Our Blog

Blog Index 

Google Launches First Datacenter Satellite: Project Suncatcher Sends Four TPUs Into Orbit to Test AI in Space

Posted on 2nd Oct 2026 12:04:45 in Artificial Intelligence, Machine Learning

Tagged as: Google, Project Suncatcher, space data centers, TPU, SpaceX, AI infrastructure

Google's plan to move artificial intelligence computing off the planet took its first real step into orbit this week, when a SpaceX Falcon 9 lifted a prototype satellite carrying four of Google's custom AI chips into space from Vandenberg Space Force Base in California.

The flight, part of SpaceX's Transporter-18 rideshare mission, marks the first in-orbit test of Project Suncatcher, the "moonshot" Alphabet first revealed in November 2025. The question the company is trying to answer is enormous: can AI data centers actually work in the harsh environment of space?

Within hours, Google said it had made contact with the refrigerator-sized satellite, which was built in partnership with Planet Labs. "Our team has confirmed contact with the satellite and it is operating as expected," Travis Beals, senior director and lead of Project Suncatcher, wrote in a blog post after the launch. He called the mission the first step in a long-term research effort exploring whether space could one day host scalable machine learning infrastructure.

Google CEO Sundar Pichai marked the milestone on X, recalling the first meeting where his team pitched the idea of putting compute in space, and noting that the launch shows "how far we've come and how far there is to go."

What Google Actually Sent to Orbit

The mission itself is modest, deliberately so. SpaceX's deployment schedule lists the spacecraft as "Project Suncatcher M1," released about 61 minutes after liftoff as one of 130 payloads on the Transporter-18 flight. The launch window opened at 11:15 a.m. Pacific time, and SpaceX reported all payloads deployed by 1 p.m. local time.

Aboard the satellite are four Tensor Processing Units, or TPUs, the Google-designed chips that power machine learning workloads inside its terrestrial data centers. This batch is Trillium-generation hardware. Over the next year, the team plans to run a version of Gemma, Google's open-weight AI model, to answer simple queries and prove the chips can compute in orbit.

There is one notable constraint: the TPUs will operate for only about 15 minutes at a time, limited by heat management. Cooling in a vacuum, it turns out, is one of the hardest problems in space computing.

The prototype draws power from solar arrays rated at around one kilowatt and sits in a sun-synchronous orbit, where its panels are almost never in shadow. Google estimates that near-constant sunlight in low Earth orbit can generate up to eight times more solar power than comparable systems on the ground, eliminating the need for heavy batteries or backup power.

Why Google Wants Data Centers in Space

"The sun puts out almost all of the power in our solar system. All of the other power sources that humanity has tapped into are just a tiny fraction of a percent," Beals said. "So in some sense, this project is about tapping into the best way to use solar power to run AI compute."

On the ground, Google can barely keep up with demand. At the company's developer summit in May, Pichai said demand for AI services exceeds supply and projected capital expenditure of $180 billion to $190 billion this year, more than six times the level of 2022. At the same time, opposition to power-hungry data centers is growing in communities around the world. In orbit, as NPR put it, there is virtually unlimited free energy and no protesters.

The long-term vision goes well beyond one satellite. Google's concept involves clusters of satellites flying in close formation and communicating by laser, functioning together as a single space-based data center. Next year, the company plans to launch two more TPU-equipped satellites with Planet to test high-speed laser links between them. Eventually, Google envisions 81 satellites flying within a one-kilometer radius, stitched together by those links.

The First Hurdle: Surviving Launch and Radiation

Before launch, Google put the hardware through a punishing test campaign. Engineers shook the satellite along all three axes to simulate the vibration of a rocket ride, where components can briefly experience forces of 50 to 100 times Earth's gravity. The Trillium chips were exposed to a proton beam at the University of California, Davis, for an equivalent of a five-year dose of space radiation. The whole satellite also spent time in a thermal vacuum chamber that simulates the environment it now occupies.

Now the real data collection begins. The mission is designed to operate for a year, gathering information on how the TPUs handle what Google calls "the physical stress of spaceflight and the radiation and thermal extremes of space."

Cooling, Lasers and the 81-Satellite Dream

Heat removal is the obstacle engineers talk about most. On Earth, data centers dump heat into air or water. In the vacuum of space, there is neither. "In a satellite in particular, using more power to do the computations means dumping more heat into the confined environment inside of the satellite," said Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University. Google says it is working on combinations of pipes and radiators to wick heat away from the chips, and it admits the radiators are among the heaviest components on the current mission, a real problem when every kilogram costs thousands of dollars to launch.

Networking is the second hard problem. Google's own peer-reviewed paper, "Toward a future space-based, highly scalable AI infrastructure system design," notes that existing network technologies probably cannot link multiple satellites into functioning clusters. The plan requires a design "significantly larger" and entailing "much closer formation flight than any previous or current satellite constellations." Google also flags optical satellite-to-ground communications as critical, and lists atmospheric turbulence, high-speed relative motion errors and precision beam tracking among the challenges to overcome. NASA's TBIRD mission, which demonstrated 200-gigabit-per-second laser links between low Earth orbit and the ground, is cited as a promising approach.

Maintenance raises harder questions still. When a chip fails in a terrestrial data center, a technician walks in and fixes it. In orbit, the same repair job becomes an operation costing vast sums. "Those all have their cost and complexity amplified by a factor of 10, maybe a factor of 100," Lucia said. "And so there has to be a big payoff."

The $200-Per-Kilogram Question

All of it hinges on economics. Google's researchers calculate that orbital data centers become competitive when launch costs fall to roughly $200 per kilogram. Current prices run between $1,500 and $2,900 per kilogram, and Google estimates that space-based AI clusters could become feasible around 2035, if costs fall fast enough.

The math leans heavily on SpaceX, which has cut launch costs by about 20 percent for every doubling of the cumulative mass it has flown. To reach $200 per kilogram, Google's paper estimates the industry would need to launch another 370,000 tonnes of payload, the equivalent of around 1,800 successful Starship launches, with components reused about 100 times. At that point, Google concludes, the annualized cost per unit of power in space could be roughly comparable to terrestrial spending.

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 expects terrestrial and orbital compute to coexist for a long time, with space handling a growing share only over decades.

The Race to Orbit Is Already On

Google is not alone in betting on orbit. SpaceX, led by Elon Musk, plans its own orbital data centers built around swarms of GPU-equipped satellites manufactured in Redmond, Washington, with solar arrays it intends to produce with Tesla. Chief operating officer Gwynne Shotwell said in September that the company will deploy "supercompute in space" in 2027, while Musk has claimed space data centers will be the cheapest way to train AI "within two years, maybe three at the latest." Blue Origin has also entered the field, and the venture-backed startup Starcloud put an Nvidia H100 chip in orbit last November and ran a version of Google's Gemini from space.

The corporate ties add another layer to the story. Alphabet holds a stake in SpaceX worth more than $82 billion, after the rocket company went public in June in a record IPO, even as the two firms compete directly in AI.

The Bottom Line

For now, Suncatcher M1 is a science experiment, not a floating data center. Four chips are a rounding error next to the hundreds of thousands of accelerators inside a modern hyperscale facility, and the economics remain firmly on the side of buildings on the ground. But if Google's bet pays off, it would reshape where the world's AI compute can live, adding one more frontier to an AI buildout that has consumed the industry for three years running. The first data points land over the next twelve months, while the real verdict is likely a decade away.

Sources

Recent Posts

Catgories

Archives

`
whatsapp me