In November 2025, a startup placed an Nvidia H100 graphics processor in orbit aboard a 60-kilogram satellite. In December, it trained a small language model on the complete works of Shakespeare. Google is preparing two prototype satellites equipped with its own chips for early 2027. And SpaceX has filed a filing with the US regulator covering one million computing satellites. The question is therefore no longer whether it's possible, but whether it's a good idea.
Throughout the summer, we documented the race for compute and its very physical constraints: electricity, land, silicon. Faced with these walls, one idea keeps coming back insistently in Silicon Valley: what if we moved it all up there? Let's look at it seriously.
Where things actually stand
The topic has gone from fantasy to engineering in less than two years. Here's the state of play.
| Player | Where it stands | Stated ambition |
|---|---|---|
| Starcloud | One H100 in orbit since November 2025, model trained in December | Second satellite in late 2026, then a constellation targeting several gigawatts |
| Google (Suncatcher) | Radiation resistance tests validated in the lab | Two prototypes in low orbit in early 2027 |
| SpaceX | Filing submitted to the regulator, nothing in flight | Up to one million computing satellites |
| Axiom, Lonestar, Aetherflux | Demonstrators and projects under study | Orbital nodes, or even a lunar data centre |
Two remarks on this table. First, the gap between what flies and what is announced is dizzying: a 60-kilogram satellite on one side, a million satellites on the other. Second, the speed of valuation raises questions: Starcloud became a unicorn seventeen months after leaving its accelerator, which says as much about investor appetite as about the technology.
The argument that holds up: energy
The central reasoning is sound, and it must be acknowledged. In sun-synchronous orbit, a satellite can remain almost permanently exposed to the Sun, with no night, no clouds, no atmosphere to filter radiation. A solar panel there produces, in certain configurations, several times what it would produce on the ground, continuously, which reduces the need for batteries.
Add to that the absence of land constraints, of neighbour objections, of grid connection, and you get a serious argument against the terrestrial walls we described regarding giant data centre projects. On Earth, a new power plant takes years of procedures. In orbit, the light is already there.
Google's CEO summarised the thesis by estimating that within about ten years, this will be seen as a normal way to build data centres. That's an entrepreneur's claim, not a neutral forecast, but it has the merit of being clear.
You often read that space is cold, so it's ideal for cooling machines. That's a physical misunderstanding. A vacuum doesn't dissipate heat: without air or water, there's no convection or conduction possible to the outside. The only available mechanism is thermal radiation, which is very inefficient. Concretely, a space computing centre needs immense radiators, often bulkier than the solar panels themselves. Cooling is not an advantage of space; it's its main engineering challenge. Any announcement that presents the vacuum as a natural refrigerator should alert you to the seriousness of the rest.
The environmental balance, honestly
This is the most interesting part, and the most poorly handled in corporate communications. Let's put the two columns side by side.
What counts in favour. Zero water consumption, whereas terrestrial cooling is very thirsty, as we documented in our article on the environmental cost of AI. Zero land artificialisation. No pressure on already strained national power grids. And primary solar energy, with no combustion.
What counts against. The first item is launch. Putting a kilogram into orbit costs fuel, and therefore emissions, in atmospheric layers where the impact is poorly understood. Multiplying by thousands of satellites changes the scale of the problem. The second item is end of life: a satellite disintegrates on re-entry, releasing metallic particles into the atmosphere whose long-term effects on the upper atmosphere are the subject of active and concerned research. The third is impossible maintenance: on Earth, you replace a faulty card; in orbit, the entire object becomes waste. The fourth, finally, is orbital congestion, with a risk of chain collisions in already saturated low orbits.
The honest conclusion is that we can't decide today. The balance depends entirely on two unknown parameters: the real lifespan of the equipment and the carbon cost of launch at the envisaged scale. A satellite that works for ten years has a very different balance from one replaced every three years.
What it could reasonably be
Let's set aside the fantasy of replacement. Nobody is going to migrate global infrastructure to orbit. Three uses, however, seem defensible.
Processing space data at the source. This is the most obvious application and already demonstrated. An observation satellite produces images that today must be transmitted to the ground for analysis. Processing them directly in orbit and only sending back the result saves a huge amount of bandwidth. The in-flight demonstrator analyses images to detect vessels in distress or monitor fires. Here, space isn't a hosting location; it's where the data is.
Batch training, insensitive to latency. Training a model doesn't require instant response. A task running for weeks can tolerate an orbital round trip. Conversational inference, the kind that answers you in a second, on the other hand, handles transmission delays poorly. The distinction is structuring.
Sovereign or isolated uses. Defence, areas without infrastructure, redundancy in case of disaster.
The serious objections
Financial analysts have listed four risks that sum up the debate well: radiation, which degrades electronics, impossible maintenance, orbital debris, and data governance.
This last point deserves attention, because it's fascinating and underestimated. Under which jurisdiction do your data fall when they're processed at 650 kilometres of altitude, above three different countries in ten minutes? European data regulation reasons in terms of territory. Orbit isn't one. It's a legal vacuum comparable to the one we described regarding the responsibility of autonomous systems: the technology arrives before the framework.
Finally, there's a more down-to-earth objection. This idea emerges precisely at a time when AI's capital needs are exploding and investors are looking for growth narratives. A space project is a formidable valuation tool. That doesn't disqualify it, but it does invite careful distinction between what flies and what is promised. A demonstration satellite is a real feat; a regulatory filing for a million units is an intention.
So, a good idea?
Our reading, and it's nuanced. The idea is neither absurd nor imminent.
It isn't absurd because the energy argument is real, the first demonstrations work, and processing data at the source in orbit has an obvious logic that has nothing to do with hype.
It isn't imminent because cooling remains a major challenge, the environmental balance depends on still-unknown parameters, and the gap between sixty kilograms in orbit and a multi-gigawatt constellation represents at least a decade of engineering, with plenty of opportunities to fail.
What this story mainly says is the scale of the wall AI is hitting. When companies seriously consider sending their machines into space to find electricity, that's a sign that the terrestrial constraint has become very serious. We wrote it in our analysis of acceleration: you can print money, not gigawatts.
There's perhaps a more unsettling question behind all this. Before asking how to produce more energy for AI, it would be reasonable to ask what it's for. A model that consumes twelve times more thinking tokens than necessary, a request sent without cache, an agent running idle: most of today's waste is software, not energy. Optimising costs less than a launcher, and requires no orbit.