Low Earth orbit is not a friendly place for a computer. A satellite in low Earth orbit (LEO) completes a lap around the planet in roughly 90 minutes, and for about a third of that lap it is in darkness. Solar panels produce nothing in shadow, so for half an hour at a stretch the whole platform runs on what the batteries stored during the sunlit pass. Power comes before compute.
It is tempting to size orbital hardware the way you size a terrestrial data centre. The environment gets there first. Eclipse, orbit choice and ionising radiation set the ceiling on what you can run, and they are one system, with orbit choice as the dial that moves the other two. Follow the orbital data centre landscape and the pattern repeats.
Why do LEO satellites lose power during eclipse, and what does that mean for AI workloads?
The mechanism is mundane: Earth gets between the satellite and the Sun, the photovoltaic array drops to zero output, and the batteries carry the entire payload until the next sunlit pass. In a roughly 90-minute LEO orbit that means 30 to 40 minutes of darkness per lap, with altitude and the beta angle fixing the exact duration.
Because the satellite spends a third of each orbit in Earth’s shadow, the power budget is two budgets. In sunlight the array has to run the payload and recharge the batteries. In shadow the batteries carry the full payload alone, so the average duty cycle is bounded by the sunlight fraction minus that recharge margin, not by your theoretical compute.
The batteries feel it. They run roughly 16 charge and discharge cycles a day, about 30,000 over a five-year mission, and deeper discharge per eclipse shortens cycle life. The same swing strains the thermals too, which is why radiative cooling matters as much as generation in an orbital data centre.
For AI workloads the consequence is a scheduling constraint. Bursty, delay-tolerant inference fits this rhythm: run heavy, checkpointed work in sunlight, then coast through eclipse on stored energy. Continuous training wants steady power the orbit never offers.
What is sun-synchronous orbit, and why is it preferred over general LEO?
A sun-synchronous orbit (SSO) is a near-polar orbit whose plane precesses about one degree eastward each day, keeping pace with Earth’s motion around the Sun. The satellite crosses any latitude at the same local solar time on every pass, which is what makes an SSO a specific, engineered orbit rather than a generic polar one.
Getting there is a tuning act. Pair inclination against altitude so Earth’s equatorial bulge nudges the plane at the right rate, and you land near 98 degrees, at altitudes around 600 to 800 kilometres.
The dawn-dusk special case rides the day-night terminator, the line between daylight and darkness. The array stays pointed at the Sun on every pass, erasing the eclipse problem and pushing the platform toward near-continuous solar exposure. Power-hungry payloads prefer it.
Sun-synchronicity is worth having, and it costs you. You are locked into a fixed local-time plane and that near-98-degree inclination, and launch windows become rigid. Orbit choice trades one hazard for another: fly a polar SSO and you pass through regions of much higher proton flux, while the thermal cycling other orbits suffer is what kills reliability over a mission.
How does ionising radiation degrade GPUs and other chips in low Earth orbit?
Radiation attacks silicon twice. Total ionising dose (TID) accumulates over the mission, shifting threshold voltages and raising leakage current until parts drift out of spec. Single-event upsets (SEUs) are the other kind: one energetic particle strikes, flips a bit, or triggers a latch-up that shorts a power device and destroys it.
TID is the slow killer. The estimated orbital dose in a target sun-synchronous LEO is about 150 rad(Si) per year, roughly 750 rad(Si) over five years from protons and galactic cosmic rays. It is a cumulative reliability limit.
SEUs are faster and stranger. A single particle flips a bit, and because advanced nodes store less critical charge per cell, the dense silicon that makes a modern GPU fast is also the silicon with the most sensitive targets. Beam tests of a shielded commercial TPU survived a full five-year-equivalent dose, but its high-bandwidth memory threw uncorrectable errors first.
The mitigation is layered: ECC memory, watchdogs, checkpoint and restart, redundancy, shielding. And because radiation tolerance and hardware reliability are claims you will have to vet, the next question is what silicon to fly.
Rad-hard processors versus commercial off-the-shelf GPUs: which is the better bet?
Rad-hard processors are the proven option. They are space-qualified and tolerant of single-event effects, but they lag years behind state-of-the-art terrestrial silicon and cost far more per unit of compute. Shielded commercial silicon carries a much smaller premium.
Commercial off-the-shelf (COTS) GPUs, Nvidia parts above all, deliver far more performance per watt and per dollar, but they are radiation-vulnerable and dose-limited. Nvidia is already shipping into this market: its Space-1 Vera Rubin module claims up to 25 times the AI compute of an H100 for orbital inference.
The realistic answer is usually the middle ground. Wrap a fragile accelerator in the shielding, redundancy and error correction described above, and it survives. Redundancy is not free: triple-modular redundancy triples the hardware requirement.
The decision is a judgement framework. Weigh mission lifetime, fault tolerance, workload type and servicing access. In-orbit servicing remains underdeveloped, so operators mostly budget spare parts and plan to deorbit and replace. That is the same build versus buy arithmetic you know from the ground, with launch mass as the binding constraint.
So the three forces are one cascade. A 90-minute orbit and its 30 to 40 minutes of darkness force a two-phase power budget and a bursty, schedulable workload. Choose a dawn-dusk sun-synchronous orbit and the eclipse nearly disappears, but you have also fixed your radiation and thermal exposure, because the same dial moves all three.
Radiation is the deeper limit. The fast, dense commercial silicon you want is exactly the silicon most likely to flip a bit or latch up, so the rad-hard-versus-COTS choice is settled by mission length, fault tolerance and servicing access rather than by any single chip.
Before you ask what hardware to fly, ask what orbit it will fly in, because the orbit answers the hardware question for you. The physics and cost threads bind the whole thing together.
Frequently Asked Questions
Can you just add bigger batteries to solve the eclipse problem?
Bigger batteries help, but they do not remove the constraint. Extra storage adds launch mass, and the solar array still has to recharge that larger battery inside the sunlit window, so deeper discharge simply shortens cycle life. The smarter fix is orbit choice and workload scheduling, not brute-force storage that the launch vehicle has to carry.
Why can’t every satellite fly in a dawn-dusk orbit?
A dawn-dusk orbit only works when the orbital plane stays aligned with the day/night terminator, and that geometry is fixed relative to the Sun. Missions needing specific overpass times, different coverage patterns or lower inclinations cannot use it. It is a special case rather than a default, and most payloads trade it away for their own operational needs.
How do the thermal swings between sunlight and shadow affect hardware?
Every eclipse drags the payload through a rapid temperature swing, and repeated cycling stresses solder joints, connectors and circuit boards until they crack or drift out of spec. Thermal design has to hold components within limits across both extremes, which means heaters, radiators and careful material choices. A dawn-dusk orbit largely removes that cycling, one reason power-hungry payloads prefer it.
Can software fixes recover a chip after a radiation hit?
Partly. Software can detect and recover from many single-event upsets using error-correcting memory, watchdogs and checkpoint/restart, so a flipped bit or a corrupted calculation does not end the mission. But software cannot undo destructive latch-up or the slow accumulation of total ionising dose, because those are physical changes to the silicon that no patch can reverse.
How much does radiation shielding actually help, and what does it cost in mass?
Shielding reduces dose, but with sharply diminishing returns, because energetic particles generate secondary radiation inside the shield itself. Every kilogram of shielding is a kilogram not spent on compute, and launch mass is the binding constraint. That is why operators lean on error correction and redundancy as much as on physical shielding.
Can a satellite’s compute hardware be repaired or upgraded once it is in orbit?
Historically no, which is exactly why hardware is designed to degrade gracefully rather than fail outright. In-orbit servicing is changing this, with robotic servicing and replacement missions now able to extend or swap payloads. For the moment, though, plan for the chip you fly to last the whole mission, and treat servicing as a welcome extension rather than a guarantee.
How can I check whether a vendor’s radiation-tolerance claims are credible?
Ask for the mission profile the numbers were measured against, not just a headline dose figure. Radiation tolerance is meaningless without the orbit, duration, shielding assumptions and specific failure modes tested. Compare the claims to recognised hardness-assurance standards, and treat any figure quoted without an orbit or mission lifetime as marketing rather than engineering.
Is “rad-hard” a fixed rating, or does it depend on the mission?
It depends on the mission. Rad-hard describes a family of parts qualified to withstand a given total dose and single-event environment, not a universal guarantee. A part rated for one orbit and lifetime may be unsuitable for a longer or harsher mission, so always match the rating to your specific orbit, duration and fault-tolerance requirement.
Is low Earth orbit the harshest radiation environment?
Not necessarily. Earth’s magnetic field shields LEO from much of the galactic cosmic radiation, but the South Atlantic Anomaly and the edges of the radiation belts create localised hot spots, and higher inclinations pass through more exposed regions. Total dose depends on altitude, inclination, shielding and mission length, so LEO is neither automatically safer nor harsher than other orbits.
How would you schedule an AI workload around the eclipse cycle?
Treat sunlight as the compute window and eclipse as the coast. Run the heavy, checkpointed work while the array is generating and the batteries are charging, then pause or drop to a low-power state in shadow. Bursty inference that can be interrupted and resumed fits this rhythm far better than a continuous training job that cannot stop.
Why not place orbital data centres in geostationary orbit, where sunlight is nearly constant?
Geostationary orbit offers near-continuous sunlight, but it sits in a harsher radiation environment than LEO and is roughly 36,000 kilometres away, which adds latency and makes servicing far harder. LEO’s closer range and lower dose are traded against the eclipse cycle. Once again, the orbit choice decides which problem you are willing to live with.
