Project Suncatcher: Google bets Gemini can run on sunlight, 15 minutes at a time

MVP, a fridge-sized satellite built by Google and Planet, carries four Trillium TPUs into orbit to run Gemini on solar power. The chips can only work for about a quarter of an hour before they have to stop and shed heat, and that limit is the real subject of the experiment.

Google’s first artificial intelligence satellite will spend part of its working life switched off on purpose. MVP, the opening flight of Project Suncatcher, carries four Tensor Processing Units that will run Gemini models for roughly 15 minutes, stop, wait for a radiator to push their heat out into space, and then start again. The pause sounds like a technical footnote. It is the question Google is sending the satellite up to answer.

Project Suncatcher is Google’s research effort to find out whether the electricity behind AI could come straight from the sun, collected in orbit, instead of from the same power grids that homes, hospitals and factories depend on. MVP is the first hardware to fly under it. Built with Planet, the San Francisco company that runs one of the largest fleets of Earth-imaging satellites, it is roughly the size of a refrigerator and draws about one kilowatt from its solar panels. Its four chips belong to the Trillium generation, the same TPU family that answers Gemini requests inside Google’s data centers on the ground.

If the schedule holds, MVP lifts off on October 1 aboard a SpaceX Falcon 9 from Vandenberg Space Force Base in California, one of many small payloads sharing the Transporter-18 rideshare mission. Google has framed the flight with unusual modesty. “Some things can only be tested in space,” the company said in the days before launch. “Putting our first TPUs in orbit next week will help us get data and learnings to inform future launches.” There is no product attached and no service waiting for the result. What comes back down is measurement data.

Why Google wants to compute where the sun barely sets

The case for going up is sunlight. In the right orbit, a solar panel can be up to eight times more productive than the same panel on the ground, according to Google’s own research, because there are no clouds, no weather and almost no night. Suncatcher relies on a dawn-dusk sun-synchronous orbit, a path that rides the boundary between day and night so that a spacecraft sees the sun nearly all the time. That also cuts the need for the heavy batteries a satellite would otherwise carry through the dark part of each orbit, and batteries are mass, and mass is the most expensive thing in spaceflight.

The pressure behind the idea is easy to see on Earth. Every new generation of AI models demands more computing, and every data center built to supply it needs a grid connection, land, water or another way to cool it, and increasingly the patience of local communities worried about their own electricity bills. The International Energy Agency estimated in 2025 that electricity use by data centers worldwide is on course to more than double by the end of the decade, to around 945 terawatt-hours, with AI as the main driver of the increase.

The largest technology companies have answered with a scramble for power. Google has signed agreements to buy electricity from small modular reactors developed by Kairos Power and, with NextEra Energy, to bring the shuttered Duane Arnold nuclear plant in Iowa back into service. Microsoft agreed to restart a reactor at Three Mile Island in Pennsylvania, and Amazon has invested in the reactor developer X-energy. All of those projects take years, depend on permits and local politics, and compete for the same transmission capacity everyone else wants. Suncatcher is the most radical version of the same search: instead of finding more power on the ground, go where the sunlight hardly stops.

A satellite the size of a fridge with the power of a microwave

For now the ambition is small by design. MVP’s solar panels deliver about one kilowatt, roughly what a microwave oven draws, and its four chips together add up to about the computing power of a single server in a ground data center. The largest AI campuses on Earth are planned in hundreds of megawatts. On paper, a single 100-megawatt facility would need the power budget of a hundred thousand satellites like MVP. Google frames the mission as a plain question that has to be answered before any of the larger arithmetic matters: can its AI hardware operate in space at all?

MVP is also a sign that the project is moving faster than first announced. When Google introduced Suncatcher in November 2025, the plan was a learning mission with Planet that would put two prototype satellites into orbit by early 2027. MVP now goes first, ahead of that pair, and its job is narrower. It is there to collect real data on how the TPUs behave under launch stress, space radiation and the sharp temperature swings of low Earth orbit before Google commits to more complex hardware.

Trillium is Google’s sixth-generation TPU, announced in May 2024 and opened to cloud customers later that year. Google has since introduced a seventh generation, Ironwood, designed for running large models at scale. Flying the previous generation has a practical logic for a first test: the chip is thoroughly characterized on the ground, so any change in its behavior in orbit can be traced to the environment rather than to unfamiliar silicon.

The partnership with Planet has history. In 2017 Google sold its satellite imaging business, Terra Bella, formerly known as Skybox Imaging, to Planet, and took an equity stake and a multi-year agreement to buy imagery as part of the deal. Planet, founded in 2010 by former NASA scientists, has spent more than a decade building and operating fleets of small satellites, so it brings the spacecraft side of the work: building them, keeping them alive in orbit and talking to them from the ground. Google brings the chips, the models and the questions. That split matches how the company has described Suncatcher from the start, as research that learns from each launch before committing to the next.

Radiation turned out to be the easier problem

The ground tests give some reason for optimism. Google shook the hardware along all three axes to mimic a rocket ride, during which individual components can experience 50 to 100 g. It then took Trillium TPUs to the Crocker Nuclear Laboratory at UC Davis and fired a 67 MeV proton beam at them while they ran AI workloads, watching for flipped bits and other errors. According to Travis Beals, senior director of Google’s Paradigms of Intelligence team, the chips withstood a total radiation dose greater than they would absorb over a five-year mission.

The published figures go further than that summary. Google’s research paper puts the expected five-year mission dose at 750 rad(Si). The chips’ high-bandwidth memory, the most sensitive component, only began to show irregularities after 2 krad(Si), nearly three times that figure. The processors were tested up to 15 krad(Si), twenty times the expected dose, and no chip suffered a hard failure attributable to accumulated radiation. That does not make space harmless for computers. Errors caused by individual particles striking the chip, known as single-event upsets, still have to be detected and corrected in software and memory.

The result matters because the space industry’s traditional answer to radiation is slow and expensive. Radiation-hardened processors such as the RAD750, which flew on NASA’s Curiosity and Perseverance rovers, are built to survive for many years in harsh environments but run at a small fraction of the speed of modern commercial chips, and they cost far more. Off-the-shelf computers have flown before, among them the Spaceborne Computer that Hewlett Packard Enterprise sent to the International Space Station in 2017. Google’s numbers suggest that commercial AI accelerators, protected by error correction rather than by special manufacturing, could be good enough for orbital work.

Heat is the bill that comes due every 15 minutes

Heat is the harder problem, and the 15-minute limit shows it plainly. On Earth, data centers blow air or pump water across chips to carry heat away, and much of a facility’s design revolves around that task. In a vacuum there is no air to move heat by convection, so MVP carries it through aluminum and copper heat pipes to a radiator that can only release it as infrared light.

Physics is unforgiving here. A radiator sheds heat in proportion to its surface area and to the fourth power of its temperature, so the only ways to dump more heat are a bigger radiator or a hotter one, and chips do not like running hot. The International Space Station handles the same problem with large external radiator panels fed by ammonia loops. On a spacecraft as small as MVP, the chips produce heat faster than the radiator can release it. So they work until they approach their thermal limit, and then they wait. Every minute added to that working window will be a direct measure of how well the thermal design is doing.

Thermal management is only one of the hurdles Google lists for a full orbital data center. A real system would also need high-bandwidth links to send results back to the ground, where the users are, without delay becoming a problem. It would have to keep running with parts nobody can replace: when a server fails on Earth, a technician swaps it; when one fails in orbit, it stays broken, so the design has to plan for a slow, steady loss of hardware. And it would depend on launch prices far below today’s. Google has not said how much the project costs.

Everything depends on the price of a kilogram

That last point is where the idea lives or dies. Google’s analysis, which projects the historical decline in launch prices forward, concluded that computing in orbit would only become roughly comparable to the energy costs of a ground data center once launches fall below $200 per kilogram. Google expects that could happen around the mid-2030s. Today’s rideshare prices are far higher: SpaceX’s Transporter program, the service MVP is flying on, has advertised rates of several thousand dollars per kilogram for small satellites. Closing that gap depends on fully reusable heavy rockets such as SpaceX’s Starship delivering the cost curve their builders promise.

The date explains the tone of the whole project. Suncatcher is a bet on where the rocket industry will be in a decade, not a product plan for next year. Google calls it a moonshot, the label it has used for long, uncertain bets such as self-driving cars, the work that became Waymo, and quantum computing. Both took more than a decade to produce anything a customer could use.

81 satellites flying a few hundred meters apart

The published design shows how far the ambition reaches. The paper, signed by Travis Beals with Google researchers including Blaise Agüera y Arcas and James Manyika, does not imagine one giant spacecraft. It describes clusters of 81 satellites flying in formation within a radius of one kilometer, at an average altitude of about 650 kilometers, with neighbors only 100 to 200 meters apart. Flying that close is what would let them behave like a single data center, linked by laser beams instead of fiber-optic cables. In a bench-scale test on the ground, Google’s researchers reached 800 gigabits per second in each direction, 1.6 terabits in total, with a single pair of optical transceivers. A working cluster would need tens of terabits per second. In 2027, Google plans to fly two satellites with Planet to test exactly those laser links in orbit.

Formation flying at that density is an open problem of its own. Satellites in low orbit are nudged constantly by the thin upper atmosphere and by the uneven pull of Earth’s gravity, and keeping 81 of them in a tight cluster would require continuous, automated station-keeping. Google’s paper includes models suggesting that modest maneuvers would keep the cluster together, but MVP, a single spacecraft, cannot test that. A dense cluster would also enter a debate that predates Suncatcher about crowding in low Earth orbit, where thousands of active satellites, most of them SpaceX’s Starlink, already share the sky.

Nothing changes for Gemini users yet, in any market. Every answer still comes from a data center on Earth, and MVP’s output is engineering data, not a new service. Nobody will be chatting with a satellite this year, and no Gemini feature depends on how the mission turns out.

Google is not alone above the clouds

Google is not the only company in this race, and it is not the first to fly AI chips. Starcloud, an Nvidia-backed start-up based in Redmond, Washington, and originally known as Lumen Orbit, launched its Starcloud-1 satellite in November 2025 with an Nvidia H100 graphics processor on board. In December of that year it ran Gemma, Google DeepMind’s open model, in orbit and trained nanoGPT, a small language model created by Andrej Karpathy. In China, the satellite maker ADA Space and Zhejiang Lab launched the first 12 satellites of what they call the Three-Body Computing Constellation in May 2025, with the stated aim of processing data in orbit rather than sending it all to the ground.

SpaceX has outlined orbital AI data centers built with Nvidia technology that could launch by late 2027, Yahoo Finance reported, and its AI1 satellite design targets about 120 kilowatts of sustained computing, more than a hundred times MVP’s power budget. Jeff Bezos has predicted gigawatt-scale data centers in orbit within one or two decades, and Sam Altman of OpenAI has also backed the idea of computing in space. The list of believers is long; the list of working systems is, for now, a handful of single satellites.

What sets Google apart is that it designs its own chips, trains its own models and runs one of the largest data center fleets in the world. If Suncatcher works, Google would not need anyone else’s processors to scale it. That is also why a small satellite with four chips matters to the company: every measurement MVP sends home feeds directly into the design of the next TPUs and the next spacecraft.

Once in orbit, MVP will fly a dawn-dusk sun-synchronous path and is expected to run tasks for about a year, although the hardware could stay up for as long as six years before it falls back into the atmosphere and burns up. The first useful numbers should come early, in how the chips came through the launch and how they behave during their first cycles of work and rest.

The figure to watch after launch is not how fast MVP computes but how long it can keep computing. If those 15 minutes stretch toward an hour, a data center in orbit starts to look like an engineering problem rather than science fiction. If they do not, AI’s power bill stays firmly on Earth, and the answer to it will have to come from grids, power plants and more efficient chips on the ground.

Tags: Google, Sam Altman, SpaceX, Gemini, data centers, Planet