# Google Tests AI Chips in Space for Sustainable Computing TPUs launch on SpaceX rocket to validate orbital datacenter feasibility - **Source:** [www.theregister.com - Articles](https://www.theregister.com) - **Published:** 2026-09-24 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/5010bbd7-5b99-4fa7-ba1f-839a05ea0ddd ## Moonshot computing initiative Google's Project Suncatcher aims to deploy tensor processing units in space, leveraging unlimited solar energy for sustainable AI infrastructure. The upcoming SpaceX launch will test whether terrestrial TPUs can withstand 50-100G forces during ascent. This follows vibration table simulations but marks the first real orbital stress test for Google's Trillium accelerators, not originally designed for space environments. ## Radiation challenges in orbit Beyond launch stresses, Google faces cosmic ray-induced bit flips and cumulative radiation damage to chips. Using UC Davis' proton beam facility, engineers bombarded TPUs to simulate years of space exposure in weeks. Each TPU requires 1.3m² radiators to dissipate heat in vacuum conditions, complicating satellite designs that must pack dozens of chips. Heat pipes and pumped coolants are being evaluated for thermal management between chips and radiators. ## Economic viability hurdles SpaceX's current $7,000/kg launch costs must drop to $10/kg for orbital datacenters to compete with terrestrial alternatives. Even with cheaper launches, Google would need laser interconnects between satellites for distributed AI workloads - a technology not yet adapted for high-bandwidth, short-range cluster communications. Competitor SpaceXAI targets 1M-satellite constellations but faces similar economic constraints despite Musk's ambitious timelines. ## Industry skepticism persists Gartner analysts deem space-based datacenters economically unviable for decades, questioning their terrestrial relevance. Unlike SpaceXAI's 250kW satellite concepts, Google takes incremental steps - first proving basic functionality before scaling. The Transporter-18 mission carries only prototype hardware, with failure points intended to guide future iterations rather than deliver immediate operational capability. ## Thermal management breakthroughs Google's vacuum chamber tests revealed unconventional cooling requirements for space-bound TPUs. Without atmospheric convection, radiators must handle concentrated heat loads from densely packed chips. Early designs suggest each satellite could require radiator surfaces larger than the compute hardware itself, creating structural and weight challenges for launch vehicles. ## Radiation hardening approach Rather than developing space-specific TPUs, Google is stress-testing commercial Trillium accelerators through proton beam exposure. This pragmatic approach contrasts with competitors' clean-sheet designs but risks unforeseen failure modes. Single-event upsets from cosmic rays could corrupt AI models during extended orbital operations despite lab simulations. ## Distributed computing paradigm Launch mass constraints force Google to consider multi-satellite clusters with laser links, creating latency challenges for AI training. Traditional inter-satellite communication systems lack the bandwidth needed for tensor operations across orbital nodes. Google must adapt existing optical technologies to maintain sub-millisecond latency between nearby satellites. ## Sustainable computing frontier While terrestrial datacenters face energy constraints, orbital solutions introduce new tradeoffs between solar abundance and launch/maintenance costs. Google's experiment could validate radiation-hardened commercial chips for space, potentially benefiting other orbital compute applications. Success would require simultaneous breakthroughs in launch economics, thermal management, and distributed AI architectures. ## Sources covering this story - https://www.theregister.com/systems/2026/09/24/googles-tpus-to-catch-some-rays-in-orbit-next-week/5298990 — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._