AI Publications

🔗https://shre.ink/Google-Space-AI-Infrastructure

1. Introduction

Google frames AI as a foundational technology that will reshape many domains — science, business, society — and asks: where can we go to unlock its fullest potential?. Their answer: perhaps the best place to scale AI compute is space.

They point out that the Sun emits more power than 100 trillion times humanity’s total electricity production, and that in certain orbits a solar panel can receive up to eight times more annual energy than on Earth. With this in mind, their “moonshot” initiative, called Project Suncatcher, envisions constellations of solar-powered satellites, each carrying high-performance TPUs (Tensor Processing Units), connected by free-space optical links — in effect, a data-centre in orbit.

2. System design and key challenges

2.1 Achieving data centre-scale inter-satellite links

One major technical hurdle is enabling high-bandwidth, low-latency communication between satellites, to replicate the kind of connectivity seen in terrestrial data centres. Google’s analysis says links supporting tens of terabits per second will be required.

To achieve that, they suggest dense wavelength-division multiplexing (DWDM) transceivers plus spatial multiplexing. But because received power drops with the square of distance, one key design approach is to fly satellites in very tight formation (kilometres or less apart) so the link budget becomes feasible.

2.2 Controlling large, tightly-clustered satellite formations

Flying satellites so close together (hundreds of metres apart) in low Earth orbit (LEO) is not trivial. The article explains how the team uses orbital-dynamics models (Hill-Clohessy-Wiltshire equations and numerical refinements) to analyze station-keeping, perturbations from Earth’s non-spherical gravity field, atmospheric drag, and more.

The example shown is a cluster of 81 satellites with mean altitude ~650 km, cluster radius ~1 km, distance between nearest neighbours ~100-200 m. The modelling indicates that modest station-keeping manoeuvres may suffice to keep the formation stable.

2.3 Radiation tolerance of TPUs

Another critical challenge: space is a harsh environment. The payloads (Google’s Trillium TPUs) must withstand total ionising dose (TID) and single-event effects (SEEs).

For typical inference workloads the observed error rates translate to very low failure probabilities (1 per 10 million inferences). However, training workloads and full system-level reliability in orbit still require further study.

2.4 Economic feasibility and launch costs

Even if all the technical pieces align, the economics need to make sense. Historically, launching mass to orbit has been prohibitively expensive. Google analyses show that if launch costs to LEO fall to ≈ US$200/kg by the mid-2030s, then a space-based data centre could become roughly cost-comparable (on a per-kW/year basis) to an equivalent terrestrial data centre.

They reference a learning-curve for launches (mass reuse, scale, competition) that could drop costs even further (< US$60/kg in an optimistic scenario). The vision is only viable if launch and operation costs decline significantly — but they appear to believe the trajectory is favourable.

3. Future directions

Google emphasizes that these findings are an early milestone — the core concepts are not blocked by fundamental physics or insurmountable economics, but many engineering challenges remain.

Next steps include a “learning mission” in collaboration with Planet Labs, slated for early 2027, to launch two prototype satellites and test the TPU hardware, optical inter-satellite links, and validate the models in orbit.

They also state that for true scale-out, future satellite design may move beyond relatively conventional discrete compute payloads — toward an integrated “system-on-chip, radiator, solar-panel” design optimized for space, much like how smartphones drove past generic PC components.