As Google's AI chip reaches space, the tech giant faces an earthly problem
It’s the company’s "long-term research moonshot."
An AI chip. Illustration by Cybernews. Images by NASA/Rohit Choudhari via Unsplash
Follow us
- Google is testing Trillium AI chips in low Earth orbit through Project Suncatcher.
- The company wants to see if space computing can reduce AI data centers’ power and cooling demands.
- Scaling the idea could require 370,000 tons of payload and far cheaper launches.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
Google has started testing its AI chips in low Earth orbit, in what could pave the way to data centers in space. But even if it works, its biggest issue might be not only how to get all the hardware needed into space, but how to do it cheaply enough.
Google’s Project Suncatcher prototype satellite reached low Earth orbit on October 1st, 2026, aboard a SpaceX Falcon 9, carrying 4 Google Trillium TPUs, Google's AI chips.
The tech giant aims to test whether these chips can survive and operate in space. To do so, they need to survive, for example, launch vibrations, radiation, and extreme thermal conditions.
Satellite developer Planet and Google also plan to launch 2 purpose-built satellites in 2027 to test high-bandwidth optical communication between satellites.
With these space experiments, Google is exploring whether space computing is a feasible way to solve AI's power problem, as data centers need huge amounts of electricity and cooling.
Space could solve AI’s power problem
The researchers estimate that appropriately positioned solar panels could generate up to 8 times more power than on Earth, while placement in space would also eliminate conventional air-based cooling.
Google's idea is to create clusters of satellites carrying many TPUs. For example, one proposed configuration uses 81 satellites in a cluster within roughly a 1-kilometer radius.
However, to reach the required scale, Google would need to send 370,000 tons of payload into orbit, according to TechCrunch.
Then, if Starship eventually carries 200 metric tons per launch, that would require 1,850 launches over 10 years, or 185 launches per year. So far, Starship has never flown more than 5 times in a year, according to TechCrunch.
But getting there cheaply is another matter
Cost is another question, as it currently costs around $3,600 to send a kilogram of hardware into low Earth orbit, while the cost needs to fall to $200 for it to become approximately comparable with terrestrial data-center power costs. According to Google's estimates, that price could be reached by the mid-2030s.
However, competition among similar projects might help these costs drop faster. For example, Blue Origin has filed plans for up to 51,600 orbital data-center satellites under Project Sunrise.
Meanwhile, Starcloud is working on launching up to 88,000 satellites and has already demonstrated an Nvidia H100 operating in orbit. Axiom Space is also pursuing orbital data centers and says its technology is moving from kilowatts toward megawatts.
In either case, while both the ambitions and the technical and cost challenges are huge, all these experiments show that the technology is moving forward. We'd need to wait for proof that the economics also work.
Google calls its Suncatcher project a "long-term research moonshot."
