Analysis
On Oct. 1, Google launched a refrigerator-sized test satellite carrying four of its custom Tensor Processing Unit chips on a SpaceX rocket -- the first real-world test of Project Suncatcher, the company's bid to run AI data centers in space. Developed with Planet Labs, the satellite is operating as expected, according to Fortune, and will spend roughly a year in low Earth orbit testing how the chips handle launch vibration, radiation and the thermal extremes of space.
Google first disclosed Suncatcher in November 2025, the same month rival startup Starcloud launched its own satellite carrying an Nvidia H100 chip and used it to run Google's own Gemma model in orbit -- a reminder the idea of space-based AI compute predates Google's hardware test by almost a year. That drive comes as Google is separately racing to put agentic AI into its enterprise products, a sign AI compute now sits at the center of nearly every part of its strategy, not just its software.
The Race To Put Data Centers In Orbit
Google is not alone in this bet. SpaceX filed with the FCC in February 2026 for a constellation of up to 1 million data-center satellites, after Elon Musk told the World Economic Forum in January that space would become the cheapest place to run AI within two to three years. Blue Origin followed in March with its own Project Sunrise plan for 51,600 satellites. On the same Oct. 1 SpaceX rocket as Google's test satellite, Robinhood co-founder Baiju Bhatt's Cowboy Space Corp. launched its own Reason-1 satellite, which plans to test beaming solar power to Earth by laser within weeks -- a different bet on the same orbital-energy opportunity.
The numbers explain why this has gone from science fiction to a live engineering question: SpaceX's launch costs have fallen roughly 95% since 2008 to about $1,000 per kilogram, per ARK Invest research. That's still expensive measured against Big Tech's own plans to spend over $700 billion combined on Earth-based AI data centers in 2026, money that funds today's buildout with none of space's solar-power or land advantages, but none of its engineering risk either.
For infrastructure investors, orbital compute is a new category of bet with no public-market comparable and a far longer time horizon than a typical data-center buildout. Google's James Manyika said space data centers may not arrive for years, comparing the project to the roughly 15 years of research Google invested before quantum computing produced results.
But the obstacles are not cosmetic. Heat is hard to dissipate in space, so Suncatcher's chips run for just 15 minutes before shutting down to cool -- nowhere near the continuous operation a real data center needs. Transporting and deploying heavy solar arrays remains expensive, and there is no way to send a technician to fix a broken satellite. Google says its chips passed radiation tests beyond the dose a five-year mission would bring, but concedes that solar flares and cosmic rays remain risks it cannot fully model on the ground.
Google plans two more satellites in 2027 to test laser communication between them, a precision challenge Manyika compared to "hitting a coin-size target from miles away while both points are in motion." Whether that works will matter more than Monday's launch headline.
