24 April 2026
Data centres in space: Can they truly replace traditional data centres?
The race to build ever larger AI infrastructure is increasingly constrained by power grid capacity and the challenge of cooling ever denser computing resources. SpaceX and others are pursuing a radical approach to address this problem: launching data centres into orbit. This idea comes with several advantages – near constant access to solar energy, and the avoidance of terrestrial land, water, and grid constraints. To assess whether these advantages can realistically translate into a viable data‑centre architecture, it is necessary to ask two questions: are space‑based data centres technically feasible, and could they ever be economically competitive?
Space-based data centres are physically possible at a limited scale
At first glance, the physics are not prohibitive. Low Earth Orbit (LEO) offers abundant solar energy, and satellites placed in sun‑synchronous orbits can achieve near‑constant illumination. This makes orbital solar power particularly attractive for computing systems that would otherwise require immense terrestrial power infrastructure.
A 100kW AI payload, as proposed in SpaceX’s ‘AI Sat Mini’ concept1, would require a large but achievable solar array. At 1AU from the Sun, there is a solar power flux of 1381W/m². A solar panel operating at 30% efficiency would provide an output power per unit area of 400W/m², so at least 241m² of solar panels would be required per satellite. Although substantial, this is not beyond the capabilities of modern spacecraft engineering, and supplying the required electrical power to an orbital compute node would not be physically prohibitive.
Another, and much discussed, technical challenge is thermal management. Terrestrial data centres dissipate heat through air or liquid cooling, whereas space-based data centres must rely on radiative cooling to balance heat absorbed from the Sun, reflected radiation from Earth’s albedo, and from its own electronics against heat emitted to space. Due to the Stefan-Boltzmann law, the radiated power per unit area scales with the fourth power of the temperature, so the required radiator area will decrease significantly as the operating temperature of the satellite increases – however, this is constrained by how hot the silicon can run. For example, using a simplified thermal-equilibrium model, at an orbital altitude of 600km, we estimate that a 100kW satellite could be cooled with a pair of radiator surfaces each with an area of 113m² at 50°C, falling to 54m² at 100°C. While significant, these radiator surfaces would still be smaller than the required solar arrays. Radiative cooling therefore poses a design constraint, but not a fundamental physical barrier.
Assuming a launch cadence comparable to Falcon 9 Starlink missions, approximately 1 launch every 3 days2, and the use of Starship deploying 60 satellites per launch (as planned for Starlink V3 satellites3), a constellation of 36 000 satellites could be sustained with a 5-year technology refresh cycle. At 100kW per satellite, this would correspond to a total computing power of 3.6GW, which is substantial but below the 10GW terrestrial data centres being planned.
By contrast deploying a fleet of one million satellites, as proposed by SpaceX4, presents a far more ambitious engineering challenge. Maintaining the same 5-year refresh rate, 200 000 satellites would have to be put into orbit every year, requiring a launch cadence of around 10 launches per day. If such a constellation size were to be achieved with megawatt powered satellites, an orbital datacentre with a terawatt of computing power could be achieved. Moreover, a fleet of one million satellites would exceed the current number of functioning satellites in orbit by around two orders of magnitude5. At this scale, orbital congestion, collision avoidance, and space-debris mitigation become significant constraints. High object densities in LEO would significantly increase the likelihood of satellite collisions and the risk of debris cascades associated with the Kessler effect6.
The technical conclusion is therefore nuanced: Space-based data centres appear physically possible at the scale of individual satellites or low-GW constellations, if Starship launches can reach the current Falcon 9 launch cadence. However, a one million satellite architecture is significantly more ambitious and would have to navigate additional constraints such as space traffic management and orbital sustainability and reach a launch cadence that has not yet been approached.
Data centres in space could be cost‑effective relative to terrestrial alternatives
The economics of space-based computing are often dismissed as speculative, yet closer inspection suggests a more competitive picture. As AI-driven demand accelerates, terrestrial data centres are being pushed to unprecedented scales, with capital intensity, power availability, and grid constraints emerging as significant constraints. This accelerated growth in scale is exemplified by initiatives such as OpenAI’s Stargate Project, which signals a new class of large scale AI infrastructure designed explicitly around energy and compute optimisation. Within this context, SoftBank’s proposed 10GW AI campus in Ohio illustrates both the extraordinary scale of next generation AI infrastructure and the growing economic strain imposed on terrestrial energy systems7. Run continuously for a year, a 10GW data centre would consume just under the amount of power consumed by the City of London in 2023 (87.6TWh vs. 93.7TWh8)!
Our high-level modelling compares the cost of delivering 10GW of compute capacity for a period of 10 years via a terrestrial facility – including upfront hardware investment9 and ongoing power costs – with an alternative based on satellite data centres. Within this framing, space-based architectures merit serious consideration as a potential pathway to circumvent terrestrial bottlenecks rather than merely an experimental fringe.
In comparison to our calculated figures, the Stargate Project is reported to include an initial investment of USD100bn, with USD500bn spent over the next four years14.
What other issues might data centres in space encounter?
Beyond engineering and cost challenges, space‑based data centres would face important performance, operational, and regulatory constraints.
Latency is one such constraint. Latency overheads are added by round‑trip delays to low Earth orbit, at around 5ms for a simple ‘bent-pipe’ systems and multiple 10s of milliseconds for multiple hops. While this increased latency may be acceptable for delay‑tolerant tasks such as large‑scale AI training, it would be unsuitable for latency‑sensitive applications.
Data sovereignty and regulatory compliance present another complex challenge. Many jurisdictions impose strict requirements on where data may be stored and processed, and current legal frameworks are largely terrestrial in nature. Determining which national laws apply to data processed in orbit would require new regulatory agreements. Until such frameworks emerge, data sensitivity may significantly limit the data centre workloads that can be migrated to space.
Conclusions
Although the concept of an orbital data centre appears both physically and economically credible, it is unlikely that data centres will migrate en masse into orbit, primarily due to orbital congestion and management challenges at scale. While orbital computing represents an intriguing and potentially valuable complement to terrestrial infrastructure, it is unlikely to replace conventional data centres in the near future. In the long-term, perhaps the bottleneck may become terrestrial, rather than space-based: rather in managing large constellations or fast launch cadences, over finding enough land and resources on Earth to host the data centres.
[1] Space News, ‘SpaceX offers details on orbital data center satellites’, 22 March 2026
[2] https://www.spacex.com/launches
[3] https://x.com/SpaceX/status/1977873370688700846
[4] Space News, ‘SpaceX files plans for million-satellite orbital data center constellation’, 31 January 2026
[5] ESA, ‘Space Debris User Portal’, 16 January 2026
[6] IEEE Spectrum, ‘Have we reached a space-junk tipping point?’, 30 September 2025
[7] AP News, ‘DOE unveils 10-gigawatt Ohio data center, gas-powered energy plan’, 20 March 2026
[8] Greater London Authority, ‘London Energy and Greenhouse Gas Inventory (LEGGI)’, LEGGI 2023
[9] We note that to calculate the upfront costs for a 10GW terrestrial data centre, we have used historical benchmark costs and assumed that costs scale with power consumption. Therefore, this does not consider any cost savings from deploying such a large-scale facility.
[10] https://www.jll.com/en-us/insights/market-outlook/data-center-outlook
[11] https://www.jll.com/en-us/insights/market-outlook/data-center-outlook
[12] https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_3
[13] https://thenetworkinstallers.com/blog/data-center-energy-consumption-statistics/
[14] https://group.softbank/en/news/press/20250122
[15] https://www.tomshardware.com/service-providers/network-providers/spacex-shows-off-massive-new-v3-starlink-satellites-expanded-technology-will-deliver-gigabit-internet-to-customers-for-the-first-time-and-enable-60-tera-bits-per-second-downlink-capacity
[16] https://ntrs.nasa.gov/api/citations/20200001093/downloads/20200001093.pdf
