It came as no surprise that when announcing the acquisition of xAI by SpaceX Elon Musk commented that within 3 years space would become the most “economically compelling” place for AI. Musk of course is well known for liking big audacious concepts and tight timelines that won’t ever be met.
But the data centre concern is real as the struggle to source enough power and water means demand far outstrips AI ready data centre supply.
So to space it is. The idea may be sci-fi in nature but it is not new. Alphabet (Project Suncatcher with Planet Labs as partner), Europe (ASCEND) and China (Xingshidai) are all at some stage of advancement in getting data centres into space. There are also a number of startups with Starcloud probably the most advanced as last November they launched a 60KG satellite with an NVIDIA H100 GPU as a prototype.
But is it really practical? The engineering challenges are solvable, with a giant bank balance. The harder problems have no clear answers.
The real killers
The first is chip obsolescence. In its latest results Alphabet announced they were almost doubling their 2026 Capex to $175-185billion and that much of this spend was for renewing chips and servers in existing data centres.
The best AI chips currently are the Nvidia Blackwell GPUs that came to market in late 2024. Then at CES this year Nvidia announced their latest chipset, the Rubin, with faster speeds and lower power consumption. It is already in production and they are expected to be on sale in the second half of this year. So that data centre you sent into space last year will be out of date next year.
Doing space walks to swap out thousands of chips or salvaging/recycling them for their minerals is possible, we’ve done it on the Hubble. But at what cost?
So what to do when the data centre is no longer usable? Send your multi-billion dollar investment hurtling into deep space? Downgrade it to use as an over-specced and very expensive email server?
Then, what about security. Actual physical security not cyber security (I’m assuming the firewall is in place and strong).
In 2021 Russia blew up one of their own satellites. It was a weapons demonstration, not an accident. They used a ground-based A-235 “Nudol” anti-ballistic missile launched from Plesetsk Cosmodrome to deliberately destroy Kosmos 1408, one of their own defunct Soviet-era satellites that had been in orbit since 1982.
It worked while also creating chaos with thousands of pieces of space junk in the process. To be precise, 1,789 trackable debris pieces, because space is that precise. But you’d have to imagine nations aiming missiles at critical data centres as there really is no way to protect them in space, or on earth. They are sitting ducks and AWS data centres in the UAE and Bahrain have targeted by Iran in the current war.
The solvable stuff
Even setting those killers aside, the engineering is no picnic.
Getting there
The big one is getting the data centre into space. This is about rockets and here prices are falling. When the Space Shuttle was the only route to space the cost was almost $55k per kilogram. The SpaceX reusable Falcon 9 charges around $2,700 per KG, Falcon Heavy around $1,500. That is still a crazy price for a data centre weighing potentially thousands of tons, depending on the size (a large data centre would be thousands of tons). The next generation Starship has SpaceX targeting $100-$200 per KG. Cheaper but easily over $1billion just to get a modest size data centre into space. Further while Starship is operational it is not currently taking heavy cargo with no current timeline for that service.
Some speculative reports suggest SpaceX could eventually get costs under $20 per KG, but you can’t even get a 1KG package across the US for that price. So that’s certainly a non-starter.
Powering up
Once in space the next point is power and solar comes to the rescue. In space solar is multiples more efficient because there are no clouds, atmosphere or night. The panels used also have a higher conversion rates of between 30%-40% (back on earth it’s a little over 20% at best).
There are also ideas floating around to use small modular reactors (SMRs) built within the data centre for power. The reality here is that SMRs don’t exist yet. Sure there is a lot of research and investment happening, including TerraPower backed by Bill Gates. But the first test plant is only expected by 2030 and that’s probably optimistic.
Keeping cool
The power is in large part for powering the AI chips. But around a third of data centre power is about cooling the chips and as we all know from school (or the movies) space is cold. But space is also a vacuum and the heat doesn’t dissipate. So you use water piping carrying the heat to radiative panels that send the heat off into deep space. This is used on the International Space Station (ISS) and the tech works. Future chips could also be designed specifically for space and optimised for radiative, not convective, cooling.
In fact Nvidia CEO Jensen Huang announced a space AI chip during the GTC keynote last week. The Space-1 Vera Rubin Module is purpose built for space AI data centres but there were no details on shielding and on cooling he commented “we have to figure out how to cool these systems out in space”. They already have a number of startups on board including Starcloud and Planet Labs from above.
The latency problem
The next challenge is latency. How long it takes the data to do the trip from earth, be processed, and return to earth. Nobody wants to be left waiting for their ChatGPT answer to critical question.
Here again SpaceX has the Starlink satellites that sit in low earth orbit about 550 km above earth (this is being lowered to 480km in 2026). Now Starlink is consumer broadband and data centres would need enterprise grade connectivity such as that from Jeff Bezos’ Blue Origin’s TeraWave. Satellite deployment here is expected late next year followed by commercial operations. There is also the European and London listed Eutelsat OneWeb that has been operating since 2023.
Radiation and shielding
The next problem is risks from radiation, which is totally different from the low-level electromagnetic fields (EMFs) found on earth. You can use aluminium to shield the data centre but this adds yet more weight that all adds to launch costs. The James Webb Space Telescope operates at L2 (1.5million km from Earth, in the direction away from the Sun), outside Earth’s magnetosphere. Its electronics were designed with significant radiation tolerance as a requirement. The Hubble launched in 1990, in low Earth orbit of 540km, has seen its instruments degraded by radiation over its lifetime and needed five servicing missions in its first nineteen years before being ‘retired’.
The verdict
Elon Musk said data centres in space within three years. This is more than a stretch. The reality is that data centres in space are definitely possible but truth is they’re probably decades off at best. Even than the costs are not likely to make sense compared with earth based data centres. It is simple cheaper and easier to build them on earth. Yes for now we have serious power and water shortages, but those are far easier to solve then data centres in space.
The technology will get there eventually. But ‘eventually’ looks a lot more like the 2040s than Musk’s 2029.
AI in the Wild
AI in the Wild is a regular column from Simon Brown.
AI is everywhere and only getting better. Record capital raises and valuations, and competing LLMs are all fun, but meaningless to our every day lives. This column will focus on how is it impacting us in the real world.
5th | 15th | 25th every month
