Your satellite saw the fire. The radio couldn’t carry it to you.

A satellite passes over a peatland fire and photographs it. Then it waits for a ground station, and when one comes, the radio can't carry everything the camera just took.

We spent a decade talking about how expensive it was to get to orbit. That got solved, mostly.

A modern radar satellite generates 2 to 5 terabytes of raw imagery a day. Researchers working on the problem put it plainly: a single satellite "captures more images than it is capable of downlinking."

Some of what it saw reaches you, and the rest is deleted to make room for tomorrow.

Why the gap keeps widening

Each new generation of satellite sensor produces roughly 100 times more data than the one before. Downlink capacity over the same period improves by about three to five times. You buy a sharper camera and inherit a bigger queue.

A satellite can only send when it's passing over a ground station. So the chain runs: capture, wait for a pass, downlink, process on the ground, someone reads it. That's hours to days. Commercial providers advertise under two hours in priority areas — that figure comes from the operators themselves.

For a lot of uses, hours is fine. For a flood moving through a valley, a boat where it shouldn't be, or a peatland fire heading for next month's haze? Two hours is a lot.

The fix everyone agrees on

Put the thinking on the satellite. Instead of sending a picture and asking someone on the ground to find the boat, the satellite finds the boat and sends you a sentence.

You need far less bandwidth, because a conclusion is small. Processing onboard drops the wait to minutes.

ESA flew Φ-sat-1 in September 2020 carrying what ESA describes as the first neural network flown in orbit. Its entire job was spotting cloudy frames and deleting them before transmission — it decided what not to send.

Doing more than that means real compute in orbit, which runs into the two things space does to electronics.

Radiation and heat

A typical Earth-observation satellite flies a radiation-hardened processor running at 0.65 to 1.6 GHz. Hardened chips lag commercial ones by years, because surviving matters more than speed. Charged particles flip bits and, over time, kill silicon.

Heat is the other one. In vacuum there's no air to carry warmth away, so every watt has to be radiated. One megawatt needs roughly 1,600 square metres of radiator — a bit bigger than an Olympic pool. A gigawatt would need 1.6 square kilometres of it.

Two companies are working on these problems from different directions.

Chips that survive up there

Zero-Error Systems spun out of NTU in Singapore in 2019 and works on the radiation half. In February 2025 it released ZSOM-F01, a radiation-tolerant system-on-module. Rather than a custom hardened chip, it wraps an off-the-shelf commercial FPGA in the company's own circuitry against latch-ups and bit flips. Customer testing opened that April.

At Space Summit 2026 in Singapore this February, ZES introduced Smart-LCL. It's a pair of radiation-hardened current limiters that cut power to a line when a particle strike triggers a latch-up. Samples are available on request.

Hardened parts are expensive and slow. If commercial silicon can be made to survive, the gap between what flies and what's current narrows.

Separately, NTU researchers published a feasibility study in Nature Electronics on whether data centres could run in orbit on solar power.

Compute on the spacecraft

This is a different problem and a different company. The two aren't connected.

Stardetect (星测未来) is a Beijing company founded in 2020 out of Tsinghua. On 11 September it announced A+ and A++ rounds worth several hundred million yuan — tens of millions of US dollars. The rounds were led by North Creation Fund, Huaying Capital and Xinshang Capital.

The company says it has 37 advanced-process GPUs operating across 39 payloads in orbit, with one unit running over four years. For scale, that's a small amount of computing by ground standards and a lot by satellite standards — most spacecraft fly nothing remotely like it.

The money goes toward 扶摇计划, a plan for 303 compute satellites of its own alongside roughly 3,000 partner spacecraft. 39 payloads are flying today.

True or Not

Two statements. One's real, one I made up.

1. A single radar satellite can generate 2–5 TB of raw imagery a day.
2. Satellites downlink roughly 80% of what they capture.

(Answer at the very bottom.)

What to do with this

If you buy satellite data: ask what the latency actually is, from capture to your screen — not the revisit rate. Those are different numbers and vendors quote the flattering one.

If you build satellites: the constraint has moved to the radio, and to what you can do with the data before it goes through.

If you fund this: the number that counts is the one already in orbit. Ask which part of a stated capability is flying today and which part is what the round pays for — in this field they are routinely quoted as one figure. Stardetect says 39 payloads are up; the 303 satellites are the plan.

The satellites already see more than they can tell you.

If you missed the last issue, go back and read it: China's new Five-Year Plan promotes space to a pillar industry — and backs it with a first-ever law.

Cheers!

— shirley

True or Not — 2 is invented. Most satellites downlink far less than they capture; that gap is the whole story.

Cover: Antennas at Awarua Satellite Ground Station by Satopsnz, cropped, licensed CC BY-SA 4.0.