On Earth, live and electronics get to pretend the universe is gentle: We sit under a magnetic cocoon.

The Earths magnetic field bends and corrals a lot of charged particles that would otherwise slam into the atmosphere and the ground. The polar lights are indicative of particles which hit the upper atmosphere and dump energy there instead of into your laptop or your DNA.

Low Earth orbit is still inside much of that protective bottle. It is not deep space: most of the time you are still inside the magnetosphere, but you pass through regions where trapped particles dip closer to Earth. The South Atlantic Anomaly is the famous example, a patch where satellites see a much higher rate of hits. Operators notice because sensors glitch, memory errors spike, and instruments get noisy.

Go higher and the protection changes. The Van Allen belts are zones of trapped particles shaped by the magnetic field. They sit above typical LEO altitudes and below geostationary orbit.

Geostationary orbit is far outside LEO, and you spend much more time in harsher particle populations and different dose conditions.

Radiation matters because a modern chip is a giant field of tiny, delicate transistors storing tiny amounts of charge. A single energetic particle can change how or even if that works.

• Bit flips: A particle passes through silicon, leaves a trail of charge, and a memory cell or latch interprets that as a 0 becoming a 1. That is a single event upset. It does not break the chip, it just corrupts state. The usual defense is error detection and correction, ECC in memory, parity, scrubbing, retries, and lots of "trust but verify" in data paths. That defense costs area, power, and latency. You carry more bits than you asked for, you spend cycles checking them, and you sometimes redo work.
• Latchup and destructive events: Some particle strikes can trigger parasitic structures in CMOS so a section of the chip effectively shorts power to ground. If you are lucky, the system detects it and power cycles that block. If you are unlucky, you get local overheating and permanent damage. The defense here is design techniques, guard rings, current limiting, fast power cutoffs, and redundancy. Redundancy costs density. You either duplicate blocks so you can route around a failed one, or you accept that some percentage of silicon will be lost over mission life and you overprovision from day one.
• Total ionizing dose: Ionizing radiation gradually traps charge in insulating layers and at interfaces. Threshold voltages shift, leakage rises, timing changes, noise margins shrink, and eventually the chip that used to pass validation at room temperature starts failing at its corners. This is why space hardware often talks about dose ratings and mission lifetimes, not just "it works today." The defenses are process choices, device layout choices, and again, guardbands.

How does space hardware survive?

Shielding and Redundant Design.

• Put material around the electronics, often aluminum as a structural and shielding compromise. It helps, but mass is the tax you pay forever. It also only helps a bit: High energy particles penetrate a lot of material, and some shielding configurations create secondary particles when the primary hits the shield. You can reduce dose and upset rates, you cannot build a perfect bunker without turning your satellite into a brick.
• A lot of rad hard parts use larger transistors, older process nodes, thicker oxides, and conservative voltages. Larger devices store more charge, so a stray deposit is less likely to flip a bit. Thicker insulators tolerate more ionization. Conservative voltages and clocks give you more timing margin as the device ages. All of this makes the chip slower and bigger for the same function.
• On top of that, Logic level hardening, software paranoia. ECC everywhere. Voters and triplication for critical state, triple modular redundancy where you do the same computation three times and take the majority result. Watchdogs, reset domains, isolation boundaries, and constant self checking. You get correctness, but you spend transistors and joules on distrust.

Every one of these defenses hits the same three budgets.

Area goes up because you replicate circuits and add check bits and voters.

Power goes up because more circuitry toggles, and because robust designs often run at higher voltages than bleeding edge consumer silicon.

Speed goes down because checking takes time, retries take time, and wide safety margins force lower clocks.

Space hardened electronics is built for survival, not speed. It is very reliable, and slow as fuck.

If we hold all that against a H100 GPU as is being used for AI, we can see that this is a lost cause without launching any satellite.

The H100s GH100 die is about 814 mm2 on TSMC 4N with about 80 billion transistors.

This kind of die does not fly into space and survives longer than an afternoon.

https://indico.esa.int/event/165/contributions/1218/attachments/1205/1425/05b_-_KIPSAT_-_Presentation.pdf

ESA talks about 65nm processes, 16 times larger structures, 250 times larger area structures, for their space hardened compute.

The same amount of transistors for a H100 would be a 0.2 m2 slab, which also means that energy goes up and clock goes down, down, down.

Some compute ins pace runs on 28 nm structures, so a 50x times area increase compared to a H100. That's 40.000 mm2 for the same amount of transistors.

Or, in other words, a GPU does not work in space.

At all.

If you try, 80 bn transistors become around 25 bn transistors for overhead (n = 3) and redundant reserve.

So even if you sent a 4 nm node chip into space, it's no longer an 80 bn monster, but only a relatively modest 25 bn transistor GPU fragment.

Running hotter cools better, to the fourth power. So even relatively modest temperature increases will pay of big time in terms of radiative cooling (100ºC -> 120ºC), but that means more leakage, more power, less clock. So your 25 bn transistors net capacity will calculate slower than on earth.

A GPU does not work in space.

We can now talk about bandwidth, latency and spectrum carrying capacity, because we also want to TALK to those data centers from earth, but that's kind of a moot point already.

We can then talk about launch costs, kilograms, lifetime, and the atmostpheric effects of that material on re-entry, but it will never come to that.