Hi all,

I'm starting something new with this newsletter: a build-in-public series on where AI, telecoms and satellites meet. One project at a time, with the engineering shown rather than hand-waved. This is the first.

I wanted to begin somewhere honest, so I started with the question every satellite engineer has to answer before anything clever happens: will the link even close? Can the signal actually get from there to here strongly enough to be useful? It's not glamorous, but everything else depends on it.

The case I find most interesting right now is a phone talking directly to a satellite, with no tower in between. That's the idea behind 5G Non-Terrestrial Networks (NTN), and it's harder than it sounds. A handset trying to reach a satellite a few hundred kilometres up is fighting around 159 dB of signal loss. Since every 3 dB halves the power, what arrives at the phone is a tiny fraction of what left orbit.

To see it properly I built a small Python tool that adds up everything helping the signal and subtracts everything hurting it, then plots the result as the satellite crosses the sky. I ran three setups through it, and the contrast told the whole story:

  • Phone, low-orbit S-band: ~5.5 dB of margin. It works — but it's tight, and it drops out when the satellite sits low on the horizon.

  • Fixed Starlink-style dish, Ku-band: ~28 dB, with room to spare.

  • Geostationary link to a big dish: a flat ~15 dB right across the sky.

That gap between the phone and the dish — roughly 23 dB — is the whole reason your phone can manage a satellite text message while real broadband still wants a dish on the roof. The dish simply has a much bigger antenna doing the heavy lifting.

The bit that caught me out was the geostationary curve. I expected it to slope like the others, and it came out almost flat. Then it clicked: from 36,000 km up, the satellite is so far away that it never really gets closer as it moves across your sky, so the margin barely changes. Obvious once you see it.

Where this goes next: if the margin depends on geometry, and geometry is predictable from a satellite's orbit, then you can start predicting the link ahead of time. That's the door into the part I actually care about — putting some machine learning behind link planning. But it only means anything once you can answer the dull question first.

Next time I'm feeding the tool real orbital data so it draws an actual satellite pass over my city instead of a textbook curve.

The full technical write-up and the open-source code are linked below if you want to dig in.

Until next time,
Shaif

Subscribe to The Signals AI

Keep Reading