I've read a lot of AI announcements this year, and most of them blur together. This one didn't. Google put a tiny AI data center into orbit, switched it on, and the first thing it did was write a haiku. Project Suncatcher is officially talking to Earth, and I think it's the most interesting weird experiment in tech right now, even if the "data centers in space" pitch still sounds a bit like a sci-fi novel's back cover.
Let me be clear up front: nobody is moving ChatGPT to orbit next year. But this test matters more than a cute poem suggests, and I'll explain why.
What Project Suncatcher actually launched
According to TechCrunch, Google's prototype satellite went up earlier this month on a SpaceX Falcon 9, sharing the ride with more than 100 other spacecraft. After roughly three weeks of commissioning, where engineers checked each subsystem, it was ready for real work.
The hardware is surprisingly modest. The satellite carries four of Google's Tensor Processing Units, the custom chips Google uses for AI, and can generate up to 1 kW of power. That sounds decent until you learn the catch: it can only run one TPU at a time, and only for about 20 minutes per session. That's less a data center and more a very expensive laptop that needs a nap.
Pre-loaded on board is Gemma 4, Google DeepMind's open model. The Verge reports that Google confirmed the TPUs booted successfully and the model answered its first prompt from space.
The haiku heard 'round the world
The first prompt was simple: write a haiku about being in space. Gemma 4 came back with this:
Blue marble below
Data streams in silent grace
Starlight feeds my core.
Is it good poetry? Honestly, it's fine. It's the kind of haiku a decent eighth grader turns in. But think about what's happening: a model running on chips that are flying around the planet, getting a prompt, doing inference, and beaming the answer back down. That full loop working on the first try is the real headline.
Google also tried other space-themed prompts and, more usefully, asked the satellite to report on the orbital conditions around it. A spokesperson told TechCrunch the satellite was operating nominally. In space-speak, that's a big thumbs up.
Why put AI chips in orbit at all?
Here's the pitch, as I understand it. AI eats electricity. Data centers on the ground fight for power, land, water and permits. In the right orbit, a satellite gets a lot of steady sunlight, and Google's long-term bet is that solar-powered compute in space could eventually be cheaper than building more giant server farms down here. Google laid out the bigger design idea in a Google Research post on a space-based AI infrastructure system.
The part that makes Suncatcher different from rivals is the plan for satellites to fly close together and talk over high-speed links, so a group of them can share one big job. TechCrunch notes that SpaceX's own orbital data center plans also use laser links, but its satellites are expected to sit too far apart to share processing well. Startups like Starcloud and Cowboy Space are going smaller, aiming to run inference on individual satellites.
The radiation problem nobody can fake on Earth
This is the nerdy bit I actually care about. Space is full of radiation, and radiation can cause "bit flips," where a stored value in memory quietly changes. The scary part is that this can corrupt data without setting off the usual alarms.
Google has already blasted its TPUs in particle accelerators on the ground. Those tests suggested radiation would be a small issue for inference, which is answering prompts, and a bigger headache for training models. Now the team gets to see what happens in the real thing instead of a simulation. The ongoing test campaign will also look at how the chips handle heat stress, which is its own nightmare when you can't just add a fan.
What comes next for Project Suncatcher
If the chips hold up, Google plans to launch two more TPU satellites in 2027. Those will generate more power and test optical links between satellites, which is the piece that would make a real orbital cluster possible.
And then there's the elephant in the launch pad. Both Google's and SpaceX's visions depend on Starship, SpaceX's huge rocket that's still in development, becoming fully and rapidly reusable. In a peer-reviewed analysis published this month, Suncatcher researchers estimated SpaceX would need to fly Starship around 1,800 times to get the kind of cost drops it saw with earlier rockets. That's not a typo. Eighteen hundred.
Worth knowing too: Alphabet, Google's parent company, holds a significant stake in SpaceX. So these two "competing" visions are a lot more tangled than they look.
Should you care?
Short answer: yes, but not because it changes anything for you this year.
Your Gemini chats aren't moving to orbit. Your phone isn't getting faster. Google itself frames Suncatcher as infrastructure meant to be economical in the 2030s. This is a classic moonshot, and moonshots fail a lot.
What I like about it is that it's an honest answer to a real problem. Everyone in AI keeps saying "we need more power," and most of the answers are bigger plants and bigger bills. Google is at least testing a genuinely different idea, with real hardware, and sharing what happens.
Here's my take on who should pay attention:
- If you follow AI infrastructure, this is the test to watch, especially the radiation results.
- If you're a space nerd, the 2027 launches with optical links are the real milestone.
- If you just use AI apps, file this under "cool, check back in five years."
The timing is also a little awkward. TechCrunch points out that Alphabet's stock dropped this week after it reported that putting AI into search is hurting ad revenue growth. Long-term moonshots are easy to love when the money is flowing. They get harder to defend when investors start asking questions. I hope Suncatcher survives that pressure, because experiments like this are how weird ideas become normal ones.
Quick FAQ
What is Project Suncatcher?
It's Google's experimental project to explore running AI computing on satellites in orbit, powered by sunlight, as a possible future alternative to building more data centers on the ground.
What AI model is running on the satellite?
Gemma 4, Google DeepMind's open model, running on Google's TPU chips.
Can it run all the time?
Not yet. The satellite has four TPUs but can power only one at a time, for about 20 minutes per session.
When is the next launch?
Google plans two more TPU satellites in 2027, with more power and links between satellites.
So I'll throw it to you: would you trust an AI data center floating above your head, or does this feel like a very expensive way to write haikus? Tell me in the comments.
0 Comments