Google Puts Its AI Chips in Orbit, Testing Space Data Centers

On October 1, a SpaceX Falcon 9 rocket lifted off from Vandenberg Space Force Base carrying a refrigerator-sized satellite with four of Google's custom AI chips bolted inside. It is the first flight hardware for Project Suncatcher, the company's moonshot bet that the future of AI computing may not be built on Earth at all.

Google's Tensor Processing Unit (TPU) boards — its custom chips for machine-learning workloads. Four TPUs are now being tested in orbit.

What Launched on October 1

The satellite rode into orbit on SpaceX's Transporter-18 dedicated rideshare mission, which carried 130 payloads to low Earth orbit, according to SpaceX's mission posts. Planet Labs, the Earth-imaging company, built the refrigerator-sized satellite and is Google's hardware partner for the project.

Shortly after the launch, Google confirmed it had established contact with the satellite. "Our team has confirmed contact with the satellite, and it is operating as expected," wrote Travis Beals, the Google senior director leading the project, in a post following the launch. Planet confirmed it had made contact and begun checking the satellite's systems before Google starts its chip experiments.

The project lead is Beals, Google's senior director for paradigms of intelligence, and the hardware at the center of the test is the company's Tensor Processing Unit — the custom AI accelerator Google designed for machine-learning workloads and has long deployed in its own data centers. This is the first time the chips have been sent into space.

The Project: Suncatcher

Project Suncatcher is a long-term research moonshot exploring whether space could one day host scalable machine-learning infrastructure, according to Google. The company first unveiled the effort in November 2025, laying out the case that orbiting data centers could tap an energy source no grid on Earth can match.

"This is the first step in a long-term research moonshot exploring whether space could one day host scalable machine learning infrastructure," Beals wrote in Google's blog post the day of launch. "Some things can only be tested in space."

The idea is deceptively simple: in low Earth orbit, satellites can access near-constant sunlight, generating up to eight times more solar power than panels on Earth, according to Alphabet's pre-launch announcement. String enough such satellites together — linked by high-bandwidth laser communications — and, the theory goes, you could run serious AI workloads on sunshine alone.

The design work is serious: Google published plans in a peer-reviewed paper in the journal Joule, envisioning clusters of up to 81 satellites flying in formation in a dawn-dusk sun-synchronous orbit roughly 650 kilometers up, where shadows are minimal. Bench tests of the free-space optical link that would connect them reached 800 gigabits per second in each direction.

But Google has been careful to frame Thursday's launch as exactly what it is: a survival test for hardware, not a data center in orbit. Beals himself described it as a "very minimal test" in comments to NPR, meant to verify the chips can run in space at all. Google SVP James Manyika has confirmed that any operationally useful version of the idea is still years away.

A SpaceX Falcon 9 rocket lifts off. Google's Suncatcher prototype rode to orbit on a Falcon 9 rideshare from Vandenberg Space Force Base.

The Hardest Test: Heat

The biggest enemy of the orbiting TPUs is not what most people expect: heat — not the cold of space.

Space has no air to carry heat away, so chips that would be cooled by fans and water in a terrestrial data hall can only shed warmth slowly, through radiation. That hard physical limit defines the entire experiment: over the coming weeks, Google plans to wake the TPUs in short bursts of about 15 minutes at a time, running a version of its open-weight Gemma AI model to answer simple queries, then powering them down before heat builds up too far.

It is a modest workload on purpose. The goal is measurement, not performance: engineers want to watch how the chips behave and gather data to feed into the next hardware design.

"If it keeps functioning for the full year as planned," as one industry analysis noted, the mission would mark the longest continuous run yet of general-purpose AI accelerators in space. Planet will finish commissioning the satellite first; then Google begins the chip experiments.

Surviving Radiation

The team did not fly blind. Before the launch, Google radiation-hardened the Trillium chips through proton-beam testing at the University of California, Davis' Crocker Nuclear Laboratory, simulating a full five-year orbital radiation dose on the ground.

The results, published in the design paper, were encouraging but honest: the chips' High Bandwidth Memory subsystems were the most sensitive component, showing irregularities after 2 kilorads of exposure — nearly three times the expected five-year mission dose of 750 rads — with no hard failures up to the maximum tested dose of 15 kilorads.

The chips also went through intense vibration testing to confirm they could survive the violence of a rocket launch. Only space, though, can provide the real exam.

Why Space?

The timing is not accidental. Artificial intelligence has created an electricity appetite that power grids are struggling to feed, and opposition to the proliferation of energy-hungry data centers is growing in communities across the United States. New terrestrial data center projects are facing protests, permitting fights, and skyrocketing power costs.

At Google's annual developer summit in May, CEO Sundar Pichai said demand for AI services now exceeds supply and forecast capital expenditures of $180 billion to $190 billion this year alone — more than six times the 2022 level — as the company builds computing infrastructure and develops its own chips.

In that context, orbit starts to look attractive. "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 told NPR. "So in some sense, this project is about tapping into the best way to use solar power to run AI compute."

The prototype's solar panels generate roughly one kilowatt — enough electricity, as observers quickly noted, to run a hair dryer. That is the scale of the beginning. The scale of the ambition is constellations drawing free, uninterruptible sunshine.

Who Else Is Racing There

Google is not alone in looking up. A startup called Starcloud has already put an Nvidia H100 AI accelerator into orbit on a satellite that trained an AI model, according to industry reporting. And SpaceX itself — which launched Google's prototype — has been explicit that it sees space-based computing as a growth business, touting "infinite real estate" for orbital data centers and, according to reports, planning large-scale deployment of orbital data centers as early as 2028.

For now, though, Google's effort is the one with flight hardware. The 2027 milestone is two more satellites testing the high-bandwidth laser links a constellation would depend on.

What This Means

The Suncatcher launch matters less for what it does than for what it signals. Every major AI company is now confronting the same ceiling: intelligence takes electricity, and electricity is getting harder to buy. Nuclear deals, gas turbines, geothermal plants — the industry has tried the terrestrial options. Orbit is the most radical answer yet proposed, and Google is the first hyperscaler to put silicon in the sky.

It also reframes the economics of the space industry. Reusable rockets made it cheap to reach orbit; the next business may be not what satellites can see, but what they can compute. SpaceX's rideshare business — 130 payloads on one Falcon 9 — shows how routine the logistics have become.

At the same time, the gap between four chips running in 15-minute bursts and a working data center is enormous. Thermal management in vacuum is brutally hard, radiation is unforgiving, and Google's own executives say the payoff, if it comes, is years off.

In low Earth orbit, satellites can access near-constant sunlight — up to eight times more solar power than on the ground, Google says.

What Happens Next

Over the coming weeks, Google's engineers will collect the in-orbit data that defines this mission: how the TPUs handle the physical stress of spaceflight, radiation exposure, and extreme temperature changes. What they learn will feed directly into the design of the 2027 pair of satellites and the laser-link tests that follow.

"As we begin our experiments, we'll use what we learn to refine our designs, and we're excited to share more as the mission unfolds," Beals wrote in the launch-day post.

Whether space ever hosts a real AI data center is still an open question. But as of October 1, the first chips in that experiment are circling the Earth — waiting, in 15-minute bursts, to find out.