On October 30, someone sent a prompt to a satellite. The model running on the chips aboard it answered as the satellite passed over Norway, and the answer came back as a haiku about the blue marble below and starlight feeding its core. It is a strange thing to read. It is also not a demo or a concept video. There is hardware in orbit right now, running an AI model, and it has been up there since the first of October.
The project is called Suncatcher, and Google calls it a moonshot. That word has a specific meaning inside Google. A moonshot is an enormous problem paired with a solution that sounds absurd, tested seriously, with the full expectation that most attempts will fail. It is a label that buys an idea permission to look ridiculous for a while.
This one has stopped looking ridiculous faster than almost anyone expected.
The idea is to put TPUs, Google's in-house AI chips, on solar powered satellites. The orbit matters as much as the hardware. These satellites ride the line between day and night, so the panels stay lit continuously, with no atmosphere in the way to scatter the light before it arrives. Clusters of satellites would be linked by laser so that they operate together as a single data center rather than as isolated machines.
What is in orbit today is modest. One prototype, about the size of a refrigerator, carrying four TPUs on roughly a kilowatt of power. For scale, the AI data centers being built on the ground draw many orders of magnitude more than that. This is a proof of concept, not infrastructure.
The model running on it is Gemma 4, Google's open weight model. Open weight means Google publishes the model for anyone to download and run themselves, which is precisely why it can live on a satellite. It does not have to call home to answer a question. The intelligence travels with the hardware.
Next come heat and radiation tests, which are the two things most likely to kill the idea. After that, twin linked satellites in 2027 to find out whether the laser connections hold up between two objects moving at orbital velocity.
Solar power is seductive right up until you account for the planet. Down here, we do not get light all the time everywhere. We have an atmosphere, and we have clouds, both of which sit between the sun and the panel. Put the array in orbit above all of that and you get close to constant power from a resource that costs nothing.
The second argument is the one that should sound familiar to anyone who has read a local news story about a data center in the last two years. Ground based AI infrastructure consumes enormous amounts of water. It carries real environmental impact. It is generating serious pushback from the communities asked to host it, and that pushback is now a genuine constraint on how fast the industry can build. Moving the compute off the planet eliminates that entire category of problem.
The networking piece is less obvious but not unproven. Laser links between satellites sound exotic, and Starlink has already demonstrated the approach at scale for a different purpose. Connectivity is not compute, but the engineering is not starting from zero.
If Suncatcher were the only effort, it would be a curiosity. It is not.
SpaceX has asked the FCC for authority to run up to a million orbital data center satellites linked to Starlink. Blue Origin has filed for more than 50,000. A startup called StarCloud has already trained a small model in orbit. China has been launching its own computing constellation since last year.
That is four serious efforts from four different directions, and the filings are not aspirational white papers. They are regulatory applications, which is what it looks like when an idea moves from the research group to the business plan.
None of this is close to done, and the obstacles are substantial.
Nothing can be repaired once it is up there. Every satellite is a sealed container on a one way trip. If something fails, the current answer is to send it back into the atmosphere to burn up. That is an expensive way to handle a hardware fault.
Heat is harder than people assume. In a vacuum there is no air to carry warmth away from a chip, and shedding heat turns out to be one of the genuinely difficult engineering problems in orbital compute.
And the entire economic argument rests on launch costs falling by something like a factor of 18. Not improving. Falling by a factor of 18. If that does not happen, the math never closes regardless of how well the hardware performs.
The interesting part is how quickly those constraints start generating their own solutions. The satellite housing, the solar panels, the communications gear, the lasers: all of that will likely outlive a TPU by a wide margin. Chips want replacing every couple of years. A satellite body might reasonably last ten or fifteen. That gap creates a strong economic incentive to make these things modular, so that spent compute modules eject and burn up or return to be recycled, while fresh ones dock and slot into place. Or to send a small repair robot up alongside the hardware, capable of handling the most common failures. None of that exists yet. All of it is a reasonable thing to expect someone to build, because the alternative is throwing away a perfectly good satellite every time a chip dies.
Here is the part worth sitting with.
A moonshot is supposed to be a twenty year proposition. The word itself is a request for patience. Google announced this one recently enough that most people filed it under distant speculation, and there is now a working prototype in orbit answering prompts.
Elon Musk has talked about orbital compute arriving within three to five years. He runs optimistic on timelines and usually by a lot. He may be closer than usual this time, because the compounding underneath both AI and launch capability is running faster than most estimates account for. Whether the real answer is five years or fifteen, the direction is not seriously in doubt.
That gap between how far away something feels and how far away it is turns out to be the expensive part. Most of us are not bad at recognizing change when it arrives. We are bad at estimating how long we have before it does.
Nothing, this quarter. No association is procuring orbital compute, and nobody should build a strategy around satellites.
What it should do is adjust your instincts. If you have a mental folder labeled someday, this is evidence that the folder empties faster than it used to. The things in it are worth revisiting on a shorter cycle than feels necessary, which is also the argument for getting full value out of the tools already in front of you rather than waiting for the landscape to settle. It is not going to settle.
The more useful move is to take off the operations hat for a moment and put on the member hat. The technology is interesting to the degree that you find it interesting. What matters considerably more is the speed of change itself, and the fact that your members are living inside it.
Some of you work in fields where it is tempting to assume this stuff lands somewhere else. Maybe that is right. Often it is not, because change reaches people through side doors. A new alternative appears in an adjacent market. A tool makes a skill your members spent years building suddenly less valuable. Or it reaches them personally rather than professionally, through their kids or their household or their own sense of where the world is heading, which is still your concern even though it never shows up in a benchmarking survey.
So the questions are these. Where are your members going? What pressures and opportunities are showing up in your field because things are moving this quickly? And what should your organization be building so that members come out ahead of the change rather than behind it?
That line of thinking starts from your members rather than from your operations, which is a slightly selfless way to approach strategy. It also tends to produce better answers, because the internal questions mostly resolve themselves once you are clear on what your members actually need from you.
A satellite over Norway wrote a haiku about starlight. The relevant part is not the satellite. It is how quickly we went from that being absurd to that being Tuesday.