Google's Suncatcher Satellite Is Online and Running AI Chips in Orbit

Google's Project Suncatcher prototype is online. On Friday, October 30, the company said its four-chip satellite is working and talking to ground stations, about a month after launch. The spacecraft is small and the test is narrow, but it answers the first question anyone would ask about putting AI hardware in orbit: can the chips survive and work up there?

What Google says is working

The satellite lifted off on October 1, 2026, from Vandenberg in California. It flew as part of SpaceX's Transporter-18 rideshare on a Falcon 9 rocket, according to CNBC. More than 100 other spacecraft shared the flight, and Planet built the spacecraft platform.

Travis Beals, the senior director leading Project Suncatcher, wrote in a blog post after launch: "Our team has confirmed contact with the satellite and it is operating as expected." A Google spokesperson later told TechCrunch that the satellite was "operating nominally."

The most vivid proof came from software rather than telemetry. After about three weeks of commissioning, engineers prompted Google DeepMind's Gemma 4 model, preloaded on the satellite, to write a haiku about space and report orbital conditions. The exchange is the easiest part of the story to picture, and it is also a real functional test: the model ran on hardware in orbit and sent its answer back to the ground.

Four chips, one at a time

The hardware is modest. The satellite carries four Google TPUs, the custom AI accelerators Google uses in its own data centers, and draws roughly 1 kilowatt of power. A TechCrunch correction confirms that only one of the four TPUs operates at a time.

Reports disagree on how long each session runs. GPB cites about 15 minutes per TPU run, while Hyper.ai says about 20 minutes. The discrepancy matters because session length is one of the few concrete measures of how the hardware is holding up.

Beals described the mission as a "very minimal test" meant to confirm that the chips can run in space. The main purpose is a radiation and thermal-stress check on real chips in orbit, not a simulation, according to AI Weekly. The mission also aims to validate whether Google's TPU hardware can support orbital inference, which means running a trained model rather than training one.

Why Google wants compute in orbit

Google first revealed Project Suncatcher in November 2025 as a moonshot to run solar-powered AI compute in space. The physics behind the idea is straightforward. In low Earth orbit, satellites can get near-constant sunlight and generate up to eight times more solar power than they could on the ground, according to Google's project blog.

That matters because ground data centers face tight limits on power and siting. An orbital facility would sidestep some of those constraints, though it would bring new ones.

Google's stated long-term goal is a constellation of satellites that shares AI workloads. The current spacecraft is a single prototype. Google's longer-range vision of 81-satellite clusters in arrays about one kilometer across, as reported by DCD, remains aspirational.

Rivals take a different route. Starcloud and Cowboy Space plan inference on individual satellites rather than on linked clusters, according to TechCrunch. Google is betting that networking satellites together will pay off, while some competitors would rather start with one machine per spacecraft. Tech-Insider also cautioned that headlines overstated the milestone, and that the prototype is not a working data center.

The cost problem that will decide it

Cooling and launch cost are the hard constraints. Radiators are among the heaviest parts of a satellite, according to the Washington Times, which is why thermal performance matters as much as the chips themselves. TechTimes framed the mission around cooling rather than radiation, a sign that observers are still sorting out which risk matters most.

Launch cost is the bigger hurdle. Google's modeling says launch prices must fall below about $200 per kilogram by the mid-2030s to approach cost parity with ground data centers, according to Global Data Center Hub. Today's price is about $3,600 per kilogram on reusable Falcon 9, the figure Google's paper uses. That gap is enormous.

Google's own peer-reviewed analysis, published earlier in October, estimated that SpaceX would need to fly Starship about 1,800 more times to match the cost declines it saw from earlier vehicles. Spread over a decade, that works out to roughly 180 launches a year.

Skeptics are not quiet about this. The Washington Times analysis found that, even under Google's assumptions, matching ground costs would leave very little room for launch and hardware costs. Beals told NPR that he does not expect orbital computing to be cheaper than ground computing within five years.

The winners and losers are easier to sort out than the technology. SpaceX and Planet, both US companies, stand to benefit from any follow-on constellation work. Consumers see no product or price change from the experiment. Most analysts read it as having no effect on data-center siting this decade.

What to watch in 2027

Google plans two more Suncatcher satellites in 2027, with more advanced chips and laser inter-satellite links. Those launches will test what the prototype cannot: whether spacecraft can share AI work with each other at all. Laser links are the piece that turns a single experiment into a network, so whether they work as planned is the question to track.

Several open questions remain. Google has not said whether the prototype will keep working through thermal and radiation stress over its planned one-year mission. The exact session length is still reported inconsistently, and the orbital altitude has not been disclosed in the reporting available so far.

Readers do not need to take any action. If you follow AI infrastructure, the useful steps are to watch for Google's published test results and the 2027 launches, and to compare any new runtime figures against the 15- and 20-minute reports.

For readers who want to see how AI infrastructure is pulling in other directions, USA 24 has covered PsiQuantum's 8x8 optical switch aimed at AI data centers, a ground-side bet on faster networking. SpaceX's reported push for 80,000 US cell sites shows how far the company's ambitions now reach beyond launch.

The next real test is whether these chips keep running through a full year of heat cycles and radiation, and whether Google can link its satellites by laser in 2027. Those answers will say far more about orbital AI than any haiku.