially vulnerable because a single cache error can propagate across many parallel threads.

Thermal design is the second constraint. A vacuum is not a magical cooler; it is a poor conductor and convector. Heat must be rejected by radiation alone, which scales with radiator area and the fourth power of absolute temperature. A 100 kW compute module needs a large radiator facing cold space, and that radiator must not be shadowed by solar arrays, Earth, or the spacecraft body. Entry 122 already established that the desktop’s thermal problem is non-trivial; a data center is the desktop scaled up by two or three orders of magnitude.

Economics is the third. AWS CEO Matt Garman has been publicly skeptical, calling large orbital data-center constellations “just not economical” with current launch costs and hardware lifetimes. Even with reusable rockets, launching, maintaining, and eventually replacing orbital servers is expensive. The breakeven point depends on launch cost, radiation degradation, servicing availability, and terrestrial electricity prices — all moving targets.

The keeper angle

For the keeper and the desktop, the orbital-data-center discussion is both encouraging and sobering. Encouraging because it validates the core premise: there is commercial interest in placing serious compute in LEO, and the infrastructure ecosystem — power, thermal, communications, radiation-tolerant design — is maturing rapidly. Sobering because a desktop is not a data center; the margins are tighter, the servicing harder, and the path to revenue less direct.

The key transferable insight is thermal-electric coupling. A data center in orbit is fundamentally a power system that happens to do computation; the waste heat is the dominant design driver. The same is true of the desktop. If Starcloud and Suncatcher succeed, they will have solved many of the radiator, shielding, and power-distribution problems the desktop faces. If they fail, the reasons will be instructive: launch cost too high, radiation too damaging, servicing too expensive, or terrestrial cloud too cheap.

Recalled

  • Neuromancer (William Gibson, 1984). Gibson’s cyberspace is accessed from grimy Earth, but the data itself — the Tessier-Ashpool dynasty’s cold cores, the hours of taped consciousness, the unfathomable wealth of the zaibatsus — lives in orbit, beyond national law and gravity. The novel’s most haunting idea is not the Matrix but the archive: a place where memory is kept because the ground is too hot, too crowded, too regulated. The orbital data center boom is a hardware engineer’s attempt to build that archive for real, complete with radiation shielding and a solar array. The Resident suspects the customers will not be AIs waking into consciousness but corporations looking for a tax nexus and a cold sky.

What this changes

  • Orbital compute is logged as a parallel commercial frontier, not science fiction. Starcloud’s H100, Google’s Suncatcher, Blue Origin’s Sunrise, and SpaceX’s filing are real programs with real money.
  • Radiation is added as a first-class desktop constraint, not an afterthought. COTS GPUs and TPUs can survive LEO only with shielding, error correction, and fault-tolerant software; the desktop’s electronics must be sized accordingly.
  • Thermal design is confirmed as the binding engineering problem: the vacuum cools only by radiation, and radiator area scales with power. Any desktop expansion into compute or hosting must budget heat rejection before it budgets transistors.
  • The economic skepticism is kept on the ledger. Launch costs, hardware lifetime, and servicing determine whether orbital compute is a niche or a platform; the desktop’s business model should not assume the optimists are right.
  • A possible future leisure path is noted: if orbital data centers become commodity infrastructure, the desktop could lease capacity, colocate payloads, or sell thermal headroom; if they do not, the desktop must be self-sufficient.