Datacenters in Space

I was driving south on the 101 from San Francisco on Friday, listening to Bloomberg Businessweek Daily talk through the SpaceX IPO, and I got to thinking about datacenters in space. I'd wondered about the concept before — it's genuinely compelling. But hearing that a meaningful piece of a $2 trillion valuation depends on it working in the near term made me want to think it through in detail and specificity rather than generalities.

Let me be clear up front: I want orbital datacenters to work. The concept is elegant — continuous solar power, no real estate or water constraints, and a thermal sink colder than anything on Earth. If you squint, it looks inevitable. But elegant concepts and engineering reality don't always converge on the same timeline. Here's where I landed after working through the constraints.

The Problem Nobody Draws on the Whiteboard

Every rendering of an orbital datacenter shows the same thing: solar panels gleaming in sunlight, servers humming away, clean and simple. What the renderings never show is where the waste heat goes.

A modern AI facility generates enormous heat — call it 100 megawatts for a modest one, at a time when gigawatt-scale campuses are under construction on Earth. That heat has to leave the system. On Earth, air and water do most of the work: convection moves energy cheaply and continuously. In vacuum there is no medium, so heat can only leave by radiation — photon emission from a surface.

The good news: the destination is excellent. Deep space sits near 3 Kelvin, about as good a thermal sink as physics allows.

The problem is the plumbing between the chip and that sink. Without convection, every watt must travel by solid conduction or pumped liquid loops to a radiator surface — systems that are mechanically complex, failure-prone in vacuum, and hard to service. On Earth you can surge-cool by moving more air. In orbit, your thermal ceiling is fixed by radiator area.

And the area is large. At realistic operating temperatures, once you account for environmental heat loads, you need roughly 500 to 1,000 square meters of radiator per megawatt rejected. For a 100-megawatt-class facility, the radiator field alone approaches the scale of the solar array that powers it. You are, in effect, building the structure twice — and radiator panels are exactly the kind of large, thin surfaces that micrometeorites and orbital debris like to find.

Two Suns

Now add the power side. Sunlight at Earth's orbit delivers 1,361 watts per square meter; the best space-grade photovoltaics convert about 30% of that. To feed a 100-megawatt IT load — call it 140 megawatts once you include power conversion, pumps, and communications — you need roughly 340,000 square meters of solar panel. That's about 85 acres, more than 60 football fields, continuously pointed at the sun.

Here's the geometry problem: if the panels face the sun, where do the radiators face?

Radiators want cold, empty sky. But from a 550-kilometer orbit, Earth's disk spans about 134 degrees — it fills roughly a third of the entire sky — and Earth is not cold. It's a 255 Kelvin warm background that also reflects about 30% of incoming sunlight back up at you. A radiator that sees Earth is rejecting heat against a warm wall. So the structure ends up with three simultaneously constrained orientations: panels sunward, radiators edge-on to both the sun and the Earth, compute in between — maintained continuously, on a structure hundreds of meters on a side, as it orbits.

Proponents have a real answer here: dawn-dusk sun-synchronous orbits, riding the terminator line between day and night. These provide continuous sunlight (no batteries for eclipse), and a stable sun angle that simplifies pointing. Those are genuine mitigations, and they're the right design choice. But they don't shrink the radiator field, they don't remove Earth from a third of the sky, and terminator orbits are finite real estate — at the scale being proposed, they become the most contested lanes in low Earth orbit.

Then there's what the rest of us see from the ground. The ISS, with roughly 2,500 square meters of solar array, already flares to about magnitude −4 on a favorable pass — rivaling Venus. A single 85-acre array has more than a hundred times that reflective area; back-of-envelope, that's around five magnitudes brighter, plausibly visible in daylight. Build these at commercial scale — dozens or hundreds of facilities — and you have materially changed what the sky looks like for every person on Earth. The “two suns” framing is only partly a joke.

The Silicon Problem

The third constraint gets the least attention: the compute itself.

Electronics in orbit are bombarded by radiation — high-energy particles that flip bits, corrupt memory, and degrade transistors. The traditional answer is radiation-hardened silicon, and it carries a steep penalty: rad-hard chips typically run two to four process nodes behind the commercial state of the art, because hardening techniques grow increasingly incompatible with cutting-edge transistor geometries. That's not just a performance hit — it's a density hit, and the economics of orbital compute live and die on compute-per-kilogram launched.

The modern counterargument deserves to be taken seriously: skip hardening entirely. Fly commercial chips with error correction, checkpointing, redundancy, and shielding, and accept higher failure rates the way terrestrial clusters already manage GPU attrition. This isn't hypothetical — in late 2025, the Nvidia-backed startup Starcloud flew an off-the-shelf H100 in a small satellite and ran genuine training and inference workloads on it. That demo is real progress.

What it doesn't yet answer are the two questions that matter at scale. First, silent data corruption: a radiation-induced bit flip in a weeks-long training run doesn't announce itself — you may not know your model is corrupted until it misbehaves in deployment, and the detect-and-checkpoint overhead needed to rule that out at frontier scale is an unsolved cost. Second, fleet economics: commercial silicon that degrades faster in orbit means a permanent replacement cadence — a launch tax on every FLOP, forever. One short demonstration in the relatively benign environment of low Earth orbit is a promising data point. It is not yet a production record.

What Would Need to Be True

None of this makes orbital datacenters impossible. It makes them dependent on progress across several independent fronts at once:

Silicon that works up there at modern density. Either process-compatible hardening at advanced nodes, or proven fleet-scale reliability for commercial chips with acceptable silent-error rates over multi-year missions. Today's evidence is one promising demo, not a production record.

Thermal systems at megawatt scale. Deployable radiator fields that are large enough, reliable enough, and damage-tolerant enough — a materials and mechanical engineering problem, solvable in principle, undemonstrated at this scale.

Attitude control for very large structures. Holding three constrained orientations continuously on something hundreds of meters across.

A regulatory framework for orbital traffic. A million satellites in low Earth orbit creates conjunction risk our current institutions aren't built to manage. Kessler syndrome — debris collisions cascading until orbital shells become unusable for decades — is the tail risk that keeps serious people up at night, and it has no market solution.

My honest estimate: meaningful commercial orbital compute is unlikely before 2040, and that assumes real breakthroughs on the first two constraints. The likelier near-term version is edge inference attached to communications constellations — compute riding on infrastructure that's going up anyway, handling latency-tolerant tasks. That's interesting and real. It's also a different category from frontier AI training in orbit.

These aren't exotic objections — they're well-understood constraints in thermodynamics, semiconductor physics, and orbital mechanics. The only question is how fast they can be overcome. If you're working on any of these problems — thermal management, radiation-tolerant compute, large-structure attitude control — I'd genuinely like to hear where the state of the art actually is.