Satellites Are Multiplying Fast. AI May Be the Only Way to Keep Them Running
With up to 70,000 new satellites possible in the next five years, human operators alone can't keep pace. A space industry veteran explains what AI can do in orbit, and why the rules haven't caught up yet.
Key points
- Goldman Sachs estimates up to 70,000 new low-Earth orbit satellites could launch within five years.
- AI could process satellite imagery in orbit, cutting the time between data collection and useful results.
- Current laws and insurance policies were written for human-controlled systems, leaving accountability gaps when AI makes a mistake.
- Gradual adoption is likely, with AI handling routine decisions first while humans approve the bigger calls.
- AI could also shorten spacecraft design and manufacturing timelines, letting smaller teams do more.
Roughly 16,000 active satellites circle the Earth right now. Market researchers at Novaspace think 43,000 more will launch over the next decade. Goldman Sachs puts it higher: up to 70,000 new satellites in low-Earth orbit (the band a few hundred miles up, where most communications and observation hardware sits) within five years alone.
The trend is what matters. Constellations are growing fast, and ground-based human management can't keep up.
Why can't people just run them from the ground?
A handful of satellites is manageable. Thousands aren't. Each one must coordinate with neighbours, dodge interference, respond to demand shifts and handle faults. Some decisions need to happen in seconds. Radio signals take real time to travel between Earth and orbit, so waiting for a ground controller to notice a problem and transmit instructions back up is simply too slow.
Writing for Space.com's Expert Voices section, Martin Halliwell, former Chief Technology Officer at satellite operator SES (where he led technology and R&D from 2011 to 2019) and now a partner at space-focused investment firm NewSpace Capital, argues this is exactly where artificial intelligence becomes essential.
What would AI actually do up there?
Two jobs stand out. The first is onboard data processing. Today, most raw satellite data travels back to Earth before anyone sorts it. An Earth-observation company tracking crop yields or wildfire smoke downloads enormous volumes of imagery and cleans it up on the ground before a single useful number reaches a customer. Our 23 July story on NASA's Google AI model tested inside a satellite showed one concrete version of this: the model sent back plain-text descriptions instead of raw imagery, cutting wildfire alert times from 90 minutes to near-instant. If the satellite's onboard computer runs an AI model, it can do that filtering itself and transmit only the relevant pieces. Faster answers, lower data costs.
The second job is managing network capacity. Demand for satellite bandwidth (the volume of data a signal can carry) shifts constantly: over cities, at disaster sites, above busy shipping lanes. An AI system could track those shifts in close to real time, steer signal beams toward wherever demand is highest and dial power up or down accordingly. A fixed beam wastes capacity; a steered one extracts more useful work from the same hardware.
Who is responsible when AI gets it wrong?
This is where the industry hits a wall. Laws and insurance policies were written assuming a named human makes every significant call. When AI points a beam at the wrong location and disrupts another operator's service, it's genuinely unclear whether blame falls on the satellite owner or the company that wrote the software.
Halliwell expects adoption to move in stages. Operators will likely let AI offer recommendations first, then hand over more routine calls while keeping humans in the loop for anything consequential. Clear written rules about which decisions AI can make alone, and which need a sign-off, will be necessary before the industry moves faster.
AI could also reshape how satellites are built, helping engineers draft structures and run tests more quickly, so smaller teams can tackle projects that once needed much larger organisations. The goal isn't to replace skilled engineers; it's to free them from repetitive work.
What happens next?
The technology is arguably ready. Building legal frameworks, cybersecurity standards and institutional trust is the harder task. As Halliwell puts it, AI may soon be capable of running more of a satellite network. The real question is how quickly the industry is willing to let it.
Watch the regulatory side, not the engineering side. The spec sheets are ahead of the rulebooks by a wide margin, and that gap is where the actual risk lives.
Common questions
Does this mean satellites will fly with no human control at all?
No. The near-term picture is AI handling fast, routine decisions while humans approve anything significant. Full autonomy would require legal and regulatory changes that are still years away.
How does this affect ordinary people on the ground?
Faster disaster alerts, more responsive mobile coverage in remote areas and cheaper satellite broadband are the practical benefits if AI-managed constellations deliver on their promise.
