Google is aiming for the stars—literally—with Project Suncatcher, its latest attempt to take artificial intelligence where no chatbot has gone before: outer space.
How do we put machine learning in space?
AI in space… A lot could go wrong.
Project Suncatcher, announced last year, is Google’s attempt at a true moonshot: what if space itself could host the next generation of machine learning? This is all to try and increase AI computing capabilities in orbit. A prototype satellite is due to launch soon to see whether its custom Tensor Processing Units (TPUs) can withstand the harsh conditions outside Earth’s atmosphere. If you’re thinking, ‘It’s impossible that there could be no problems with putting advanced AI hardware into space,’ then you’re not alone.
In low Earth orbit, satellites are exposed to near-continuous sunlight—potentially allowing them to generate up to eight times more solar power than their counterparts on the ground. Image huge numbers of satellites, linked as a network, carrying out AI processing above the clouds. At first,, this sounds like the opening of a sci-fi horror story, but is still an interesting concept.
Before we worry about rogue satellites getting interested in world domination, there’s one more important question: can AI chips actually function in space? The Suncatcher team will find out, since their hardware is being transported on SpaceX’s upcoming Transporter-18 mission. The company will collect data in orbit to see whether Google’s TPUs can withstand the pounding vibrations, cosmic radiation, and thermal stress of a journey beyond the stratosphere.
The journey to low Earth orbit is a white-knuckle, ten-minute ride like a rollercoaster, exposing the satellite—and its valuable chips—to forces amounting to as much as 100 times the force of Earth’s gravity. To put its prototype through the AI version of a NASA bootcamp, the Suncatcher team carried out vibration tests, radiation exposures, and trials in thermal vacuum chambers.
But orbit isn’t just about launch-day stress. Out there, satellites are bombarded by cosmic rays and solar events that could make even the hardiest smartphone throw in the towel. So, the team blasted their TPUs with proton beams at UC Davis—yes, like something out of a superhero origin story—to see how they handled ‘bitflips’ and other digital gremlins. The outcome? The chips held up surprisingly well, shrugging off more radiation than they’d get in five years circling the Earth.
Of course, space has its own weird problems. For example: how do you cool a chip when there is absolutely no air? TPUs, like all powerful processors, become very hot. On Earth, fans handle the situation, but in the vacuum of space, there’s no airflow. Instead, the only way to lose heat is through radiators. The team has come up with a combination of heat pipes and radiators—something you might imagine as interstellar plumbing—which has been tested in a special chamber that simulates space conditions.
Google plans for satellites to be loaded with dozens of TPUs and to orbit in synchronized groups. To keep them aligned, the satellites will have to communicate by aiming lasers at one another.
Although laser communication in space is not new, achieving it at high speeds over short distances between moving satellites is a new and highly difficult challenge. It’s like trying to hit a coin from miles away while both you and the coin are moving through space. Practical tests of these laser links are due in 2027, when Google intends to place two satellites into orbit and will have to hope for the best (and perhaps cross their fingers).
Will an AI-powered space network mark the beginning of humanity’s bright future, or will it become the plot of the next big disaster movie? Right now, Project Suncatcher is focused on finding out what works and what fails—literally and figuratively—when AI is placed in orbit. The approach followed is one of careful engineering, cautious optimism, and a touch of cosmic-sized ambition.