The Answer to Orbital Compute's Silicon Problem Isn't Rad-Hard
Vincent Pribble
SubscribeSign in
The Answer to Orbital Compute's Silicon Problem Isn't Rad-Hard
Vincent Pribble<br>Jul 08, 2026
Share
Orbital data centers and large low-Earth-orbit constellations run on commercial silicon. Not radiation-hardened parts: the same memory and compute you would find in a terrestrial server or a modern car. That choice is largely already made, and it raises a question the industry has not answered cleanly. How do you qualify commercial silicon for the radiation environment of space when the failure-rate data for these parts mostly does not exist?<br>Last month I argued that this data gap is real and unpriced, that “COTS plus system-level mitigation” often means sizing error budgets against failure rates nobody has measured. That piece described the problem. This one is about the answer.<br>Thanks for reading! Subscribe for free to receive new posts and support my work.
Subscribe
The short version: you do not qualify a million-satellite fleet by beam-testing every die. You test the right dies, model the rest, and accept that the qualification paradigm built for rad-hard parts does not carry over. Working through why is where most of the value is, so let me build it up properly.<br>The commercial-silicon bet is already placed
Since that piece ran, the clearest evidence has come from the newest operator, not the biggest one. Starcloud put an NVIDIA H100 in orbit in November, trained a small language model on it, and ran Google’s Gemma. In February it filed with the FCC for a constellation of up to 88,000 orbital data center satellites, and in March it closed a $170 million Series A. Google’s Project Suncatcher will fly TPUs on two prototype satellites with Planet by early 2027. And the SpaceX filing I opened the last piece on, up to one million satellites between 500 and 2,000 kilometers built by scaling up Starlink V3, was accepted for review by the FCC’s Space Bureau within a week of submission.<br>None of these are rad-hard programs. There is no radiation-hardened H100, and no QML-V version of a Trillium TPU. Starcloud’s own CEO was candid that an H100 is probably not the right chip for space, and that the point was to show state-of-the-art terrestrial silicon can run in orbit at all.<br>One detail from that mission deserves more attention than it got. A second GPU on the same spacecraft, an Nvidia A6000, failed during launch. To be careful about what that is: a launch failure, not a radiation failure, and no root cause has been published. Which is rather the point. Two commercial parts flew, one survived, and the public evidence on how terrestrial silicon holds up across the full mission environment is thin enough that a sample of two counts as a headline.<br>I made the caveat last time and it still holds: committed capital is not engineering validation. Taking that as read, the interesting question is no longer whether the bet is being placed. It is whether there is a defensible way to underwrite it.<br>Why the bet is reasonable, and where it stops
The case for commercial silicon in low Earth orbit does not rest on anyone’s logo. It rests on four conditions, and all four have to hold.<br>The first is a benign environment, which I sketched last time and will now put numbers on. A satellite at 550 kilometers behind a millimeter or so of aluminum accumulates roughly 1 krad per year of total ionizing dose. Over a five-year Starlink-class life, that is under 30 krad. A heavily shielded compute satellite sees far less: Google’s Suncatcher analysis puts its design point near 150 rad per year behind 10 millimeters of aluminum, a few hundred rad across the mission. Geostationary orbit, by contrast, can deliver up to 50 krad per year and over a megarad across a twenty-year life. The hardness budget that makes sense for a fifteen-year GEO satellite is largely unnecessary mass on a five-year LEO satellite you plan to deorbit and replace.<br>One detail in the SpaceX filing cuts against the easy version of this argument, and it needs saying. The proposed shells run from 500 to 2,000 kilometers, at 30 degree and sun-synchronous inclinations. The upper end of that range is not benign LEO. Somewhere above roughly 1,000 kilometers you are climbing into the inner proton belt, where dose rates and proton-driven upset rates rise by orders of magnitude, and sun-synchronous orbits transit the South Atlantic Anomaly several times a day. A constellation spanning that range cannot carry a single qualification posture. The lower shells can fly the commercial bet on the terms this section lays out. The upper shells are a different radiation environment wearing the same filing. That is not a problem for the case I am making. It is the case: qualification rigor has to be priced per shell, not per constellation.<br>The second condition is short mission life. The third is constellation spares: when a node fails, the fleet routes around it and a replacement is already...