Google is about to do something genuinely historic: launch a data center into space. The catch? It can only work for 15 minutes at a time before it overheats. In space. The place that is famously cold. Science has never felt more personal.
Welcome to Project Suncatcher, the plan that makes "cloud computing" suddenly, hilariously literal. On October 1, a refrigerator-sized satellite called MVP will lift off from Vandenberg Space Force Base aboard a SpaceX Falcon 9 rideshare. Its cargo: four of Google's Trillium TPU accelerators, ready to run Gemini from low Earth orbit.
The Fridge That Thinks
Planet Labs, the imaging company, supplied the satellite body and Google dropped its chips into a frame that already existed. That is the aerospace equivalent of buying a used van and bolting a supercomputer into the back.
Here is the spec sheet, and it is magnificent:
- Size: about a refrigerator. Not a server rack. Not a modular pod. A fridge.
- Chips: four Trillium TPUs, the same accelerators crunching numbers in Google's Earth-bound data centers.
- Power: roughly one kilowatt of solar, which 24/7 Wall St. translates as "enough to run a microwave or a hair dryer."
- Work schedule: 15-minute bursts of AI inference, followed by a cooling timeout while the radiators catch up.
Let that sink in. Every other tech giant is buying gigawatts and building campuses the size of small towns. Google's first orbital data center has the power budget of a kitchen appliance and the work ethic of a student pulling an all-nighter in 15-minute sprints.
Physics Says No Hot Air in Space
The overheating problem is not a design flaw, it is an occupational hazard of orbit. AI accelerators dump enormous heat into a tiny area, and on Earth we simply blow air over them. In space there is no air, so the heat has nowhere to go. The radiators struggle, the chips throttle, and the whole experiment clocks out for the day.
Google says its engineers have been testing the hardware in a thermal vacuum chamber that simulates both the temperature and the airless void of orbit. The good news: the Trillium chips survive radiation doses larger than a five-year space mission would throw at them. Reassuring, right up until you remember there is zero radiation shielding on board and one flipped bit can corrupt a calculation in a blink.
The cooling fix is a soft thermal interface material linking the chips to aluminum and copper heat pipes that feed a radiator. Basically the radiator in your car, except your car never has to track a coin-sized target from miles away at orbital speed.
Oh yes, about networking: the endgame is a constellation of satellites linked by lasers. Google says holding those links requires extraordinary precision, the laser equivalent of threading a needle during an earthquake. The two-satellite laser test is scheduled for 2027.
Space Is the New Server Rack
Why go through all this trouble? Because on Earth, AI's appetite is becoming a political problem. Data centers already swallow 4 to 5 percent of US electricity and could reach 17 percent by 2030, according to EPRI. Communities are suing, banning, and flat-out rejecting new facilities. Orbit, by comparison, offers near-constant sunlight with up to eight times more solar energy than the ground, free cooling, and zero zoning battles. NIMBYs cannot stop you in orbit, but physics absolutely can.
Google's senior vice president James Manyika is refreshingly honest about the timeline: "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years." He compared the project to self-driving cars, which took Google 15 years to make real. So the fridge in the sky is a science experiment with a very long runway, and Google seems fine with that.
The long-term sketch is gloriously ambitious: roughly 81 satellites packed within a one-kilometer radius at about 650 kilometers of altitude, a constellation behaving like one hyper-local cluster. It is a theory on a whiteboard today, but it is a theory with four TPUs on a launch manifest.
Analysts project the space data center business at $28 billion by 2040, though rival estimates range from $3.8 billion to $8.4 billion. Forecasts for a market that barely exists are guesses dressed in decimal points. Google has set no official target, which may be the closest thing to humility this industry has ever shown.
Google is not alone in the orbit play. Starcloud launched a test satellite last December and raised $170 million this year, and both SpaceX and Blue Origin have floated their own orbital compute ambitions. The space race for AI is officially a thing, and it fits in a kitchen.
Here is the scoreboard, if you are keeping one at home:
- Earth data center: works 24/7, needs 100 megawatts, needs a zoning permit, needs a friend in city hall.
- Orbital data center: works 15 minutes, needs 1 kilowatt of sunlight, needs zero permits, overheats anyway.
Nobody is pretending Suncatcher is ready to host your next big model. It is a proof of concept, an expensive one, designed to find points of failure so future missions can avoid them. As Google puts it, this first launch is about seeing what works. Sometimes what works is humbling.
Still, give credit where it is due: Google found a way to make a data center that is both literally in the cloud and, at 15 minutes of work per day, somehow still more productive than a spreadsheet full of humans staring at the same numbers. When the satellite finally stays cool long enough to answer a query, it will deliver the most expensive "let me think about that for a sec" in human history.
The launch window for this beautiful disaster is October 1. Grab the popcorn, keep an eye on the sky, and remember: every great leap for mankind starts with a small step, a bigger ego trip, and a fridge that refuses to work a full shift.
Comments