Google's four-TPU MVP satellite tests whether radiation and heat rule out orbital AI compute

The cyclotron result is a claim about one particle source. For it to hold in orbit, the Crocker beam has to resemble the radiation the satellite actually meets, and the word "usually" in Google's finding that restarts fixed the flips has to become a measured rate ^[6]. A bit flip turns a stored zero into a one, or the reverse, and shows up as an error in software or in a calculation ^[7]. A restart clears only the errors something noticed, and the reports do not say how MVP detects a flip before it becomes a visible fault.

Restarts also cost time on a machine that already works in short bursts. Heat leaves the TPUs by conduction. The chips sit on a motherboard against sheets of thermal interface material, aluminum and copper layers pull the heat away, and a radiator panel emits it to space ^[9]. There is no air in orbit, so there are no fans ^[8]. About 1 kW shared across four chips caps each at roughly 250 W, even if nothing else drew power ^[23]. A forced cool-down after every run means the radiator sheds heat more slowly than the chips make it at full load. The metal layers absorb the difference until they reach their limit, and Gemini requests wait until the stack cools ^[14].

Now scale it with Beals' own estimate. One gigawatt across 10,000 satellites is about 100 kW each ^[21], roughly a hundred times MVP's power budget ^[22]. At that power, a run-and-cool cycle would leave launched hardware sitting idle between runs. According to The Decoder, launch costs would also have to fall to about $200 per kilogram before the economics work ^[24]. Jeff Bezos thinks orbital data centers could take up to 20 years to beat ground sites on cost ^[20]. Google has not disclosed how much it is spending on the project ^[18].

I think the sequencing is right. Google had planned two satellites for 2027 and instead installed its hardware in an existing satellite from Planet Labs, a company it has invested in ^[17]. The refrigerator-size craft was built and tested at Planet Labs in San Francisco ^[10], so the unknowns reach orbit on a platform Google did not have to design first. Technicians marked screws to see whether they shifted, a check about as old as screws, then shook the satellite; the chips stayed intact and no screws loosened ^[15].

MVP is the first orbital test of Project Suncatcher, and Google is treating the launch as an early experiment ^[2][11]. "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years," James Manyika, Google's senior vice president for research, said ^[3]. He compared the effort to Google's driverless cars, which it was testing "like 15 years before anything showed up," Tom's Hardware reported ^[5]. MVP is expected to operate for a year and could stay in orbit for up to six before burning up on re-entry ^[16]. The concepts after it are already drawn: two more satellites next year, formations of more than 80 spacecraft, and a custom satellite as long as a soccer field ^[12].