A typical AI-focused data centre already sits closer to heavy industry than to the old image of a quiet server room. The International Energy Agency put the comparison plainly in its 2025 Energy and AI report: one such facility can consume as much electricity as 100,000 households, while the largest under construction can use around 20 times that amount.

That single fact explains why a once eccentric idea is now being discussed in boardrooms, research labs and regulatory filings: moving some computing infrastructure into orbit.

The argument is simple at first glance. A satellite in the right orbit can see near-continuous sunlight. It does not need to buy electricity from a strained regional grid. It does not need a local water supply for evaporative cooling. It does not compete with towns for land, substations or transmission lines.

But that is only the sales version of the idea. The engineering version is less forgiving. In orbit, the power bill may disappear from the monthly accounts, but the cost has not gone away. It has been moved into solar arrays, radiators, launch mass, radiation protection, communications, replacement hardware and orbital risk.

The result is worth taking seriously, but it should not be read as the final word. Orbital data centres are still an early-stage proposal, with a few prototype efforts and a growing stack of feasibility papers, not a mature alternative to terrestrial computing.

Why the idea is suddenly being taken seriously

The pressure on ground-based data centres is no longer abstract. The IEA estimated that data centres used around 415 terawatt-hours of electricity in 2024, about 1.5 percent of global electricity consumption, and projected that figure could more than double to around 945 terawatt-hours by 2030. In the United States, the agency said data centres account for nearly half of projected electricity demand growth between now and 2030.

Those numbers do not mean AI is the only reason electricity demand is rising. Air conditioning, industry and electric vehicles also matter. But AI-focused facilities create a distinctive planning problem because they concentrate large power loads in particular places. A 100-megawatt facility is not just another office building. It is a permanent industrial customer that wants high reliability, dense power delivery and cooling capacity from the moment it switches on.

That is why space has become tempting. Google Research described its Project Suncatcher concept in November 2025 as a possible long-term way to build machine-learning infrastructure from solar-powered satellites linked by free-space optical communications. The company’s researchers proposed a dawn-dusk, sun-synchronous low-Earth orbit, where satellites could receive near-continuous sunlight while keeping latency lower than higher orbits.

Google is not alone. Associated Press reporting in 2026 described SpaceX and xAI plans for solar-powered orbital data centres, while also noting Starcloud, Google and Blue Origin as part of the wider field of companies exploring related infrastructure. Some proposals are still closer to ambition than hardware, but the direction is clear: AI compute demand has become large enough that space is no longer being treated only as a science-fiction backdrop.

The Sun is not the whole system

The most persuasive part of the orbital argument is energy supply. Solar panels above most of the atmosphere receive sunlight more consistently than panels on Earth. They avoid clouds, weather, night cycles and the land-use constraints that shape terrestrial solar farms.

Google’s Suncatcher paper argues that panels in certain orbits can receive much more yearly solar energy than a mid-latitude panel on Earth, and that a constellation could convert that energy directly into computation instead of trying to beam power down to the surface. That is an important distinction. Space-based solar power has long struggled with the problem of transmitting electricity back to Earth. Orbital computing sidesteps that by moving the load to the power source.

The difficulty is that computing does not only need power. It also produces heat.

On Earth, data centres rely on air, water, chilled liquid systems and large mechanical infrastructure to carry heat away from chips. In space, there is no air to convect heat away. Vacuum is useful for some things, but it is not a magic cold bath. Heat has to be moved through the spacecraft and then radiated away as infrared energy from panels.

That turns radiator area into one of the central constraints. A 2026 paper by Slava G. Turyshev, Orbital Data Centers: Spacecraft Constraints and Economic Viability, modelled a representative one-megawatt high-sunlight system and found that it would need about 5,640 square metres of photovoltaic area and about 2,500 square metres of radiator area before other spacecraft mass was included.

Those are not decorative panels. They are core infrastructure. Every square metre has to survive launch, deployment, micrometeoroids, thermal cycling, radiation and orbital debris. Every kilogram has to be paid for before it can do useful work.

Launch costs decide whether the maths closes

This is where the utility-bill argument meets the launch manifest.

Google’s analysis suggested that if launch costs to low-Earth orbit fell below about $200 per kilogram by the mid-2030s, the cost of launching and operating a space-based data centre could become roughly comparable to reported terrestrial data-centre energy costs on a per-kilowatt-year basis. That is a conditional statement, not a present-day price.

Turyshev’s analysis was more cautious. For a system around 40 kilograms per kilowatt, he found that a terrestrial infrastructure benchmark of $10,000 to $40,000 per kilowatt leaves only about $250 to $1,000 per kilogram for the combined cost of launch and spacecraft build, before communications, operations, utilisation and lifetime penalties are included. The paper notes that this allowance is already below the public Falcon 9 dedicated low-Earth-orbit launch-price benchmark, even before the spacecraft itself is built.

That is why the phrase “free solar power” can mislead. Sunlight in orbit is free. The machine required to turn it into reliable compute is not.

At even modest scales, the hardware bill is measured in millions. A one-megawatt orbital system would require tonnes of solar arrays, radiators, batteries or storage margins, structural hardware, computers, shielding, radios, optical links and propulsion. If the architecture grows toward the tens or hundreds of megawatts associated with serious AI workloads, the question becomes less about whether sunlight exists and more about whether the orbital supply chain can build, launch, replace and coordinate enough hardware cheaply enough.

The bottleneck may be getting data up and down

Power is only one constraint. Communication may be just as hard.

A 2026 review by Minghao Sun, Zehui Chen, Jinbo Hou, Kezhi Wang and Xiaoli Chu, Toward Communication-Efficient Space Data Centers, argues that space data centres are not limited in the same way as ground facilities. Ground facilities are often constrained by power, land and site availability. Orbital systems are constrained by the capacity of links between Earth, satellites and other satellites.

The paper points to a mismatch between the enormous internal data exchange rates of terrestrial data centres and the much lower capacity of practical space-to-ground links. That matters because many AI workloads are not tidy, isolated calculations. Training, fine-tuning and large-scale inference can involve huge flows of data, model weights, updates and results. If the work requires constant movement of raw data between Earth and orbit, the network can become the limiter.

That is why many near-term concepts focus on workloads that make sense in space: processing satellite imagery before downlink, analysing sensor data near where it is collected, or running tasks that tolerate delay. General cloud-style computing for ordinary terrestrial users is a harder target.

Cooling, maintenance and debris are not footnotes

The orbital case also has a reliability problem. On Earth, a failed server can be replaced. A cooling pump can be serviced. A rack can be upgraded when a better chip arrives. In orbit, hardware is exposed to radiation, energetic particles, thermal cycling and debris, and routine maintenance remains far from ordinary.

Google’s paper reports promising radiation tests for its Trillium TPUs, but it still identifies thermal management, high-bandwidth ground communication, on-orbit reliability and repair as major future challenges. The point is not that the hardware cannot work. It is that a data centre is not a single chip surviving a test beam. It is an entire industrial system remaining useful over years.

There is also the sky itself. Large constellations add objects to already crowded orbital regimes. They require collision avoidance, end-of-life disposal and coordination with astronomy, satellite operators and regulators. The larger the proposed constellation, the less convincing it is to treat debris and visibility as side issues.

Some researchers are exploring different physical architectures. A 2025 University of Pennsylvania-led preprint proposed a tether-based design for solar-powered orbital AI data centres operating in dawn-dusk sun-synchronous orbits, with photovoltaic panels, radiative cooling and integrated shielding arranged to provide multi-megawatt computing. That kind of work is useful because it turns the idea from slogan into mass budgets, thermal surfaces and orbital dynamics.

It also shows why the problem is difficult. Space offers sunlight, but it removes many ordinary conveniences. There is no cheap repair crew, no easy cooling tower, no nearby substation and no simple way to add another building next door.

A possible layer, not a replacement planet

The most plausible early role for orbital data centres may not be replacing Earth’s data centres. It may be adding a specialised layer for tasks that already belong near space: Earth observation, satellite coordination, disaster response, communications routing, remote sensing and delay-tolerant AI workloads.

That is a narrower claim than the grand version of the idea, but it is also more believable. If data is collected in orbit, doing some of the computation there can reduce downlink pressure. If a satellite network needs rapid coordination, local processing may make operational sense. If launch costs fall and optical links improve, the range of useful tasks could expand.

What the concept does not do is abolish cost. It changes the shape of it.

On Earth, the visible problem is electricity: grids, substations, cooling water, land, permits and local opposition. In orbit, the visible problem becomes mass: solar arrays, radiators, shielding, communications, reliability, replacement and the continuing cost of access to space.

That makes orbital computing neither fantasy nor easy answer. It is a serious response to a serious energy bottleneck, but one that has to pass through the unforgiving accounting of spacecraft design. The Sun may be available above the atmosphere almost all the time. Everything needed to turn that sunlight into dependable AI compute still has to be built, launched, cooled and kept alive.

Sources