The Moonshot Is a Confession

Commentary on “Google Wants to Put AI Data Centers in Space”  ·  The New York Times

The story reads as science fiction arriving ahead of schedule: Google wants to put its AI chips in orbit. The useful way to read it is not as a plan for space. It is a confession about Earth.

Project Suncatcher, run with the satellite firm Planet Labs, will launch two prototype satellites carrying Google's TPUs by early 2027. The research behind it imagines clusters of 81 satellites flying in roughly one-kilometre formation at an altitude of 650 kilometres, in a dawn-dusk sun-synchronous orbit that never passes into the Earth's shadow and therefore draws continuous solar power, "unaffected by weather, land use, water availability, or grid infrastructure." The satellites would talk to each other by laser at up to 10 terabits per second. Google's chips survived a radiation test simulating five years in low orbit, with memory the weak point and the error rate judged acceptable for inference if not yet for training. Heat would be moved passively to radiator panels. And the economics only close if launch costs fall from something like $1,500 to $2,900 per kilogram today to around $200, which Google guesses might happen by 2035 if SpaceX's Starship reaches 180 flights a year. Sundar Pichai calls it a moonshot.

Read back the list of things space is meant to escape: weather, land, water, grid. That is not a list of space problems. It is the exact list of terrestrial constraints this series has spent every issue describing.

The signal is terrestrial

When the most capable infrastructure operator on the planet concludes that one plausible route to more compute is to launch it off the planet, that is a measurement. It measures how binding land, water, grid interconnection, and community consent have become on the ground.

Nobody models leaving Earth because orbit is convenient. You model it because the alternative — more transformers, more transmission, more permits, more "No" at more councils — has become slow and contested enough that a constellation of satellites reads as a reasonable comparison case. Suncatcher is the clearest evidence yet of how expensive the terrestrial constraint has become. It is a data point about the grid, dressed as a data point about space.

But the constraint follows you up

Space removes land, water, and the grid, and hands back a different set of problems that are, if anything, harder. Radiation degrades memory. Launch cost dominates the whole case and the whole case rests on it falling by an order of magnitude on the back of someone else's moonshot.

And heat, the subject of recent pieces here, does not get easier in orbit. It gets harder. On Earth a facility can reject heat into air or water. In a vacuum there is nothing to conduct it into; heat can only be radiated, slowly, from dedicated surfaces. The one consequence every terrestrial data centre externalises over its fence line becomes, in space, a first-order engineering constraint with no neighbourhood to absorb it. Space does not repeal the laws that make data centres hard. It trades the constraints we have partly solved for ones we have not.

A research paper is not a data centre

Two prototypes, memory rated acceptable for inference but not yet trusted for training, and a business case that turns positive around 2035 if a different company's rocket reaches a flight rate it has never achieved. This is a serious research programme and an honest one, and Google presents it as exactly that. It is not capacity for this decade.

The compute the world is fighting over between now and 2030 will be built where compute is always built: on land, on grids, near water, beside communities whose consent is now part of the critical path. Suncatcher changes none of that. It only tells you how much the people with the most resources wish it were otherwise.

The Canadian read

There is a quieter lesson in the orbit Google chose. It went looking for a place with abundant, continuous, weather-independent power and no competition for land or water. That place exists on Earth too, in more forms than the discourse admits. It just requires siting compute where clean power is plentiful and people are not, instead of clustering it in the metros where the grid is already tightest.

The terrestrial version of Suncatcher's insight — go where the power is, not where the real estate is — is available now, at nothing per kilogram to orbit. The countries and operators that take it seriously will have built real capacity by the time the first orbital cluster is anything more than a paper and two prototypes.

What this means going forward

Suncatcher will be reported as the future of data centres. It is better understood as a verdict on the present. The frontier has become expensive enough on Earth that leaving Earth is now a line item in a research budget.

That should concentrate attention rather than scatter it. The advantage for the next decade goes to whoever can do the unglamorous terrestrial work — firm clean power, real interconnection, honest handling of heat, consent earned in advance — faster than the queue and the opposition. Space is the moonshot. The ground is the business.

Google has not really told us it is going to space. It has told us how badly the grid on Earth is holding it back. That is the sentence worth keeping.

VOLTEDGE

Reference: “Google Wants to Put AI Data Centers in Space” · The New York Times · read the article