Orbital Data Centres: The Future of AI in Space — A Timeline

Google’s Suncatcher prototype reaches orbit In orbit

Google said its first Project Suncatcher satellite, built with Earth-imaging company Planet, had launched on SpaceX’s Transporter-18 rideshare from California. The team confirmed contact and said the spacecraft was operating as expected. Over the following weeks it is gathering data on how Google’s Tensor Processing Units cope with launch, radiation and temperature swings.

Reports described a fridge-sized satellite with about a kilowatt of solar power. Google has said further satellites are planned for 2027, and published a peer-reviewed design paper in the journal Joule.

Rules and regulators start to catch up 2026

Orbital AI data centres were on the agenda of the 77th International Astronautical Congress in Antalya, Turkey, alongside crowded orbits and the race to the Moon. In the US, the Office of Space Commerce opened a pilot certification programme for novel in-space activities, including orbital data centres, with expressions of interest due by 5 October 2026.

Law firms and policy analysts are working through basic questions: which country’s data-protection law applies to a server in orbit, how such satellites are licensed and insured, and who is liable for debris.

BCG prices the gap: 2.5–3x more expensive Reality check

Boston Consulting Group’s Space-Based Data Centers: Cost and Outlook to 2040 estimated that an orbital data centre today costs 2.5 to 3 times as much over 20 years as one on the ground, narrowing to about 1.5x by 2030–35 in realistic scenarios. It called cooling the principal bottleneck: roughly 400 m² of radiator for a 100 kW satellite.

BCG still saw a niche worth $240–320 billion a year by 2040, 10–15% of the AI data-centre market, mainly latency-tolerant inference, processing of space-generated data and sovereign workloads.

Laser links and studies multiply Plan

Starcloud agreed to fit SpaceX’s Starlink mini laser terminals on its next 25 satellites, so they can pass data through the Starlink network instead of waiting for a ground station. In Europe, the European Space Agency awarded Edge Aerospace a contract to study orbital data-centre architectures and a roadmap.

The US GAO weighs the idea Reality check

The US Government Accountability Office published Science & Tech Spotlight: Data Centers in Space. It cited a Department of Energy projection that data centres could use up to 12% of US electricity by 2028, driven by AI, and said moving some computing to orbit could ease demand for power, water and land and speed up processing of data collected in space.

It listed the obstacles plainly: cooling in a vacuum, radiation damage that could force more frequent replacement, launch and manufacturing costs, collision risks including for crewed missions, and interference with astronomy. It asked whether launch capacity, space law and research funding are ready.

NVIDIA builds chips for space; Starcloud becomes a unicorn Plan

At its GTC conference NVIDIA announced a space computing line-up headed by the Space-1 Vera Rubin Module, which it says offers up to 25 times the AI inference compute of the H100, alongside the smaller IGX Thor and Jetson Orin. It named Axiom Space, Kepler, Planet, Starcloud, Sophia Space and Aetherflux as users of its platforms.

Two weeks later Starcloud raised $170 million at a $1.1 billion valuation to build Starcloud-2, designed to carry an NVIDIA Blackwell GPU, an AWS server blade and bitcoin-mining chips, with about 100 times Starcloud-1’s power and a large deployable radiator.

The mega-filings: up to one million satellites Filing

SpaceX asked the US Federal Communications Commission to authorise up to one million satellites between 500 and 2,000 km to act as solar-powered orbital data centres, saying they were “the most efficient way to meet the accelerating demand for AI computing power”. Days later SpaceX confirmed it had acquired Elon Musk’s AI company xAI, in a deal valuing the combined company at about $1.25 trillion.

Blue Origin announced TeraWave, a network of about 5,400 satellites for data-centre connectivity, and Starcloud filed for a constellation of up to 88,000 satellites.

Axiom’s first orbital data-centre nodes launch In orbit

Axiom Space said its first two orbital data-centre nodes had reached low Earth orbit, flying as part of Kepler Communications’ optical relay constellation. Their laser links, compatible with the US Space Development Agency’s Tranche 1 standard, let other satellites send data to the nodes for processing without routing it through the ground.

Axiom uses commercial off-the-shelf hardware and standard container software, and says it aims to grow from kilowatts to megawatts of capacity over the coming years.

The first AI model trained in orbit In orbit

Starcloud said its satellite had trained NanoGPT, a small language model by Andrej Karpathy, on the complete works of Shakespeare, the first training of a language model in space, and had run and answered queries with Google’s open Gemma model. Starcloud describes this as running ‘a version of Gemini’.

Starcloud-1 puts an NVIDIA H100 in orbit In orbit

Starcloud, a start-up founded in 2024 as Lumen Orbit, launched Starcloud-1 carrying an NVIDIA H100, a data-centre AI accelerator about 100 times more powerful than any GPU previously flown, according to the company. The satellite flies along the boundary between day and night so its solar panels stay lit.

Getting a chip designed for air-conditioned server halls to work in vacuum, radiation and extreme temperature swings was the real test.

Google announces Project Suncatcher Plan

Google unveiled Project Suncatcher, a research moonshot asking whether constellations of solar-powered satellites carrying its TPUs could scale machine learning in space. It said its Trillium TPUs survived radiation testing well, and modelled clusters of 81 satellites flying within about a kilometre of each other, linked by lasers.

Google’s analysis suggested the economics could work if launch prices fall below about $200 per kg by the mid-2030s. It planned two prototype satellites with Planet by early 2027.

Axiom tests AxDCU-1 on the ISS In orbit

Axiom Space installed AxDCU-1, a prototype data-processing unit running Red Hat Device Edge, on the International Space Station in 2025 to test cloud services, AI and machine learning, data fusion and space cybersecurity.

The ISS is a forgiving test bed: the station supplies power, cooling and communications. A free-flying orbital data centre must provide all of those itself.

China launches the Three-Body Computing Constellation In orbit

China launched the first 12 satellites of the Three-Body Computing Constellation, billed by ADA Space as the world’s first dedicated orbital computing constellation. Together they offer about 5 peta-operations per second, 30 TB of storage and laser links of up to 100 Gbps, and process remote-sensing data in orbit.

The wider Star-Compute programme plans about 2,800 satellites and 1,000 POPS.

A data centre goes to the Moon In orbit

Lonestar Data Holdings flew its small ‘Freedom’ data-storage payload on Intuitive Machines’ IM-2 lunar mission. The company said the payload met its tests in cislunar space and lunar orbit, with stable power, temperature and memory readings, but the Athena lander tipped over after touching down far from its target in March, cutting the mission short.

From papers to start-ups Plan

Starcloud was founded in January 2024 and published a white paper proposing gigawatt-scale orbital data centres built from modular solar and radiator units. In Europe, the ASCEND feasibility study led by Thales Alenia Space for the European Commission examined whether orbital data centres could cut the carbon footprint of computing, concluding that it would depend on much cleaner, cheaper launchers.

Off-the-shelf servers survive the ISS In orbit

Hewlett Packard Enterprise sent the Spaceborne Computer, built from commercial off-the-shelf servers rather than radiation-hardened parts, to the ISS on SpaceX’s CRS-12 mission. It ran for more than a year, reached about one teraflop, and relied on software to manage radiation errors. It returned to Earth in 2019 and a second version followed in 2021.

Satellites learn to think for themselves Plan

Satellites have always carried computers to run the spacecraft and its instruments, usually slow, radiation-hardened chips. As sensors grew more capable, operators began filtering and compressing data on board to save scarce downlink bandwidth.