Google AI chips in space will move from concept to hardware on 1 October, when Project Suncatcher is due to send four TPU processors into orbit. The experiment is tiny compared with the orbital data centres companies have discussed, but it is meant to answer a more basic question first: can modern AI hardware survive and work reliably up there?
The test is being carried out with Planet and is scheduled to fly on SpaceX’s Transporter-18 mission. Google says the hardware will be exposed to the radiation, thermal swings and vacuum conditions that future space-based computing systems would have to tolerate.
This is not an operational data centre in orbit. It is closer to putting a small piece of one through a real space environment before anybody tries to build the ambitious version.
What Project Suncatcher is testing
Google’s long-term idea is to place computing hardware where solar energy is abundant and connect orbital systems with high-speed optical links. The near-term test is much less grand. Engineers want to see how the TPU hardware behaves under launch vibration, radiation and temperature conditions that cannot be reproduced perfectly on the ground.
Ars Technica reports that the first spacecraft carries four TPUs in a small prototype system and that the chips will run AI workloads in bursts while the satellite manages the challenge of removing heat in space.
Heat is easy to overlook because space is cold in popular imagination. Electronic components still generate heat, and vacuum removes the convenient option of cooling them with moving air.
Why put AI computing in orbit at all?
The attraction is energy. Data centres on Earth need electricity, grid connections, land, cooling systems and often large volumes of supporting infrastructure. In suitable orbit, solar panels can receive long periods of intense sunlight without local clouds or night-time weather interrupting generation.
That does not make space computing automatically greener or cheaper. Rockets, satellites, maintenance, communications and replacement hardware all carry costs. A data centre that fails on Earth can be repaired by technicians. A failed orbital computer is much harder to reach.
Project Suncatcher is therefore testing whether the physical disadvantages can be reduced enough for the energy advantages to matter.
Radiation is one of the first obstacles
High-energy particles can disrupt electronic components and gradually damage them. Google says ground tests of its Trillium TPUs exposed the chips to radiation doses intended to represent more than five years in the target environment.
Ground qualification is necessary, but it cannot replace orbital experience. Launch itself introduces severe vibration, and thermal behaviour can change once hardware is operating in vacuum. The 1 October mission gives engineers real data rather than another simulation.
If the chips perform poorly, that is useful information too. The purpose of an early demonstrator is to find the uncomfortable engineering problems while the system is still small.
The vision is much bigger than four chips
Google has previously described a future in which groups of satellites carry larger numbers of processors and communicate using optical links. That architecture would have to move data quickly enough that separate spacecraft could behave more like parts of one computing system.
It would also have to solve deployment and replacement at a scale very different from one demonstration satellite. The cost of launch has fallen dramatically over the past decade, but sending tonnes of computing hardware into orbit remains a different proposition from delivering racks to a terrestrial warehouse.
LiveAIWire has examined the wider push towards orbital data centres. Suncatcher is useful because it turns that speculative industry idea into a specific piece of hardware that can now fail or succeed in public.
Earth-bound data centres are creating pressure for alternatives
The experiment arrives as AI companies face scrutiny over the power and water demands of terrestrial computing. LiveAIWire recently reported on California rules requiring more disclosure around data-centre resource use.
Space does not remove resource consumption. It moves some of the energy generation and cooling problem while adding launch, manufacturing and orbital-infrastructure problems. A credible environmental comparison will eventually need the full life cycle rather than a dramatic image of solar panels above the atmosphere.
For now, four chips are enough. Before anybody argues about thousands of orbital accelerators, Google first has to show that a handful can survive launch, radiation and heat while doing useful computing.
A small launch with a very large question behind it
The 1 October mission is interesting because the physical event is modest and the implication is enormous. Four processors will not transform the AI industry. They will test whether one of the stranger proposed answers to AI’s infrastructure problem deserves another round of investment.
If the hardware performs well, the next steps will still require optical networking, larger power systems, reliable thermal design and a convincing economic case. If it performs badly, engineers will learn which assumptions need to change.
Either way, Project Suncatcher is moving orbital AI computing out of slide decks and into an environment where physics gets the final vote.
What Google AI chips in space still have to prove
The orbital test has to show more than survival. Engineers need predictable computing performance, manageable heat and reliable communications. A chip that works for a short burst is encouraging, but a future orbital data centre would have to operate repeatedly and coordinate with other hardware for years.
That is why the four-chip mission is useful. It gives Google a comparatively small place to discover which part of the idea is hardest before the company attempts anything resembling a commercial system.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.