How much credence should we give to this proposal? I will first explain the technical challenges that make it seem very complicated and costly for a benefit that is unclear, before offering some hypotheses about the reasons behind these announcements.

Artificial intelligence involves a lot of computation

Large-scale machine learning is at the heart of generative artificial intelligence (AI) applications (such as ChatGPT and similar systems). This requires a large number of computing processors, which consume electricity and release the energy they consume as heat. On Earth, a large AI computing data center can consume 100 megawatts (MW) or more; there are even plans to build data centers with a capacity of 1,000 MW.

From a technical standpoint, this raises four issues in space:

• Electronic components are exposed to cosmic rays, which can cause malfunctions;

• We need to generate electricity;

• The heat generated must be dissipated;

• The facilities are not easily accessible for maintenance work.

Radiation interferes with electronic components

Recently, an Airbus A320 veered off course in mid-air because its flight control computer had been struck by cosmic rays. In fact, in space—and, to a lesser extent, at high altitudes (commercial airliners fly at altitudes of about 10 kilometers)—electronics are subjected to extreme conditions.

At worst, cosmic radiation can destroy certain components; at best, it can temporarily disrupt their operation. For this reason, suppliers design special radiation-resistant processors for use in spacecraft, such as the LEON and NOEL-V, but these have modest computational performance (for example, the NOEL-V, though modern, is about ten times slower than a single core in my laptop, which has 12 cores). If components intended for conventional terrestrial applications are used in space, radiation can cause failures that require each processor to be restarted—ranging from once every few weeks to several times a day, depending on operating conditions—with the frequency increasing as the processor’s feature size decreases (i.e., higher performance).

Generate enough electricity

Tech giants are currently discussing plans to build onshore data centers that consume around 1,000 MW. By way of comparison, reactors at French nuclear power plants have nominal electrical outputs (the amount they can produce at 100% of their normal operating capacity) ranging from 890 MW to 1,600 MW. In other words, such a data center would consume the entire output of one of the smaller French reactors.

However, in space, generating electricity requires the use of solar panels or more exotic and rarely used methods—such as nuclear microreactors or radioisotope power generators—the latter being used for probes traveling to regions far from the Sun, where it would therefore be difficult to use solar panels.

Today, the solar panels on the International Space Station generate about 100 kilowatts (kW) of power—in other words, 1,000 times less than what a 100-MW data center consumes. Depending on the orbit, it may be necessary to rely on batteries (which have a limited lifespan) to power the satellite during periods when it passes through Earth’s shadow, or to accept that each satellite will only operate part of the time, which poses other problems.

Dissipate heat

It may seem surprising, given how cold space is, that it’s difficult to dissipate heat there. On Earth, we dissipate heat from data centers directly into the air, or via liquids, and then release that heat into the air through a cooling tower. In space, there is no air to which heat can be transferred, whether by conduction or convection.

Thus, the only way to dissipate heat into space is through the light radiation emitted by any object. When an object is very hot—such as iron heated to white heat—this radiation is intense (and in the visible light spectrum). In contrast, for objects such as a running computer or a human body, this radiation (in the form of infrared light, invisible to the human eye but detectable with special cameras) is not very intense. Therefore, large radiator surfaces are needed to dissipate heat into space. Managing heat dissipation is no easy task in a satellite…

Very down-to-earth problems

Let’s turn to some more practical issues. When there’s a problem at a data center on Earth, we send a technician. In space, that would require a space mission. While some tasks could be performed by robots, the level of complexity is orders of magnitude greater than that of maintenance in a building on Earth. Moreover, solar panels and other components have a limited lifespan. Finally, communicating with a satellite is more complicated and offers lower data throughput than installing a fiber-optic connection in a well-served area on Earth.

Of course, there would also be the issue of the considerable amount of equipment that would need to be transported into orbit, as well as the cost of launches and assembly.

We can also mention the pollution of the sky—which interferes with astronomical observation—caused by the passage of satellite constellations, as well as the pollution of orbits by debris from destroyed satellites.

In short, even if it were technically possible to perform artificial intelligence calculations in an orbiting satellite (or on a lunar base), it would come at a considerable cost and involve significant challenges. In the arguments put forward by those advocating for space-based data centers, it’s hard to find a good reason for so many complications. Among the justifications offered is the idea of escaping national laws by operating in space.

So why are we even talking about putting data centers into orbit?

The interesting question, rather than whether it would be possible to build an AI data center in space, is therefore who benefits from discussing projects in the media that are more akin to science fiction than to realistic industrial development. It is, of course, risky to claim to identify the objectives behind such communications, but we can offer a few hypotheses.

The U.S. space industry, particularly SpaceX, is fueling the idea of space as the final frontier, with its sights set on establishing a presence on Mars—or even colonizing the planet—regardless of the fact that it is cold (an average of −63 °C at the equator), has a very thin atmosphere, and offers no protection against cosmic radiation—in other words, it is extremely hostile to life.

The artificial intelligence industry, for its part, promotes the idea of surpassing the human brain.

Both of these industries have a pressing need for capital—for example, OpenAI has $96 billion (81.2 billion euros) in debt. To attract investors, they need narratives that inspire people. Both are fueling the “ fear of missing out ” (or FOMO)—the fear of missing out on a major development and becoming obsolete.

In fact, this strategy is working. The proof is that I wrote this article, which—even though it’s meant to explain just how unrealistic these projects are—ends up giving them even more publicity…