Google's orbital haiku is a stunt; the data center is the question

Is a haiku from orbit a breakthrough or a very expensive postcard? Both, and the second description is the more useful one. Google's poem is a good story. It is not evidence that data centers in space make sense.

Here is what happened. Google's prototype satellite, part of Project Suncatcher, launched on October 1 and reached low Earth orbit. After commissioning, the team uplinked a prompt to the onboard Gemma 4 model, which ran on a TPU. As the satellite passed over Svalbard, Norway, the reply came down, a "Hello World" milestone in Google's words. The Register reports that the prompt asked for a haiku from the perspective of a TPU orbiting Earth.

Google then tried limericks and telemetry prompts. The tests that matter, TPU behavior under orbital thermal conditions and radiation, are still ahead.

The haiku proves the chip turned on, not that the idea works

Strip away the poetry and the result is this: a chip powered up, took a prompt and answered. That is a necessary step, and a sensible one. But a working "hello" says nothing about whether such a chip survives for years, or whether it can do so economically. Even a specialist outlet, HW Busters, noted that the hard tests, heat and radiation, have not started yet.

The Register had fun with the economics. Those roughly 20 tokens would cost about $0.000012 through Google's cloud API, while a satellite of perhaps 500 kg, at a retail launch price of around $7,000 per kg, makes it a very expensive haiku. That figure is the outlet's own back-of-the-envelope estimate, based on Starlink-class mass, not a Google number. Treat it as a joke with a point.

Radiation is the honest part of Google's own data

Credit where it's due: Google has published numbers that are not flattering. In lab tests with a 67 MeV proton beam, the Trillium TPU's high-bandwidth memory was the most sensitive component. It showed irregularities after a cumulative dose of 2 krad(Si), which Google says is about three times its minimum requirement of 750 rad(Si).

On silent data corruption, Google's researchers estimate roughly one failure per 3 million inferences at one inference per second. They call that likely acceptable for inference. They add that the impact on training jobs "requires further study."

We think that caveat is the whole story. Inference of a chatbot reply can tolerate an occasional glitch. The large training runs that drive the AI energy debate are the workloads where errors compound. A proton beam on the ground is also not low Earth orbit over years. Until the satellite's own data comes back, this is a promising lab result and nothing more.

The cost case is a forecast, and the environment is missing from it

Google's argument rests on launch cost. If getting a satellite to low Earth orbit falls to $200/kg, it calculates, the amortized launch cost (about $810–7,500/kW/year) could be comparable to US terrestrial data center electricity spending (about $570–3,000/kW/year). The Register adds that Google expects SpaceX could get there around 2035. That is a prediction about someone else's rocket, and a wide range at that. The top of the space range is more than double the top of the ground range.

The pitch also includes an energy advantage: according to EconoTimes, Google estimates that panels in low Earth orbit could generate up to eight times as much power as on Earth. Fine. But the documentation we have says little about the two questions we would press hardest: how you shed heat in vacuum at data center densities, and what a constellation of many small satellites adds to orbital debris and to launch emissions. Google anticipates many small linked satellites and plans a two-satellite mission in 2027. Those questions deserve published answers before anyone talks about scale.

What to demand before believing it

Ground data centers have real problems, and the political fights over them are visible in places like Hermantown, Minnesota, where a council spent hours hearing public comment on a Google project. That is exactly why orbit is tempting: it looks like a way around local opposition. We would be wary of that framing. Moving a problem off the map is not the same as solving it.

Our verdict: enjoy the haiku, then ask for the boring results. We want radiation and thermal data from the actual satellite, ideally peer-reviewed rather than only a blog post. We want an honest lifecycle comparison with ground facilities, including launch emissions and debris risk. And we want the 2027 two-satellite mission to be judged on those numbers, not on its poetry. Until then, orbital AI compute is a research question, not an infrastructure plan.