The Data Centre That Leaves Earth
In the last post I argued that the AI infrastructure boom has quietly become a bet on physical infrastructure, and that the binding constraint is not silicon but the unglamorous business of getting electricity to a specific place at a specific time through equipment that takes years to permit and build. That was the pessimistic reading, and I stand by it. This is the other end of the same argument, because when a constraint gets tight enough, somebody always starts looking for a way around it, and the way around this one is genuinely strange.
We may be about to build the first industrial infrastructure in human history that is economically better off leaving the planet.
I am aware of how that sounds. It is also not science fiction any more, which is the part that made me want to write it down.
Google has put a spacecraft behind the idea
Project Suncatcher is Google's programme to work out whether AI compute can be deployed at scale in low Earth orbit, and it has now progressed from a paper to an actual mission. The first flight is designed to test Google's TPU hardware against radiation, launch loads and the thermal environment of space, with a second stage planned for 2027 to test high-bandwidth laser communication between satellites.
Whether Google can put a TPU in orbit was never in doubt. Of course it can. The question that makes this interesting is whether a collection of solar-powered satellites, connected by optical links, can behave like a data centre rather than like a fleet of computers that happen to be in the same neighbourhood. That is a much harder question, and if the answer turns out to be yes, the consequences run a very long way past AI.
Why the terrestrial answer is running out of room
The conventional response to rising AI demand is to build more data centres, which sounds like a construction problem and is actually a queueing problem. You need land, planning approval, enormous quantities of electricity, substations, transmission, cooling, and a connection to a terrestrial network, and almost every one of those has a lead time measured in years rather than months.
The limiting resource is not the chips. It is the infrastructure around the chips, and the economics get progressively stranger when you notice that the most valuable component in the building is ultimately a machine whose primary input is electricity. We are spending a great deal of time and capital building an industrial system whose job is to convert stored chemical or nuclear energy into electricity and then carry that electricity to the place where the computation happens.
Which is roughly the point at which somebody looks up.
The Sun is already there
In the right low Earth orbit a solar array has access to substantially more useful sunlight than the same panel on the ground, and can spend most of its time illuminated. Google's analysis estimates that appropriately positioned panels could produce up to eight times the output of comparable terrestrial ones, in an orbit chosen to provide near-continuous sunlight.
I want to be precise about why that matters, because it is easy to make it sound like magic. The Sun is not free because space is special. The Sun is free because we do not have to build the Sun. On Earth we construct an enormous apparatus for generating electricity and moving it to where it is needed, and in orbit the energy source is simply overhead already. The engineering problem reduces to capturing it, converting it, using it, and then getting rid of the waste heat.
That is still hard. It is just a different hard, and different hards are where the interesting engineering usually lives.
The lasers are the part being underestimated
A satellite with a processor on board is an edge computer, and we have had those for years. A thousand satellites with high-speed interconnects between them is something else entirely.
Google's proposed architecture flies the satellites in tight formation and links them with free-space optical communications, and the reason for the close spacing is unglamorous physics: the shorter the distance, the easier it is to close the link budget at very high bandwidth. Google has already demonstrated 800 Gbps in each direction with a ground-based prototype and is working towards satellite-to-satellite testing.
Once that works, the architecture stops looking like a fleet and starts looking like a computer whose motherboard happens to be an orbital formation. That is a genuine change in kind rather than degree, because the network is no longer carrying data between machines. The network has become part of the machine.
The economic loop that makes it self-reinforcing
Here is the part I find most compelling, and it is an argument about industry rather than about space.
Demand for orbital compute creates demand for satellites, solar arrays, processors, optical terminals, launch capacity, autonomous rendezvous, servicing and eventually orbital manufacturing. Every one of those industries, as it matures, reduces the cost of putting the next thing into orbit. Lower costs make larger constellations viable, larger constellations create more demand, and more demand funds the next round of investment. AI infrastructure ends up helping to finance the infrastructure required to build more infrastructure in space.
Google's own modelling assumes launch costs could fall below $200 per kilogram by the mid-2030s under sustained learning rates, at which point it suggests orbital data centre economics could become comparable with reported terrestrial energy costs on a per-kilowatt-year basis. I want to be careful with that number, because it is a projection resting on an assumed learning rate and not an established commercial result. It is enough to make the proposition technically interesting. It is not enough to make it a plan.
But the structural point survives the uncertainty in the figure. Launch does not have to become cheap before orbital industry can start, because orbital industry is itself one of the mechanisms that makes launch cheaper.
The first useful applications will be unglamorous
The first orbital computers do not need to replace AWS, and expecting them to is the quickest way to conclude that none of this works. They need to solve problems where processing the data where it is collected is already the better answer.
Earth observation is the obvious case. A satellite collects enormous quantities of imagery and traditionally sends the raw data down for somebody on the ground to process. If the useful output is only that there is a ship at a particular location, transmitting several gigabytes in order to discover that fact on the ground is a strange use of a very expensive downlink. Compute the answer where the sensor is and send the answer.
This is already happening commercially. India's TakeMe2Space plans to launch a small orbital computer carrying NVIDIA Orin NX processors so customers can run models in orbit and transmit the analysis rather than the underlying data, and the company reports 23 customers across agriculture, mining, supply chain and insurance. That is not a hyperscale data centre and nobody is claiming it is. What it demonstrates is the economic principle underneath all of this, which is simply that compute can be worth more than bandwidth.
Then it gets stranger
Extend that to a constellation capable of processing its own sensor data across Earth observation, weather, astronomy, navigation, space traffic management, communications, scientific simulation and defence sensing, and the shape of the system changes. Instead of a sensor sending data to Earth, a data centre producing a decision and the decision going back up, you get a sensor, onboard compute, a decision and an action, with latency falling from seconds or minutes to milliseconds.
IEEE Spectrum has pointed at exactly this with increasingly dense constellations: as orbital traffic grows, collision avoidance becomes an onboard computational problem rather than something that can reliably be closed through a ground loop. At sufficient scale the distinction between the communications network and the computer network stops being meaningful, and the network is the computer.
Which is where Starlink stops looking like a phone company
The easy mistake is to treat satellite communications and orbital compute as separate industries that happen to share an altitude.
A large communications constellation already supplies orbital nodes, power, inter-satellite links, ground stations, launch cadence, orbital operations, navigation, software-defined networking and automated fleet management. That is most of the hard part. Add substantial compute and the network starts becoming a distributed computational platform rather than a pipe. Starlink already uses laser links between satellites and its next generation increases network capacity considerably.
Follow that through and the eventual architecture looks less like a satellite operator selling connectivity and more like an orbital cloud provider. Your workload does not care which physical satellite executes it. It asks the network for compute, the constellation decides where to run it, the result moves through the optical mesh, and Earth becomes one endpoint of a much larger computational system rather than the centre of it.
The cooling problem is the honest objection
There is an enormous asterisk on all of this, and it deserves more attention than the enthusiasts usually give it.
Space is cold, but vacuum is a terrible heat sink. There is no air to carry heat away, so everything a processor generates has to be radiated, which is why Google is explicitly working on heat pipes and radiators. That produces a design constraint that inverts the terrestrial one. On Earth we obsess over power density and cooling plant. In orbit the chain runs from power density to heat generation to radiator area to mass to launch cost, and every watt you add has to be paid for twice.
The radiator becomes part of the computer. So does the solar array, the optical network and the orbital mechanics. This is not data centre engineering with a rocket on the front. It is systems engineering at a scale we have not really attempted before.
The long game, which is not about AI at all
If a substantial orbital industrial ecosystem does get built, we end up with power generation, computing, communications, robotics, manufacturing, servicing and transport all operating off-planet, and at that point the question quietly inverts. We stop asking why anyone would put industry in space and start asking which industries are foolish to keep on Earth.
That is a far more interesting question, and it has real answers. Some things stay here indefinitely. Energy-intensive computation is an obvious candidate to leave, and so eventually are some forms of manufacturing, materials processing, scientific instrumentation and resource extraction. The Moon stops being primarily a place to put astronauts and starts looking like an industrial resource base, and the Solar System stops being a list of destinations and becomes a list of locations.
This is probably how it actually starts
I do not think an interplanetary civilisation begins with Mars, and I am fairly sure it does not begin with anyone declaring that humanity must become multiplanetary.
I think it begins with an accountant. Somebody works out that electricity, land, cooling, regulation and transmission make a particular industrial process cheaper somewhere else, except that this time somewhere else is five hundred kilometres up. Then somebody notices that the communications infrastructure needed to run that facility also makes orbital computing cheaper, and that orbital computing justifies more launches, and that more launches bring launch costs down, and that cheaper launch justifies more infrastructure, which makes the next set of applications viable.
The economic boundary moves outward one decision at a time. Earth, then Earth and orbit, then Earth and orbital industry and the Moon, and eventually Earth as one node in a Solar System economy. Nobody has to announce any of it. It happens because the economics keep pointing the same way, and economics is remarkably persistent.
What is worth watching
None of this proves that orbital hyperscale compute will work, and I would not want the enthusiasm to carry more weight than the evidence does. The remaining problems are serious: radiation, thermal management, reliability, servicing, debris, communications, manufacturing, launch economics, and the plain fact that replacing a failed server is considerably harder when the server is in orbit than when it is in a rack you can walk to.
But none of those problems violates the laws of physics, which is a meaningfully different situation from most speculative infrastructure. Google is testing hardware, thermal systems and optical networking. Commercial operators are already putting computation in orbit. Launch costs are falling. Communications constellations are building the orbital networks. And the AI industry has generated a demand for electricity and computation large enough to make previously absurd concepts worth the cost of investigating.
So the story here is not AI in space, which is the version that makes headlines. It is that AI may end up providing the economic justification for building something humanity has never built: a permanently operating industrial network powered directly by the Sun, connected by lasers, increasingly autonomous, and no longer confined to the surface of one planet.
The first interplanetary civilisation almost certainly will not announce itself. It will look like a business decision, and we will only notice afterwards, when a meaningful fraction of our computation, energy collection, communications and manufacturing has stopped happening here. At that point we will not have started becoming a spacefaring civilisation. We will already be one.