Topic
AI energy and infrastructure: data centres in space, efficient computing and responsible use
AI's growth depends on electricity, cooling and hardware, and on how people learn to use it. Three TechDrive Zürich guests look at those limits from the research side: Bilge Demirköz, a professor of physics at METU, on AI data centres in space; Daniel Dobos of Swisscom, who spent much of his career building detectors at CERN; and Yulia Sandamirskaya of ZHAW, who works on neuromorphic computing.
This page collects what they said, with a link to each conversation. It summarises their views; the episodes have the full context.
Would putting AI data centres in space solve the cooling problem?
Bilge Demirköz argues that the debate gives too much attention to electricity supply and too little to removing waste heat. Space offers solar power, but computing equipment still produces heat, so she asks for the complete system to be examined: power generation, radiators, orientation and the surrounding environment. She questions the idea of effortless cooling in space; she does not claim orbital data centres are impossible.
Daniel Dobos adds a risk from his CERN experience with electronics in high-radiation environments: radiation can flip a bit from zero to one, and solar storms add further disruption. He asks how an AI system would detect that a calculation had been altered, especially when it supports critical decisions. He recognises advantages such as access to solar energy, but wants the engineering risks to get equal attention.
What happens when many companies launch satellites at once?
Bilge Demirköz's concern goes beyond whether one spacecraft works. Collision fragments can strike other objects and create more debris, the cascade associated with Kessler syndrome, and launches and re-entry have cumulative effects of their own. Her point is that commercial plans need to be assessed together, because their consequences are shared. She presents collision cascades as a risk, not an inevitable event with a fixed timetable.
Does better AI have to mean more energy?
Daniel Dobos describes today's progress as partly throwing computing at our ignorance: large amounts of computing and energy compensate for an incomplete understanding of how learning works. He points to the human brain, at roughly 20 to 50 watts, as a reason to expect more efficient approaches informed by physics and neuroscience, and offers it as an illustration rather than a like-for-like benchmark.
Yulia Sandamirskaya makes a similar case from neuromorphic computing. A fruit fly performs acrobatics on about 100,000 neurons, and biological systems act in the real world on a few watts through sparse, time-based processing. In her view, scaling static models adds connections but misses that kind of processing.
How should children and educators approach chatbots?
Bilge Demirköz compares AI to fire: extraordinarily useful, but something to hand over with judgement. As an educator, she worries about tools that replace the effort of learning or keep reinforcing what a user wants to hear, particularly for children who cannot yet question the answers.
Daniel Dobos describes sitting with his children while they explored image and text generators. They came to ask, on their own, whether the tools could be used to harm someone, and then realised that someone else could do the same to them. He wants that exploration to happen with an adult present to talk it through.
When is AI worth adding at all?
Bilge Demirköz asks what AI improves when an efficient existing method already meets the need, and suggests comparing performance, reliability and resource use before replacing it. Where she is enthusiastic, the common thread is verification: her lab's local Qwen model still produces errors the team has to catch, and an AI-assisted filter design looked unfamiliar but, in her account, worked when built.
Episodes on this topic
AI Data Centres in Space: A Physicist’s Reality Check
Episode 12 · Bilge Demirköz · Middle East Technical University (METU)
We Throw Computing at Our Ignorance, and It Works
Episode 9 · Daniel Dobos · Swisscom
Nobody Has Left a Robot Working Overnight
Episode 8 · Yulia Sandamirskaya · ZHAW Zurich University of Applied Sciences