AI in Extreme Situations Why Shoot AI to the Moon When You Can Sink It?

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

The idea of establishing data centers in Earth orbit is not new, nor is it currently economically viable. Tech celebrities are dreaming of AI from space, but there is other uncharted territory to be discovered: the sea.

The increasing energy requirements of data centers, particularly due to AI workloads, have led to ever more radical infrastructure ideas in recent years. There is talk of locations in cool regions, then again of locations in areas where the weather is largely stable, such as deserts. Then there is talk of data centers near renewable energy sources—or near decommissioned nuclear reactors that could be put back into operation.

A vision that is as fascinating as it is controversial regularly crops up in conversations and brainstorms: data centers in space. The idea behind it seems plausible at first. There is almost unlimited solar energy available in orbit. There are no space problems like on Earth. Well, that's easy to believe if you don't necessarily want to deal with space debris.

And yes, the waste heat could theoretically be released into space. Well ... it's not that easy because of the vacuum. Anyway, let's stick to the green playground for now. At the same time, data could be processed directly where it is generated—for example on earth observation satellites. These are all valid ideas that are already being implemented.

Calculating in Space

However, when it comes to shooting data from Earth into orbit in order to process it there and send it back again, this is precisely where criticism comes in, including from analyst firm Gartner. In a recent analysis, the experts came to the conclusion that orbital data centers will not serve terrestrial requirements in the foreseeable future and that companies would be better off focusing their investments on infrastructure on Earth.

The reasoning is less visionary than physical and economic. Transportation costs dominate every calculation, as IT hardware, energy supply, cooling and structure must first be brought into orbit. Even if launch costs fall sharply, mass and volume remain decisive cost drivers. Added to this are extreme environmental conditions such as radiation, vacuum and temperature cycles, which require specially hardened components and further increase costs. Our space fan Hendrik Härter has already explained the challenges that electronics would have to face when Jeff Bezos shared his vision of AI data centers in space at Italy Tech Week 2025.

First the Construction, Then the Maintenance

Another key issue is maintenance. While servers in terrestrial data centers can be continuously replaced, any repair in orbit would require complex missions or sophisticated robotics. Without maintenance, systems would quickly contribute to additional space debris; a problem that is already being critically discussed today due to its dangerous nature.

At the same time, data transmission also remains a limiting factor, as large volumes of data would have to be reliably transmitted between earth and orbit. Against this backdrop, the idea of transferring a relevant proportion of global cloud capacities into space currently makes neither economic nor practical sense. Even the often cited advantage of cooling is put into perspective on closer inspection, as heat in a vacuum can only be dissipated via radiation, which requires large radiator surfaces and additional mass.

There is also an aspect that is often overlooked in visionary discussions: space problems cannot really be solved by orbital computing. Data centers require enormous infrastructure—energy conversion, communication systems, structural modules—which also take up space and resources. The physical location merely shifts from the earth's surface to orbit, without the fundamental scaling issues disappearing.

Why Look Up ... And Not Down?

On the other hand, a concept that also seems unusual but is physically much more favorable is much closer to practical implementation: data centers in the sea. Underwater data centers use the high heat capacity and stable temperatures of seawater for cooling, which can significantly reduce one of the biggest energy consumers of traditional data centers. At the same time, many large conurbations are located close to the coast, so that latency times remain low and existing fiber optic infrastructure can be used.

Projects that have already been carried out show that this approach is not just theoretical. Microsoft tested a sealed data center module on the seabed off the Scottish coast for several years with "Project Natick". Among other things, the results indicated higher hardware reliability, which was attributed to the stable environment and the sealed design. Underwater data centers have also been tested or announced in China, including in connection with offshore energy projects. Such concepts enable direct coupling to renewable energy sources such as offshore wind power, which opens up additional efficiency potential.

The Other Side of the Coin

Of course, underwater data centers are not without their challenges. Maintenance work is understandably more complex than on land. Material durability in salt water must be guaranteed in the long term. At the same time, ecological effects must be assessed. However, these disadvantages are considerably reduced compared to orbit. If maintenance in space is considered a solvable problem in many visions, then this applies all the more to systems that are located a few hundred meters below the surface of the sea and can be reached with existing offshore technology.

Perhaps the future of data centers does not lie above our heads, but under the surface of the water. While data centers in space currently remain primarily a long-term vision with special niche applications, underwater data centers could become a realistic addition to existing infrastructure much sooner. They address specific problems such as cooling, energy efficiency and space availability without having to overcome the fundamental physical and economic hurdles of space travel.

Whether we will actually need the data center boom at the end of the day, in the next 5, 10, 15, 50 years, or whether ways can be found by then to significantly reduce AI's hunger for energy and space, is of course another matter. (sb)

AI in the industry

ETMM Dossier – Artificial Intelligence