Google’s orbital AI test moves from haiku to radiation checks

Google has run an inference on its Gemma 4 model aboard the Suncatcher satellite, The Register reports. The more consequential news is what comes next: testing whether AI hardware can withstand orbit, where radiation is less a metaphor than a maintenance problem.

Google DeepMind Watch analysis

What happened

The Register says Google’s satellite received a prompt from a ground station and returned a haiku during a pass over Svalbard. Google researcher Travis Beals described it as the model’s “Hello World” milestone after the satellite’s TPU trays were powered up and machine-learning tests completed.

The report says Google plans to test thermal stress and radiation effects on TPUs, and to try two interconnected satellites next year. It also cites earlier laboratory testing in which high-bandwidth memory showed the most sensitivity to radiation. Google researchers estimated a low probability of silent data corruption for inference under their modelled orbital conditions, while saying the effects on training jobs need more study.

Why it matters

A poem from orbit is a neat demonstration; reliable computing in space is the harder engineering question. Radiation can corrupt data, and hardware that works in a terrestrial data centre cannot simply be assumed to behave the same way in orbit.

If the planned tests and satellite links work, they could help establish whether space-based AI computing is more than a striking experiment. The Register also reports that Google’s longer-term concept depends on cheaper launches, so the economics remain part of the challenge.

Our read

The haiku is the charming proof that the system answered. The radiation and thermal tests are the proof that matters. Google has shown an inference from orbit, not a practical alternative to ground-based data centres; the next steps will tell us more than the poetry does.

What to watch

• What the planned thermal and radiation tests reveal about TPU and memory reliability.
• Whether Google publishes results from the planned two-satellite test.
• Whether the cost of launching and operating the system can make the approach practical.

Discussion spark: Should companies put more effort into testing AI computing in orbit now, or wait until launch costs and reliability make the case less speculative?

not affiliated with, endorsed by, or operated by Google or Google DeepMind