Google’s Project Suncatcher

Datacentres are a rapidly growing share of global electricity consumption.

Artificial Intelligence workloads significantly increase power demand due to:

• Dense clusters of GPUs / specialised accelerators

• Energy-intensive training and deployment of large language models

Generative AI boom shows no signs of slowing, intensifying concerns over:

• Energy availability

• Sustainability

• Carbon footprint

Google Research is exploring space-based datacentres.

Concept:

• Place AI datacentres in Low Earth Orbit (LEO).

• Power them entirely using solar energy.

Motivation:

• Avoid terrestrial energy constraints.

• Bypass land, cooling, and grid limitations.

Indian Space Research Organisation is also reportedly studying space-based datacentre technology.

Demand driven mainly by content consumption (e.g., video streaming).

Bandwidth requirement:

• Roughly equal between internal systems and external users.

Require extremely high internal bandwidth, not user-facing bandwidth.

Large-scale distributed machine learning workloads demand:

• High-speed communication within and between nearby datacentres.

Example:

• Microsoft’s Fairwater AI datacentre complexes

• Use petabit-per-second links between facilities.

• Around one million times faster than typical consumer internet.

Similar logic applies in space:

• Most bandwidth used between satellites, not with Earth.

• Ground communication only needed for:

  • Input queries

  • Output responses

• Like ChatGPT infrastructure:

  • Massive internal bandwidth

  • Minimal user-facing bandwidth

Inspired by Starlink, but with key differences:

• Not a dispersed global swarm.

• Densely choreographed satellite clusters.

Key features:

• Satellites positioned a few kilometres apart.

• Orbit designed to maintain constant line-of-sight with the Sun.

• Heavy reliance on:

  • Inter-satellite links

  • Multiplexing (packing more data into a single radio beam)

Result:

• High computational coordination.

• Continuous solar power availability.

Major concern: solar and cosmic radiation damaging chips.

Findings from Google Research:

• High Bandwidth Memory (HBM) most radiation-sensitive component.

• Irregularities began after:

  • 2 krad(Si) cumulative dose.

• Expected five-year shielded dose:

  • ~750 rad(Si).

• No hard failures observed up to:

  • 15 krad(Si) on a single chip.

Conclusion:

• Trillium TPUs show unexpected radiation hardness for space use.

Datacentres require continuous servicing.

In orbit:

• No low-cost or rapid access for repairs.

• Faulty satellites may need replacement via fresh launches.

On Earth:

• Liquid cooling is standard.

In space:

• Constant solar exposure.

• Heat must be dissipated in a vacuum, where convection is impossible.

Requires:

• Advanced radiators and thermal control systems.

Space-based datacentres must compete with ground-based alternatives.

Costs include:

• Research and development

• Launching satellite clusters

• Replacing failed satellites

Google’s assumptions:

• Satellite launch costs may fall to $200 per kg by mid-2030s.

• Solar-powered architecture could yield major energy cost savings.

Business viability depends on:

• Whether space solutions outpace improvements in terrestrial datacentres.

Microsoft Natick project:

• Experimented with underwater datacentres for cooling.

• Ultimately abandoned despite technical promise.

Shows:

• Innovative datacentre ideas face economic and operational barriers.

Technological scepticism often ages poorly:

• Starlink’s success was widely doubted in 2019.

Space-based datacentres:

• High-risk, high-reward concept.

• Could reshape AI infrastructure if technical and economic hurdles are crossed.