Google will put its own AI chips into space next week, and the hardware is modest. The satellite is called MVP. It’s about the size of a refrigerator and carries four Tensor Processing Units, the same chips Google uses in its ground data centers. A SpaceX Transporter-18 rideshare will take it up.
Plenty of people will see the specs and roll their eyes. A single AI rack on the ground pulls tens of kilowatts, and operators are already planning for racks that get close to a megawatt. MVP’s solar panels produce around one kilowatt, which is about what a microwave oven uses.
I’d argue the size is beside the point. Google says as much in its announcement: “Some things can only be tested in space.” This flight is about finding out which of those things will break.
Fifteen minutes on, then a cool-down
The spec I keep coming back to has nothing to do with raw compute. Google plans to run Gemini on the TPUs in short bursts of about 15 minutes each, then turn the chips off so the radiators can get rid of the heat.
Most pitches for orbital data centers lead with energy, and they have a point. Solar panels above the atmosphere get more sunlight and don’t lose it to nightfall. In orbit, collecting power turns out to be the easy half. Dumping heat is the hard half.
Ground data centers get rid of heat through air and water, and that water use is a big reason neighbors show up at zoning meetings to oppose them. Space has no air or water to carry heat away. Google is using heat pipes and radiators to draw heat off the chips and radiate it outward, and it still has to show that setup works in orbit. When four chips need a break every quarter hour, I’d guess the radiator is the limiting part. Selling “unlimited solar” doesn’t mean much if the machine can only use that power in short bursts.
MVP is really a cooling test with some AI chips attached. Success depends on the satellite surviving launch vibration and on a cooling system that works without air or water doing what the engineers expect.
Why the schedule moved up
Google wasn’t planning to fly anything this soon. The original plan called for two purpose-built satellites in 2027, but the company decided to put its chips on an existing satellite and launch earlier. If the TPUs come through intact, the 2027 two-satellite test is still the next big step.
The competition explains the hurry. Starcloud, a startup backed by Nvidia and Y Combinator, launched its first AI-equipped satellite in December 2025. SpaceX isn’t waiting either. Its first Starmind satellite, AI1, is supposed to carry a space-hardened version of Nvidia’s Vera Rubin NVL72 rack system, and the company wants to start launching AI satellites in late 2027.
Google seems to have decided that messy flight data now is worth more than a perfect prototype two years from now. Owning its own chip design helps a lot with that. Google doesn’t have to wait for a supplier to give permission before putting TPUs on a rocket.
Google’s odd position: backer, customer and competitor
Most of the coverage leaves this part out. Google is racing SpaceX, and it is also financially tied to SpaceX in several ways.
Google owns 6.1% of SpaceX. SpaceX is also launching Google’s satellite. On top of that, Google rents compute from SpaceX here on Earth, under a multiyear agreement for roughly 110,000 Nvidia GPUs at Colossus 2 that would cost $920 million a month once fully ramped.
So Google is SpaceX’s investor, launch customer, compute customer, and potential competitor all at once. Relationships like this are common in the industry right now, and we laid out several of them in our look at who really controls AI. The same company can be a customer at one layer, an investor at another, and a rival at a third. Orbit adds another layer where that can happen.
It also explains why Google can’t afford to sit this one out. If orbital compute ever works at scale, whoever owns the rockets would also control access to the best locations for it.
The economics are still a decade away
Google has been clear about what would need to happen for this to pay off. Its Suncatcher paper projects that launch costs could fall below $200 per kilogram by the mid-2030s, and at that price, operating in orbit could roughly match a ground facility’s energy costs per kilowatt-year. That’s a forecast about the 2030s, and nobody should read it as a business plan. Google also hasn’t said how much it’s spending on the project.
Leadership is talking about long timelines too. Sundar Pichai has said that roughly ten years from now, building data centers in space may start to look routine. Heat is only one of the technical problems. A cluster in orbit would need networking comparable to a ground data center, which means keeping fast-moving satellites precisely aligned while their orbits slowly drift. Google’s longer-range research describes about 81 satellites within a kilometer of one another, connected by laser links so they operate as one computer. And if a chip dies 500 kilometers up, nobody has a good way to replace it yet.
Some workloads fit orbit better than others. JLL analysts say training and batch jobs, along with processing data generated in space, can tolerate delays, while real-time inference will probably stay on the ground close to users. MVP happens to be running inference, which works for a test but probably not for a business. The more realistic early use is processing satellite imagery before it’s sent down, the same logic behind moving compute onto factory floors with industrial edge computing.
The case for trying anyway
Nobody would pay for this if building data centers on Earth were easy, and it’s getting harder. According to the IEA, natural gas supplied over 40% of US data center electricity in 2024, with wind and solar around 24%, nuclear near 20%, and coal about 15%. Electricity bills are rising faster than inflation in many parts of the country; some places blame data centers for that, and local opposition is disrupting the plans of the largest tech companies. By July 2026, community pushback had stopped around $130 billion in AI data center projects.
That’s Suncatcher’s real rationale. Space isn’t clearly better, but building on Earth keeps getting harder. CNBC made a version of this argument: the best long-term case may be that ground construction keeps getting pricier while land, water and power remain political fights, even as launches get cheaper. When your compute costs run into the trillions, a long-shot bet that avoids grid queues and water permits can make sense as a hedge.
Three numbers to watch
Wall Street barely reacted. Alphabet rose 0.5% in early trading after the news. That seems about right to me, because the launch matters less than the data that comes back afterward.
First, watch how long the TPUs can run before they have to shut down to cool, and whether that window gets longer. Second, watch how the chips handle months of radiation, since a few days in orbit won’t tell us much. Third, watch whether the 2027 mission can actually connect hardware across satellites.
If fifteen minutes becomes several hours, orbital AI becomes a serious conversation. If it stays at fifteen minutes, we’ll have learned, at considerable expense, that cooling was always going to be the hard part.
Related: Who Owns Artificial Intelligence? Ownership vs Control
