AI has only been around in everyday life for a couple of years, but it’s resulted in a huge need for data storage and processing. The race to plug it into everything means that data centres are being built at a frantic pace all over the world, and they consume staggering amounts of electricity. Training LLMs requires massive computing runs that can last weeks or even months. For example, training GPT-4 consumes 50 gigawatt-hours of electricity. Asking an AI chatbot a question uses ten times more energy than finding the answer with a simple web search.
Then there’s the water problem. High-density AI chips run so hot that air cooling isn’t enough. New facilities are turning to water evaporation. It’s more effective and energy-efficient than recirculating water, but a large data centre consumes millions of litres of water a day, which is a huge drain on local water supplies.
There can also be a local impact. Would you like a massive warehouse-like facility in your neighbourhood with 24/7 lighting and the sound of air conditioners continuously running? A local example is eThekwini, which is planning to build a massive 400 MW AI data centre. This power guzzler will burn up the equivalent of 25% of its current electricity supply. Not surprisingly, the local communities are up in arms.
So what’s the solution? Why not put the data centres into space? Some people think this is less crazy than you think it is. The idea is that instead of building an AI data centre on earth, it could be built in orbit in the form of a swarm of satellites interconnected by optical data links.
Elon Musk (who else) says that data centres in space could run on satellites in a ‘sun-synchronous’ orbit so they will be able to continuously capture unlimited solar energy. He has already merged SpaceX with his AI startup, xAI with this aim in mind, and applied for a licence to build an orbital data centre consisting of up to a million AI satellites.
Other companies are also exploring the idea. Google’s Project Suncatcher plans to use a constellation of solar-powered satellites that would run on its own TPU chips and transmit data to one another via lasers. Data centre company, Lonestar aims to put a small data centre in orbit round the moon in 2027, and is working on storing data securely up there, where it would be protected from the risks on earth. It already has customers lined up, including the State of Florida and the Isle of Man government.
The USA startup, Starcloud has launched a fridge-size satellite carrying an ordinary Nvidia GPU of the type used in AI data centres, and used it to train a small AI language model, NanoGPT on the works of Shakespeare. Meanwhile, Nvidia is working on a chip designed to survive the rigours of space, having just launched the Vera Rubin space module that can run AI models from orbit. Others are working on modular space stations that could host computing hardware.
But does it make sense? Computing hardware must be protected from radiation through shielding or error-correcting software. Getting hardware into space is another challenge. Rocket launch costs are a significant barrier to building large orbital data centres, not to mention the need to replace onboard chips every five years. SpaceX delivers payloads into orbit at a price of around $1500/kg with its Falcon Heavy rocket. The Suncatcher team estimates that liftoff costs would need to fall to under $200 per kilogram by 2035 for their vision to make sense. They predict that this could happen by the early 2030s.
Another challenge is cooling. On earth, servers are cooled with air or water. In space, there is no air and no easy way to dissipate heat. To cool off, orbital platforms need large radiators that can dump heat into the vacuum of space, adding significant mass that has to be launched on rockets.
I am wondering what to make of all this. For now, space-based computing is still what Google calls a ‘moonshot’. It’s ambitious and technically demanding, but it could reshape the future of AI infrastructure. What’s possible today is beyond anything I imagined ten years ago. Look at solar panel costs, which have dropped by over 90% since the early 2000s. When humans have a problem they solve it, but there are tradeoffs. When I gaze at the stars I don’t want to see hundreds of thousands of mini Starlink satellites. But on the other hand, I don’t want a data centre in my back yard. Where will we be in ten years time?
| Tel: | +27 11 543 5800 |
| Email: | [email protected] |
| www: | www.technews.co.za |
| Articles: | More information and articles about Technews Publishing (SA Instrumentation & Control) |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version