Google recently made strides in space technology by launching its prototype orbital compute satellite aboard a SpaceX rocket. This event marks the first time Google has sent one of its advanced chips into orbit, featuring a Google Tensor Processing Unit (TPU) that competes with Nvidia’s GPUs. Built by Planet Labs, the satellite is designed to test the TPU’s functionality in the harsh conditions of space.
To maximize performance, the satellite will provide a kilowatt of continuous power and activate its TPU in 15-minute bursts. This approach helps prevent strain on the satellite’s power and thermal management systems. Google is also working with Planet Labs on a future demo, set to launch next year, which will include two satellites specifically designed for advanced computing tasks. These upcoming satellites will utilize a laser link for communication, further boosting their capabilities.
Described by Google executive Travis Beals as a “long-term moonshot,” this project aims to establish the foundation for future space infrastructure and AI workloads. Google envisions a network of 81 satellites flying in close formation, processing data simultaneously. Beals highlighted the significance of bandwidth and latency between the TPUs, stating, “we’re trying to look ahead to not just what workloads exist today, but where they will be in five years.” This forward-thinking approach is essential, as the rockets required to efficiently scale up orbital data centers are still under development.
Alongside the launch, Google published a peer-reviewed version of its white paper on orbital data centers in Joule. This document rigorously analyzes how computing can be integrated into orbit, although the researchers clarified that it does not serve as an economic feasibility study. Nevertheless, it provides insights into Google’s expectations for decreasing costs of launching payloads into space over time.
Looking ahead, Google predicts that SpaceX will achieve a launch cost of around $200 per kilogram by 2035. This estimate is based on a learning curve that SpaceX has reportedly experienced, with a reduction of about 20% in costs per year since the Falcon 1 rocket’s debut. To reach these lower launch prices, the authors of the paper estimate that SpaceX’s Starship will need to deliver an impressive 370,000 tons of payload into orbit, which translates to about 1,800 launches over the next decade–averaging 180 launches each year.
However, that target presents a significant challenge for a vehicle that has yet to exceed five flights in a single year. SpaceX has ambitious plans, with CEO Elon Musk suggesting that Starship could potentially achieve an hourly flight rate by 2029, although past promises from Musk have not always come to fruition as expected.
On the technical side, Google’s research suggests that its chips are likely to endure the radiation found in space. Initial tests revealed that the chips required reconfiguration after receiving more shielding than they would encounter in actual space conditions. Updated evaluations indicated a slightly higher error rate in the chip’s logic circuitry, yet the company remains confident that their chips can handle large inference workloads over the satellite’s anticipated five-year lifespan. Beals noted, “The error rate is very low if you’re thinking about typical inference operations, right? Like one in a million.” However, the chips may face difficulties during mega-scale training runs, where thousands of chips operate simultaneously for extended periods.
Google’s launch of its orbital compute satellite marks a significant step in testing the viability of advanced chips in space, while highlighting the ongoing challenges of achieving cost-effective launches. As the company navigates the complexities of scaling up satellite operations and addressing the limitations of current rocket technology, the outcomes of these experiments could shape the future of computing infrastructure in orbit.
