Google's TPUs to catch some rays in orbit next week

Google's TPUs are now being tested for space deployment, potentially altering AI infrastructure economics.

Google is set to launch its Tensor Processing Units (TPUs) into orbit on a SpaceX rocket next week as part of its Project Suncatcher initiative. This test aims to evaluate how its custom AI hardware withstands the stresses of launch and the harsh conditions of space, with the long-term goal of developing sustainable, solar-powered datacenters in orbit for AI processing. The TPUs, specifically the Trillium accelerators, have undergone Earth-based tests simulating launch forces and radiation. Challenges include surviving extreme gravitational forces during ascent, radiation-induced errors, and effectively dissipating heat in a vacuum, which requires significant radiator area. The mission, scheduled for October 1 on the Transporter-18 rideshare mission, will assess the viability of these components and identify failure points for future iterations. Google is also exploring high-speed laser-based communication for inter-satellite connectivity to enable distributed AI training and inference.

Google's foray into orbital datacenters, Project Suncatcher, signifies a bold, albeit speculative, move to harness solar power for AI. While analysts like Gartner's Bill Ray have labeled the concept "peak insanity," citing decades of unviability and lack of terrestrial application, Google is proceeding with a pragmatic, moonshot approach. This test flight will reveal critical data on hardware resilience and heat management, essential steps before any large-scale deployment can be considered, and it positions Google as a key player in the nascent space-based computing race.

The success of this proof-of-concept is contingent on overcoming significant technical hurdles, from surviving launch G-forces to managing heat radiation and potential cosmic ray interference. Should these challenges be met, it could pave the way for distributed AI computation beyond Earth's atmosphere, offering a path towards energy-efficient AI processing, though the economic feasibility remains heavily dependent on drastically reduced launch costs, a prospect still met with skepticism.