Google's Suncatcher satellite is up, with four chips and one kilowatt

Google's first Project Suncatcher satellite is in orbit, carrying four tensor processing units, about the computing power of one data-centre server. The New York Times reported that the satellite, named MVP, was set to fly on a SpaceX Falcon 9 from Vandenberg Space Force Base on 1 October. The Register confirmed on Friday that the satellite had reached space. Solar panels supply about one kilowatt, roughly a hair dryer's worth, to the chips.

The test is narrow. The four processors are meant to run Google's Gemini models for about 15 minutes at a time, then shut down and cool. MVP is supposed to answer simple queries and operate for a year, though it may circle Earth for up to six years before drag pulls it down to burn up. James Manyika has said Google's expectations for this first craft are tempered. The company has also cast doubt, in its own research note, on whether the idea can scale.

The paper, "Toward a future space-based, highly scalable AI infrastructure system design," puts a number on that doubt. SpaceX and others have said orbital data centres start to make sense when launch cost falls to about 200 dollars a kilogram. Google's researchers note that SpaceX has cut the cost of launching a kilogram by about 20 percent with every doubling of the mass it has already launched. Keeping that rate would require about 370,000 extra tonnes of cumulative mass, on the order of 1,800 successful Starship flights, plus reuse of components about 100 times. Only then, the paper says, is 200 dollars a kilogram plausible under reasonable assumptions.

Google does not yet know the mass of a future data-centre satellite. It uses Starlink's second-generation craft, about 575 kilograms, as a comparison point. The conclusion in the paper is conditional: if launch cost reaches about 200 dollars a kilogram or less, the annualised cost per unit of power in space could sit near what operators spend on the ground. That is a cost comparison, not a claim that the heat, the radiation, or the network problems are solved.

Those problems are the part a four-chip test does not touch. A ground data centre rejects heat into air and water. A satellite rejects it by radiation into vacuum, which is why the Gemini runs are capped at 15 minutes. Chips also take radiation hits that servers in Oregon do not. Linking thousands of satellites into one training run needs formation flying and a network that does not exist at that density. The Register's summary is blunt: orbital scale needs SpaceX to fly on the order of 1,800 Starships, and it needs researchers to solve networking, design and formation flying.

The commercial context is SpaceX's own pitch. When Elon Musk took SpaceX public in June, data centres in orbit were part of the offering. Sceptics listed power, heat, launch cost and latency. Google's flight does not settle that argument. It puts four TPUs above the atmosphere on a rocket Musk's company flew, with one kilowatt of solar power and a 15-minute duty cycle.

Vandenberg, rather than Cape Canaveral, fits a sun-synchronous or high-inclination test better than a crewed-station launch. The Times had the satellite on a Falcon 9 there on 1 October, loaded with four TPUs and enough power for simple queries. Friday's confirmation that it is up means the radiation and thermal legs of the test can start. Those legs are measured in weeks. A passing grade would still leave the 200-dollar launch-cost condition, and the 1,800 Starship equivalent behind it, untouched. Google has put the hardware in orbit. It has not put a data centre there.