Google's plan to build AI data centers in space leaves the drawing board next week. On October 1, a refrigerator-sized satellite called MVP will lift off from Vandenberg Space Force Base in California aboard a SpaceX Falcon 9, carrying four of Google's Tensor Processing Units into low Earth orbit. It will spend about a year there answering the question behind Project Suncatcher: can the chips that train and serve Gemini survive, and do useful work, beyond the atmosphere?
The launch, first reported by The New York Times on September 24 and confirmed the same day in a Google blog post, is the program's first hardware in orbit. It comes months ahead of the schedule Google set when it unveiled Suncatcher in November 2025, when the company said its first two prototype satellites, built with Planet Labs, would fly by early 2027.
"Can our TPUs survive and operate in space? Well, we're going to find out," Google CEO Sundar Pichai wrote on X on Thursday, adding that the project was "hitching a ride" on SpaceX's Transporter-18 rideshare mission. SpaceX CEO Elon Musk replied with two rocket emojis.
A Server-Sized Test
MVP is modest by design. Its four TPUs deliver roughly the computing power of a single data-center server, and its solar panels produce about one kilowatt. Planet, which has spent more than a decade building small Earth-imaging satellites, co-developed the spacecraft, and Google's AI hardware rides aboard it. Google has not published the satellite's mass or dimensions.
The mission is essentially a survival test, and the ground data so far has been encouraging. In a post published September 24, Travis Beals, the Google Research senior director who leads the Paradigms of Intelligence group, wrote that the ride to orbit lasts about 10 minutes and puts the spacecraft through acceleration of up to 10 g. Individual components such as the TPUs can feel 50 to 100 g. Google shook the satellite along all three axes to mimic a launch. "Tests like this rarely go as planned, so we were pleasantly surprised that the hardware held up to the force," Beals wrote.
Radiation is the next hurdle. Google ran AI workloads on its sixth-generation Trillium TPUs while firing a 67 MeV proton beam at them at UC Davis's Crocker Nuclear Laboratory. According to Google's published research, the chips, including radiation-sensitive high-bandwidth memory, computed without errors up to a cumulative dose of 2 krad(Si). That is well above the roughly 750 rad(Si) a shielded satellite would absorb over a five-year mission. Google saw no permanent failures even at 15 krad(Si). Ground beams can't fully reproduce orbit, though, which is why the hardware has to fly.
Heat may be the tougher problem. With no air in a vacuum, the TPUs have to shed heat through heat pipes and radiators. On this first flight that system has limits: the chips are expected to run Gemini workloads for about 15 minutes at a time and then pause while the thermal system catches up. "This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions," Beals wrote.
Why It Matters
Suncatcher exists because AI's biggest bottleneck is increasingly electricity rather than chips. Hyperscalers are competing for grid connections, land, and water, and Google is simultaneously pursuing geothermal power with Fervo Energy and small modular reactors with Kairos Power. Google says a satellite in a dawn-dusk sun-synchronous orbit, around 650 kilometers up, gets near-constant sunlight and can generate up to eight times more solar energy per panel than one on the ground.
The long-term design is far more ambitious than one satellite. Google's research describes clusters of 81 satellites within roughly a one-kilometer radius, flying 100 to 200 meters apart and linked by free-space optical lasers. Ground tests have already reached 1.6 terabits per second per transceiver pair. The economics depend on falling launch prices. A paper co-authored by Blaise Agüera y Arcas, Beals, and James Manyika estimates that launch to low Earth orbit could drop below $200 per kilogram by the mid-2030s. At that price, Google argues, orbital compute could compete per unit of compute with the power and cooling costs of ground data centers.
The competitive picture is awkward. SpaceX, which is carrying Google's hardware, is pursuing orbital AI compute of its own and has described an initial AI1 satellite design with about 120 kilowatts of sustained compute. That is more than 100 times MVP's power budget. Pulling its first flight forward by several months lets Google show real hardware in a race where most of the competition so far consists of FCC filings and slide decks. Skeptics point out that radiators, orbital congestion, collision risk, and the impossibility of in-orbit repairs remain unsolved at scale, and the U.S. Government Accountability Office has flagged several of the same concerns.
What to Watch
The first milestone is simple: a clean Transporter-18 launch on October 1, followed by confirmation that all four TPUs power on. After that, the useful signal will come from any telemetry Google chooses to release on bit-flip rates, thermal cycling, and how the 15-minute duty cycle holds up over months. Those numbers will shape the two Planet-built satellites planned for 2027, which are meant to prove the laser links an 81-satellite cluster would depend on. If MVP runs its full year without major failures, expect Google to compress its roadmap again. If it doesn't, the question of whether orbital compute is a moonshot or a mirage will get considerably louder.
“Can our TPUs survive and operate in space? Well, we're going to find out.”— Sundar Pichai, CEO, Google