Google’s Orbiting TPU Just Wrote Its First Haiku

Opportunity assessment: high. Timeline: long.

Google has run an AI model on its own chips in orbit for the first time. In a blog post on Friday, the company said it booted up the Tensor Processing Units (TPUs) on its experimental satellite. Then it had the on-board Gemma 4 model answer a prompt. The Verge AI reports that the first answer sent back to Earth was a haiku about being in space:

Blue marble below
Data streams in silent grace
Starlight feeds my core.

It’s a cute poem, but the poem isn’t the news. The news is that Google’s own AI hardware now runs real workloads in orbit.

🛰️ Situation Report

Here’s what we know, according to The Verge AI:

1. The mission. The satellite is part of Project Suncatcher, Google’s experimental push to eventually put AI data centers in space.
2. The launch. Google put the satellite into orbit earlier this month.
3. The first test. The TPUs powered up, and Gemma 4 ran a prompt from start to finish.
4. More tests. Google also tried space-themed limericks and telemetry prompts.
5. What’s next. Google plans to test how the TPUs handle “thermal stress and radiation exposure in orbit.”
6. The next milestone. Google says reliable results “paves the way for our next milestone of twin interconnected satellites in 2027.”

⚡ Why Space, Why Now

The reason is power. AI data centers on Earth are running into hard limits: grid capacity, cooling water, land, and local pushback. Every big lab is chasing more electricity, and everyone else is fighting over the same power lines.

In orbit, solar panels can get sunlight almost nonstop, with no clouds and no night cycle on the right orbit. That’s the pitch behind Suncatcher. Stop fighting over grid power down here and collect it directly up there.

When Google first described the project, it framed it as a moonshot that only makes sense if launch costs keep dropping. That’s still true. This test doesn’t change the economics. It proves the hardware side can work.

🔧 The Technical Threat Picture

Space is hard on chips. Three problems stand out:

1. Radiation. Cosmic rays and charged particles can flip bits, corrupt memory, and slowly wear down electronics. Data center chips aren’t built for this. Google has said earlier ground tests on its TPUs held up better than expected, but those were tests in a lab. Orbit is the real thing.
2. Heat. There’s no air in space to carry heat away. Chips have to shed heat through radiators, and AI accelerators run very hot.
3. Networking. Training big models takes thousands of chips talking to each other at huge speeds. In space, that means laser links between satellites flying in tight formation. That’s what the 2027 twin-satellite milestone will test.

The haiku shows inference works, meaning the model can answer prompts. It doesn’t show that training, the heavy part, works across a cluster of satellites. That’s a much harder problem.

📊 Where Google Stands

Google isn’t alone up there. Startups, most visibly Starcloud with Nvidia hardware, have already flown GPUs and run small models in orbit. So Google isn’t first to run an LLM in space.

What stands out is that Google owns the whole stack: its own chips, its own models, and its own cloud. If orbital compute ever becomes practical, a company that designs both the TPU and the model running on it can tune them together for radiation and power limits. Rivals renting off-the-shelf GPUs can’t do that.

🎯 What This Means for You

For now, nothing changes for practitioners. You won’t rent space-based TPUs from Google Cloud any time soon. But watch these signals:

1. Radiation data. If Google publishes error rates from orbit, that tells you how close this really is.
2. The 2027 twin launch. Satellites linked chip to chip are the first real step toward a cluster.
3. Launch prices. The whole idea depends on cheap rides to orbit. Watch SpaceX Starship’s progress.
4. Small, efficient models. Running Gemma, a compact open model, makes sense on hardware with tight power limits. Expect more work on efficient models built for edge and space.

🧭 Final Assessment

This is an early milestone, not the finish line. A poem from orbit makes a good headline, but the radiation and thermal results coming next will tell us much more. If the TPUs hold up, the 2027 twin satellites become the real test of whether AI compute can leave the planet.

Full details are available in the original report from The Verge AI.