The pitch is seductive in its simplicity: AI needs more power than terrestrial grids can supply, so move the data centres into orbit, where the sun never sets and the electricity is free. SpaceX, Blue Origin, and a growing constellation of startups are now racing to make that vision real. The problem, according to the scientists and engineers who would have to make the physics work, is that the vision skips several chapters of thermodynamics, economics, and orbital mechanics that have not yet been written.

SpaceX filed with the Federal Communications Commission on 30 January for permission to launch up to one million satellites into low Earth orbit, each carrying computing hardware that would collectively form what the company described as a constellation with “unprecedented computing capacity to power advanced artificial intelligence models.” The satellites would operate at altitudes between 500 and 2,000 kilometres, in orbits designed to maximise time in sunlight, and route traffic through SpaceX’s existing Starlink network. SpaceX requested a waiver of the FCC’s standard deployment milestones, which typically require half a constellation to be operational within six years.

Seven weeks later, Blue Origin filed its own application. Project Sunrise proposes 51,600 satellites in sun-synchronous orbits between 500 and 1,800 kilometres, complemented by the previously announced TeraWave constellation of 5,408 satellites providing ultra-high-speed optical backhaul. Where SpaceX’s filing emphasised raw scale, Blue Origin’s emphasised architecture: the system would perform computation in orbit and relay results to the ground through TeraWave’s mesh network.

The startup ecosystem is moving even faster. Starcloud, formerly Lumen Orbit, raised $170 million at a $1.1 billion valuation in March, becoming the fastest unicorn in Y Combinator history just 17 months after completing the programme. The company launched its first satellite carrying an Nvidia H100 GPU in November 2025 and filed with the FCC in February for a constellation of up to 88,000 satellites. Aethero, a defence-focused startup building space-grade computers with Nvidia Orin NX chips wrapped in radiation shielding, raised $8.4 million and is testing hardware on orbit this year.

The commercial logic rests on a genuine problem. Global data centre electricity consumption reached roughly 415 terawatt-hours in 2024 and the International Energy Agency projects it could exceed 1,000 TWh by 2026, with accelerated AI servers driving 30 per cent annual growth. In Virginia alone, data centres consume 26 per cent of total electricity supply. Ireland’s share could reach 32 per cent by year’s end. The grid constraints are real, the permitting delays are real, and the political resistance to building more terrestrial capacity is real.

What is also real, scientists argue, is the physics that makes orbital computing spectacularly difficult at any meaningful scale. The most fundamental challenge is heat. In space, there is no air to carry heat away from processors, only radiative cooling, which requires vast surface areas. Dissipating just one megawatt of thermal energy while keeping electronics at a stable 20 degrees Celsius demands approximately 1,200 square metres of radiator, roughly four tennis courts. A several-hundred-megawatt data centre, the minimum threshold for commercial relevance, would require radiators thousands of times larger than anything ever deployed on the International Space Station.

Radiation presents the second structural problem. Low Earth orbit exposes unshielded chips to cosmic rays and trapped particles that induce bit flips and permanent circuit damage. Radiation hardening adds 30 to 50 per cent to hardware costs and reduces performance by 20 to 30 per cent. The alternative, triple modular redundancy, means launching three copies of every chip, three times the cooling, three times the electricity, and three times the mass. Starcloud’s approach of flying commercial GPUs with external shielding is an interesting experiment, but no one has demonstrated that it works at scale or over hardware lifetimes measured in years rather than months.

Latency is the third constraint. A million satellites spread across orbital shells from 500 to 2,000 kilometres cannot achieve the tight coupling required for frontier model training, where inter-node communication latencies must remain in the microsecond range. Low Earth orbit introduces minimum latencies of several milliseconds for inter-satellite links and 60 to 190 milliseconds for ground-to-orbit round trips, compared to 10 to 50 milliseconds for terrestrial content delivery networks. That makes orbital infrastructure potentially viable for inference workloads, not for training, which is where the overwhelming majority of AI compute demand currently sits.

Then there is cost. IEEE Spectrum estimated that a one-gigawatt orbital data centre would cost upwards of $50 billion, roughly three times the cost of an equivalent terrestrial facility including five years of operation. Google has said that launch costs must fall to under $200 per kilogram before space-based computing begins to make economic sense. SpaceX’s current Starlink economics operate at roughly $1,000 to $2,000 per kilogram. Some analysts argue the true threshold for competing with terrestrial refresh economics is $20 to $30 per kilogram, a figure no credible projection places within the next two decades. The economics look even less favourable when set against the deep-tech funding landscape on the ground, where terrestrial infrastructure projects can draw on established supply chains and proven unit economics.

Even OpenAI’s Sam Altman, who explored a multibillion-dollar investment in rocket maker Stoke Space as a potential SpaceX competitor for orbital data centres, has publicly called the concept “ridiculous” for the current decade. Altman told journalists that the rough maths of launch costs relative to terrestrial power costs simply does not work yet, and he pointedly asked how anyone plans to fix a broken GPU in space.

The astronomical community adds a separate objection entirely. The vast majority of the roughly 1,000 public comments on SpaceX’s FCC filing urged the commission not to proceed. If approved, the constellation would place more satellites than visible stars in the sky for large portions of the night throughout the year, further militarising and commercialising an orbital environment that is already straining under the weight of existing megaconstellations.

None of this means orbital data centres will never exist. SpaceX’s Starship, if it achieves its cost targets, could fundamentally change the mass-to-orbit economics that currently make the concept unworkable. Starcloud’s incremental approach of flying small payloads and iterating on radiation performance is the kind of engineering pathway that occasionally produces breakthroughs. And the terrestrial grid constraints driving the interest are not going away.

But the gap between filing an FCC application for a million satellites and actually making orbital computation economically competitive with a warehouse full of GPUs in Iowa is not measured in years. It is measured in physics problems that the current pace of AI infrastructure investment cannot shortcut, no matter how many billionaires are willing to try. The question scientists are asking is not whether space data centres are theoretically possible. It is why, given the magnitude of the unsolved engineering, anyone is treating them as a near-term solution to a problem that requires near-term answers. The sky, it turns out, is not the limit. The radiator is.