The Case for Orbital Data Centres

The critics of orbital data centres are focused on what goes up and comes back down. That is the wrong frame. The right frame is what stays on the ground if we do not build them: more land stripped for server farms, more rivers drained for cooling towers, more grids strained past breaking point to feed an AI buildout that is not slowing down. The environmental argument against putting compute in space falls apart the moment you compare it to the environmental cost of the alternative.

"The question is not whether orbital compute has a footprint. Everything has a footprint. The question is whether it is smaller than the footprint of not doing it."

---

1. The Physics Is Not Optional

The sun delivers 1,361 watts per square metre to a surface in low Earth orbit. No clouds. No night in a sun-synchronous orbit. No atmosphere absorbing 30% of the energy before it reaches a panel. On the ground, a solar farm in a good location produces an average of 150 to 200 watts per square metre after accounting for weather, angle, and the rotation of the Earth. In orbit, you get 1,361 watts continuously. That is not a marginal improvement. It is a structural one.

Modern multi-junction gallium arsenide cells deployed in space convert roughly 29 to 30% of incident radiation, producing around 400 watts per square metre of panel in direct sunlight. SpaceX's AI1 satellite, revealed in June 2026, is targeting 150 kilowatts of compute power from a 70-metre wingspan solar array. That is 150 kW from a platform that fits in a rocket fairing. A terrestrial data centre achieving the same power output from rooftop solar would need roughly 750 to 1,000 square metres of panels and a very sunny day.

The energy problem for AI is real and it is not going away. Training a frontier model already costs tens of millions of dollars in electricity. The next generation of models will cost more. The grid in most developed countries is not growing fast enough to absorb that demand. Orbital compute does not ask permission from a utility company. It generates its own power from a source that has been running for five billion years and shows no signs of stopping.

---

2. No Land. No Water. No Neighbours.

A large terrestrial hyperscale data centre consumes between three and five million gallons of water per day for cooling. In regions already under water stress, that is not a neutral act. The facilities also require hundreds of acres of land, years of permitting, and the sustained opposition of every local planning board, environmental group, and property owner within range. Then they need a power line, which requires its own permitting battle.

An orbital data centre uses none of this. The vacuum of space is a 2.7 Kelvin heat sink. Heat rejection happens via radiator panels facing deep space, no water involved, no river drawn down, no aquifer tapped. The land footprint is a launch pad used for an hour. The planning permission is an FCC filing.

This is not a small advantage. It is the difference between a technology that can scale and one that cannot. The terrestrial data centre buildout is already running into hard limits: power availability in Northern Virginia, water rights in Arizona, grid capacity in Ireland. These are not problems that engineering can solve. They are political and geographic constraints that orbital compute sidesteps entirely.

---

3. The Environmental Footprint Is Genuinely Small

Opponents of orbital data centres point to the atmospheric effects of satellite reentry: aluminium oxide nanoparticles, trace metals, carbon soot from rocket launches. These are real phenomena and worth monitoring. But the numbers need context.

At full buildout of every constellation currently filed with the FCC, roughly 1.24 million satellites, the combined solar panel area is approximately 744 square kilometres. Against Earth's 127.5 million square kilometre sunlight interception disk, that is 0.00058%. Six parts per million. The threshold where atmospheric scientists begin to care about sunlight reduction is around 0.1%, a level associated with major volcanic eruptions. Orbital compute at full scale does not get within two orders of magnitude of that threshold.

The reentry chemistry story is more nuanced. Aluminium oxide nanoparticles from vaporising satellites have been detected in the stratosphere and are worth ongoing research. At the scale of currently approved constellations, models suggest alumina deposition roughly 18 times the natural meteoric baseline. That sounds large until you consider that natural meteoric input is 15,000 to 40,000 tonnes per year of total material, and the ozone layer has survived billions of years of it. The peer-reviewed uncertainty range on ozone depletion from satellite alumina spans from negligible to 9%. It could be nothing. It could be something. It is not known, and that means it warrants study, not a moratorium on a technology that has not yet been built.

Compare this to what terrestrial AI infrastructure actually does to the environment right now, with certainty:

• A single large data centre emits between 100,000 and 500,000 tonnes of CO2 equivalent per year, depending on the grid it draws from
• The global data centre sector consumes roughly 200 to 250 terawatt-hours of electricity annually, a number growing at 15 to 20% per year
• Cooling water withdrawals from US data centres are estimated at over 600 billion litres per year
• The land cleared for a single hyperscale campus runs to hundreds of acres, often on agricultural or ecologically sensitive sites

The orbital alternative produces its own power from sunlight, uses no water, clears no land, and injects trace quantities of material into an atmosphere that has been absorbing meteoric debris at far greater rates for geological time. The environmental case is not against orbital compute. It is for it.

---

4. Starlink Already Proved the Model

Before Starlink, satellite internet was slow, expensive, and available to almost nobody outside a niche industrial market. The prevailing expert consensus was that megaconstellations were economically unviable and the physics of low-latency broadband from orbit could not compete with fibre.

That consensus was wrong. Starlink now serves tens of millions of users across more than 100 countries. It provided connectivity to Ukraine when Russian forces destroyed terrestrial infrastructure. It brought broadband to rural communities in Canada, Australia, and sub-Saharan Africa that had waited decades for a terrestrial provider that never came. It drove launch costs down by roughly an order of magnitude through reusable rocket development that the entire space industry now benefits from.

The people who said it could not be done were not lying. They were reasoning from the technology and economics of the previous generation. That is what incumbents always do, and it is always wrong about the next generation.

---

5. The Geopolitical Stakes Are Not Abstract

China has its own orbital infrastructure ambitions. The Guowang constellation is a state-backed programme with filed capacity for 13,000 satellites in LEO. China's state-owned space sector is not subject to the environmental objections, public planning processes, or competitive pressures that shape Western commercial space. They are building because they understand that whoever controls the compute controls the AI, and whoever controls the AI shapes the next century of economic and military power.

The argument that the West should slow down orbital compute development for atmospheric reasons while China accelerates is not an environmental argument. It is a unilateral disarmament argument dressed in scientific language. The atmosphere does not care which country's satellites are burning up in it. If the choice is between Western commercial operators developing orbital compute with private capital and genuine competitive pressure to do it efficiently, or ceding that infrastructure to state actors with no such pressure, the environmental calculus strongly favours the former.

---

6. Critics Are Always Wrong About This

Steam engines were going to ruin the countryside. The internal combustion engine was going to poison the cities. Aviation was going to be too dangerous and too expensive to matter. Nuclear power was going to make large parts of the planet uninhabitable. The internet was going to destroy attention spans, privacy, and civil discourse.

Some of those predictions had elements of truth. Cars did change cities. The internet did change attention spans. But the technologies were built anyway, because the benefits were real and the alternatives were worse, and because the people making the predictions consistently underestimated the capacity for the technology itself to evolve in response to the problems it created.

Orbital data centres are early. The reentry chemistry is understudied. The engineering is not fully solved. The economics are unproven at scale. All of that is true and none of it is an argument against building. It is an argument for building with rigorous monitoring, open data on atmospheric impacts, and the regulatory framework to adjust course if the science demands it. That is how every transformative infrastructure technology has ever been developed. There is no reason to expect this one to be different.

---

7. The Vision Is Worth Taking Seriously

SpaceX's ambition with AI1 and the broader orbital compute programme is not primarily about making money, though it will. It is about removing the fundamental constraint on what compute can do. The history of computing is a history of removing constraints: vacuum tubes to transistors, mainframes to personal computers, spinning disks to solid state, on-premise to cloud. Each transition looked unnecessary or impossible or dangerous to the people invested in the previous paradigm. Each transition made the previous one look quaint.

Orbital compute removes the last constraint that ground cannot solve: energy at scale, unconstrained by land, water, permitting, or grid politics. If it works, the consequence is not a marginal improvement in data centre economics. It is a step-change in the amount of compute the human species can run, which means a step-change in what AI can do, which means a step-change in the rate at which every other hard problem gets addressed.

The atmosphere deserves monitoring. The reentry chemistry deserves study. But the vision deserves serious engagement rather than reflexive opposition from people whose mental model of the risk was formed before the first AI1 design was published. The physics is right. The need is real. The alternative is worse.