Google sends AI into space: the first Suncatcher prototype goes into orbit

Google has sent its first artificial intelligence computer into orbit. The Project Suncatcher prototype, built in collaboration with Planet, launched on October 1st aboard SpaceX’s Transporter-18 mission. Google confirmed that it has established contact with the satellite and that the system is working as expected. It is the first concrete test of a project with which the Mountain View group wants to understand if, in the future, part of the computing power necessary for artificial intelligence can be moved into orbit. In the coming weeks, the satellite will collect data on the behavior of Google’s TPUs during spaceflight and their ability to withstand radiation and extreme thermal conditions.

The idea behind Project Suncatcher it starts above all from energy. In low Earth orbit, Google explains, a satellite can have almost continuous access to sunlight and panels can produce up to eight times more energy than similar installations on Earth. Looking ahead, multiple satellites could be linked together to handle much larger AI workloads.

The first obstacle is surviving space

Before even thinking about a real orbital data center, however, it is necessary to verify that the hardware can survive the journey and the space environment. According to Google, during the approximately ten minutes needed to reach orbit the satellite can be subjected to accelerations up to ten times the gravitational one, while individual components can undergo stresses of between 50 and 100 grams. For this reason, the company subjected the satellite to vibration tests and exposed the TPUs to proton beams at the Crocker Nuclear Laboratory at the University of California at Davis. Early results, Google reports, indicate that the chips used in the project can withstand a higher total radiation dose than expected during a five-year mission.

Another obstacle is heat. Space is extremely cold, but in a vacuum there is no air that can carry away the heat produced by the processors. Google is therefore experimenting with systems based on heat pipes and radiators to disperse the energy generated by the TPUs. The problem is far from solved. TechCrunch reports that the first satellite will use the TPU for sessions of about 15 minutes at a time, precisely so as not to put excessive stress on the power and thermal management systems. “We have done ground tests, but no test is completely comparable to reality,” explained Travis Beals, head of Project Suncatcher.

Satellites connected with lasers

The long-term goal is much more ambitious. Google imagines satellites each equipped with dozens of TPUs and capable of working together in close groups. To do this they will have to exchange huge amounts of data with very low latency. The solution studied is a laser communications network. Google compares the precision required to hitting a coin-sized target with a laser beam miles away while both dots are moving. In 2027 the company plans to put two satellites into orbit to test this very link.

TechCrunch also reports that, in the long-term scenario discussed by the research group, Google has considered configurations with dozens of satellites capable of processing data in parallel. However, Beals himself defines Suncatcher as a “long-term moonshot”, a research project intended to develop over a long period of time.

How much would it cost to put a data center in space?

The economic question probably remains the main question mark underlying every operation. Based on scientific work published by Google, the group believes a sharp decline in launch costs over the next decade is plausible. The peer-reviewed paper on the project, now published in the journal Jouleanalyzes precisely the conditions necessary to make a large-scale spatial computing infrastructure possible.

According to calculations reported by TechCrunch, to maintain the cost reduction trajectory hypothesized by the researchers, it would be necessary to transport quantities of material into orbit that are enormously higher than the current ones, with hundreds of launches per year. This is a theoretical scenario and not an operational plan already decided by Google.

Because they won’t replace terrestrial data centers

Even if the technology works, data centers in space are unlikely to replace those on Earth. According to an analysis by the Information Technology and Innovation Foundation, orbital infrastructures are better suited to complement terrestrial ones than to replace them, at least in the near future.

ITIF lists cooling, radiation exposure, communications capability, difficulty of maintenance, cybersecurity, launch costs and the sustainability of the orbital environment as major obstacles. Space systems are also much more difficult to repair or upgrade than a data center on Earth. We are therefore moving towards a specialization: orbital systems could deal with asynchronous and energy-intensive processing, while terrestrial data centers would retain the advantage in real-time and latency-sensitive tasks.

Google is not alone in this race

Project Suncatcher fits into an industry that is attracting a growing number of companies. Reuters reported that on the same Transporter-18 mission also launched MOI-1A from Indian startup TakeMe2Space, a satellite equipped with Nvidia Orin NX processors designed to process collected data directly in space, instead of sending large amounts of raw information to Earth.

TakeMe2Space told Reuters it already has 23 customers in the agricultural, mining, logistics and insurance sectors and plans to launch larger, connected satellites starting in 2027. For Google, however, we are still in the experimental phase. “Some things can only be tested in space”, explained the company when announcing the entry into orbit of the prototype

Suncatcher’s first goal is therefore much simpler than a space data center: to understand whether the chips that power artificial intelligence on Earth can continue to work so far away from us.