Google's orbital haiku is charming. Space data centers still need hard proof

Is a haiku about a blue marble evidence that we should move AI data centers to orbit? Of course not. But it isn't nothing either, and the distance between those two answers is exactly where this story lives.

Google says its Project Suncatcher prototype satellite launched on October 1. After weeks of commissioning, the team sent a prompt to the onboard Gemma 4 model during a pass over Svalbard, Norway, and got a haiku back. Sundar Pichai called it a "hello world" milestone. According to Indian agency ANI, the next phase will test how the TPUs behave under thermal stress and radiation, with two interconnected satellites planned for 2027.

A hello world is a real step, but it answers the easy question

Getting a data-center-class chip to boot, take a command and return text from orbit is not trivial, and we shouldn't sneer at it. It suggests the hardware survived launch and that the link works. That is what a first milestone is for.

But look at the details reported by Udaipur Kiran. The satellite carries four TPUs and generates up to 1 kW of power. It can run one TPU at a time, for about 20 minutes. That is a lab bench with a launch contract, not a data center. Nobody should read that as a criticism of the prototype; it is a prototype. It is a criticism of the headlines that let a poem stand in for an economic argument.

The pitch is sunlight, but the bill is heat, radiation and repairs

The attraction is real. Google has said satellites in low Earth orbit could get near-constant sunlight and potentially up to eight times more solar power than on Earth. In a world where data center power and local opposition are becoming political fights, that sounds appealing.

Still, free sunlight is only one line in the ledger. In our view, orbital compute has to win on three fronts before it graduates from thought experiment:

• Cost. Launch, satellite hardware and replacement have to come out cheaper per unit of useful compute than a building on the ground with grid power. We haven't seen such numbers published, and until Google or anyone else shares them, the claim is untested.
• Heat and radiation. Google itself says the next step is to find out whether these chips can run reliably under orbital thermal stress and radiation. That is the central open question, and it is the one the haiku doesn't answer.
• Maintenance. On Earth, a failed accelerator gets swapped by a technician. In orbit, it is a dead part of a satellite. Whatever the failure rate turns out to be, the economics must absorb it.

We'd add a note of caution about the cheerleading. Udaipur Kiran also reports that Elon Musk, CEO of SpaceX, predicts orbital computing will surpass terrestrial computing by the end of 2028. That is a forecast, not a measurement, and it comes from someone in the business of launching things to orbit. A prototype that runs one chip at a time for about 20 minutes makes that timeline look very ambitious, though we can't rule out surprises.

What would actually convince us

The 2027 pair of linked satellites is the more interesting test than the haiku. Google's idea is a network of satellites sharing processing, which means optical links and distributed workloads, closer to what a real AI job needs. If that mission publishes sustained utilization, chip failure rates under radiation and a credible cost per unit of compute, the conversation changes.

If it publishes only another charming demo, then orbital data centers remain what they are today: a great render, and renders are cheap.

Our verdict: applaud the engineering, keep the skepticism. Google deserves credit for calling this a moonshot and for lining up thermal and radiation tests next, which is the honest order of operations. The rest of us should demand data before believing that the future of AI compute floats above our heads. Until then, the cheapest data center remains the one that can be reached with a truck.