Google’s First AI Chips in Orbit Come Online and Write a Haiku

By Economy Media Newsroom · October 30, 2026

Google’s prototype satellite for testing artificial intelligence chips in orbit is online and has answered its first prompts, TechCrunch reported Friday, marking an early step in the company’s long-range effort to move some AI computing off Earth. Asked by engineers to write a haiku about being in space, the AI model on board obliged.

What happened

The spacecraft is the first hardware from Project Suncatcher, a Google research effort announced in November 2025 to explore whether networks of solar-powered satellites could one day run machine learning at scale. It was built with Planet, a satellite-imagery company, and reached orbit on Oct. 1 aboard SpaceX’s Transporter-18 rideshare mission. Taiwanese tech outlet TechNews reported that the Falcon 9 rocket lifted off from Vandenberg Space Force Base in California, and TechCrunch said it carried more than 100 other spacecraft.

Travis Beals, the Google senior director who leads the project, wrote in a company post that day that the team had made contact and that the satellite was “operating as expected.”

According to TechCrunch, about three weeks of commissioning followed, as engineers checked the spacecraft’s subsystems. This week they sent the first prompt to Gemma 4, a Google DeepMind AI model loaded on the satellite before launch. It answered: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Google also tried space-themed limericks and asked the spacecraft about conditions in orbit, and a spokesperson told TechCrunch it was operating nominally.

The hardware is modest. The satellite carries four Tensor Processing Units, or TPUs, the custom AI chips Google also uses in its data centers on the ground, and generates up to 1 kilowatt of power. TechCrunch noted in a correction that only one chip can run at a time. Accounts of how long each can run differ slightly: TechCrunch reported about 20 minutes, limited by power demands, while NPR reported 15 minutes at a time because of heat. Beals told NPR the mission is “a very minimal test,” with a goal of about a year of operation.

Why it matters

The experiment comes as technology companies spend heavily to keep up with demand for AI. Sundar Pichai, Google’s chief executive, said at the company’s developer conference in May that demand exceeds supply and projected 2026 capital spending of $180 billion to $190 billion, more than six times the 2022 level, NPR reported. Data centers on Earth, meanwhile, face growing opposition over their power use.

Google is not alone, and it was not the first to run AI in orbit. Starcloud, a venture-backed startup, launched a spacecraft last November carrying an Nvidia H100 chip and demonstrated a version of Google’s Gemini model from space, according to NPR. SpaceX, led by Elon Musk, says it expects to deploy “orbital AI compute satellites” as early as 2028, and Musk has called space the “only way to scale” data centers.

The designs differ. Google wants clusters of satellites flying close together and sharing large workloads over laser links. TechCrunch reported that SpaceX’s spacecraft are also expected to use lasers but would orbit too far apart to share processing effectively, while Starcloud and another company, Cowboy Space, expect to run AI tasks on individual satellites.

The hurdles

Scientists and Google’s own researchers describe several unresolved problems:

• Radiation: Charged particles can cause “bit flips” that silently corrupt data without tripping typical monitoring tools. TechCrunch reported that particle accelerator tests suggested minimal effects on running AI models but a bigger problem for training them. The orbital data will test that finding.
• Heat: With no air or water in orbit to carry heat away, more computing means more heat in a confined space, Brandon Lucia, an electrical and computer engineering professor at Carnegie Mellon University, told NPR. Beals said radiators are among the mission’s heaviest components, and added weight raises launch costs.
• Repairs: Fixing hardware in orbit is far harder than in a building on the ground. Lucia said costs and complexity could grow “by a factor of 10, maybe a factor of 100.”
• Launch costs: TechNews reported that Google says prices must fall to roughly $200 per kilogram by 2035 for the idea to make economic sense. A peer-reviewed analysis by Suncatcher researchers estimated that SpaceX would need to fly Starship, its heavy-lift rocket still in development, about 1,800 times to reach cost reductions like those seen with its earlier rockets, TechCrunch reported.

Rocket fuel emissions and possible pollution from satellites burning up on reentry are also concerns, NPR reported. Alphabet, Google’s parent company, holds a significant investment in SpaceX, according to TechCrunch.

What comes next

Testing will now focus on how the chips handle thermal stress and radiation. If they perform well, Google plans to launch two more satellites in 2027 with more advanced TPUs, greater power generation and optical links between spacecraft, TechCrunch reported. TechNews said the company’s envisioned final design involves more than 80 satellites in tight formation.

Even Google’s project lead is cautious about the timeline. Beals told NPR he does not expect space data centers to be cheaper than those on Earth within five years and expects the two to coexist for a long time. He described the project as a “moonshot,” and TechCrunch reported that it aims to make orbital computing economical in the 2030s.