Google Sends AI Chips to Orbit in Data Center Moonshot

Google has launched a prototype satellite carrying its custom artificial-intelligence chips into orbit, an early step in a long-term research effort to determine whether space can host the enormous computing infrastructure that AI demands. The refrigerator-sized spacecraft, developed with aerospace company Planet Labs, launched aboard a SpaceX Falcon 9 rideshare mission from California on October 1, and Google says it has confirmed contact and that the satellite is operating as expected, according to NPR.

The mission, called Project Suncatcher, marks the first time Google has sent one of its advanced Tensor Processing Units (TPUs) into space. It reflects a broader industry bet that the soaring energy and land demands of AI may eventually push some computing off the planet entirely.

What Google Put in Orbit

The satellite carries four TPUs — Google-designed chips for machine learning, comparable to Nvidia’s widely used GPUs. According to Scientific American, the spacecraft will run versions of Google’s Gemini AI models in orbit. NPR reported the chips will operate for roughly 15 minutes at a time, a limit imposed by the challenge of shedding heat in the vacuum of space.

The mission is designed to operate for one year, gathering data on how the TPUs withstand the physical stress of launch and the radiation and thermal extremes of the orbital environment. Google described the effort in a blog post as a “long-term research moonshot” intended to test hardware survival, cooling, and interconnectivity.

Google placed the satellite in a sun-synchronous orbit, where its solar panels will sit in near-constant sunlight, eliminating the need for heavy batteries or backup power. The company argues that low Earth orbit can generate up to eight times more solar power than the surface of the Earth, making space an attractive — if distant — source of energy for computation.

Why Space, and Why Now

The appeal is partly about power and partly about politics. Data centers on Earth are increasingly straining electricity grids and drawing local opposition, and space offers both near-unlimited solar energy and, as NPR put it, “no protesters.” Google CEO Sundar Pichai said in May 2026 that demand for AI services exceeds supply, and the company projects capital expenditures of $180–190 billion this year — more than six times its 2022 level.

“The sun puts out almost all of the power in our solar system,” Travis Beals, the Google senior director who leads Project Suncatcher, told NPR. “So in some sense, this project is about tapping into the best way to use solar power to run AI compute.”

But Beals is candid about the timeline. “I don’t see this being something where it’s cheaper to do this in the next five years,” he said. “I think it will take longer than that.” He described the effort as a “moonshot” that is “not going to be a completely straightforward journey.”

The Hard Engineering Problems

The obstacles are substantial. In space there is no airflow, so heat can only be radiated away — Google is testing a combination of heat pipes and radiators, which are among the heaviest components on the mission, and heavier payloads cost more to launch. Launch itself subjects components to vibration and forces up to 10 times the pull of gravity, with individual chips enduring as much as 50 to 100 g.

Radiation is another concern. Google tested its Trillium TPUs in a proton beam facility at the University of California, Davis, running AI workloads to see how the chips tolerated radiation doses exceeding what a five-year space mission would deliver. The company found an error rate of roughly one in a million for typical inference operations — acceptable for many tasks, Beals told TechCrunch, but “already problematic” for the massive training runs that use thousands of chips for months.

Independent experts remain cautious. Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University, told NPR the idea sounds “very sci-fi,” though much of the underlying technology has been proven at a research level. Repairing or maintaining hardware in orbit, he warned, multiplies cost and complexity “by a factor of 10, maybe a factor of 100. And so there has to be a big payoff.”

The Economics Remain Unproven

Google released a peer-reviewed white paper in the journal Joule alongside the launch, arguing that SpaceX’s reusable rockets have driven a roughly 20% annual decline in launch prices and that costs could reach about $200 per kilogram by 2035. Meeting that projection, however, would require Starship to lift roughly 370,000 tons of payload — about 1,800 launches over a decade, or 180 a year. Google acknowledged its analysis is not an economic feasibility study.

Beals told NPR that technical possibility alone is not enough. “It doesn’t help a lot if this is technically possible,” he said, “if it’s always going to be too expensive to be practical.”

The company’s long-term vision is a network of satellites, each carrying dozens of TPUs, flying in close formation and processing in parallel. TechCrunch reported the concept envisions 81 satellites connected by laser communications links. Google plans to launch two more satellites in 2027 to test those inter-satellite laser connections.

A Crowded Field

Google is not alone. The startup Starcloud launched a spacecraft in November 2025 carrying an Nvidia H100 chip and demonstrated a version of Google’s own Gemini AI from orbit; by March 2026 it had raised $170 million at a $1.1 billion valuation. SpaceX, in which Google is a major investor, has said in an SEC filing that it expects to begin deploying “orbital AI compute satellites” as early as 2028. A cluster of smaller players — Axiom Space, Lonestar, Satlyt, and Cowboy Space Company — are pursuing related concepts, and the Transporter-18 rideshare that carried Google’s prototype lofted more than 100 payloads in total.

Elon Musk, SpaceX’s CEO, has argued that space is the “only way to scale” data centers to meet future AI demand, a view that Google’s Beals treats more cautiously. He told NPR he expects terrestrial and orbital computing to coexist for a long time: “In the really-long run, perhaps the set of things you do in space gets larger and larger until that’s where most of the compute is happening. But I think for a long time you’ll have both coexisting.”

What to Watch

The Project Suncatcher prototype’s year in orbit will yield the first real-world data on whether Google’s chips can survive space and whether the company can cool them efficiently enough to matter. The next milestone — two satellites linked by laser in 2027 — will test whether individual spacecraft can be stitched into something resembling a data center.

None of this is close to commercial. For now, the experiment is less a product than a hypothesis: that the sun’s energy, and the absence of anyone to object to it, may eventually make the cosmos a sensible place to run the world’s AI.