On October 1, 2026, Google, in partnership with Planet Labs, is preparing to launch its first experimental satellite named MVP under Project Suncatcher. The launch will be carried out using a SpaceX Falcon 9 rocket from Vandenberg Base in California.  

In reality, the first MVP satellite is very tiny. It carries only 4 tensor processors from Google, which in terms of computing power is equivalent to just one standard ground server.  

The spacecraft will be powered by solar panels that generate only 1 kW of energy, which is like a regular household hairdryer.  

The solar panels displayed on the test stand of Google’s MVP ("SUN V1") satellite. Once in orbit, these wings will unfurl to collect sunlight for powering AI chips. Photo: Philip Cheung for The New York Times.

There is no air in a vacuum, so ordinary coolers or liquid cooling do not work. Therefore, a special multi-layered thermal conductor system by Google will be used, which allows the chips to run continuously for only 15 minutes, after which they will have to be turned off for cooling. But it is worth specifying here that Google plans to launch around 80 or even more satellites that will pass a so-called relay race. When one satellite heats up and goes to cool down, it transfers the task to a neighboring satellite using laser inter-satellite communication channels. The calculations do not stop, they will simply be transferred along the orbit.   

Google’s experimental MVP ("SUN V1") AI satellite on a test rig at the Planet Labs facility. Photo: Philip Cheung for The New York Times.

The greatest danger remains cosmic rays, which cause glitches in the code. Instead of expensive heavy protection, Google plans to simply reboot the chips remotely in case of errors.  

And the most important question is what is all this for and why exactly space? As they state, satellites in low Earth orbit can receive 8 times more solar energy than on Earth. This could potentially solve the problem of the massive power consumption of AI on Earth, which currently overloads power grids. Google itself predicts that assembly-line satellite production and cheaper launches from SpaceX will allow space data centers to match ground ones in cost by the mid-2030s. Ground data centers require billions of dollars for land acquisition, laying power lines, and taxes. In space, everything is free, and the sun shines almost around the clock. Also, a data center in space does not fall under the jurisdiction of any specific country. It cannot be shut down by a court decision, have its servers seized, or be de-energized due to a political conflict. 

In order not to idealize Google and not to give them extra respect points, I suggest writing a bit of negative text about this approach to data centers, namely that transmitting data from Earth to space and back in low orbit takes time. This means that such an AI will not be suitable for instant chat responses like ChatGPT, Gemini, or Claude, but rather for climate analysis, molecular biology, or training new large models. In short, it is simply for heavy tasks that require time, and also where an instant response is not needed, but pure and high-quality analysis is.

A Planet Labs engineer seen through the mirrored panel of the experimental satellite during final assembly. Photo: Philip Cheung for The New York Times.

And how can we go without one of the biggest problems that the space industry constantly faces, which is debris. The emergence of fleets of hundreds of satellites increases the risk of collisions in orbit, and also if an AI satellite fails due to radiation and does not reboot, it will turn into an unguided pile of metal.  

Thanks for reading!