- Published on
Five nines in orbit, and the reason is permits
- Authors
- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · Five nines in orbit, and the reason is permits
0:00 / 0:00
Watch this article · Five nines in orbit, and the reason is permits
Larry Page reportedly has a number for how much computing ends up in space: five nines. As in 99.999%.
The number arrived secondhand. Patrick O'Shaughnessy, who hosts Invest Like the Best, told it on the podcast: someone asked Page how much of compute would be orbital, and he said five nines. Martin Casado, a general partner at a16z, retold it on the same episode. Within a day the clip was a post with close to a million views, and Elon Musk had endorsed it — "Larry Page has made many accurate predictions and this is one of them."
At five nines, one unit of compute in every 100,000 stays on Earth. That is not a forecast so much as a position: the ground stops being where compute lives.
The argument is not about space
Casado's case is unusually blunt. "The actual space thing actually pencils out," he said, "but the reason it makes the most sense is just regulatory. It's just basically regulatory. There's 500 —" and he went on to describe proposed bills aimed at stopping data centre construction across the United States, Texas included.
He added that the operators he talks to say it is very hard to break ground right now, and that the choice, if permitting keeps tightening, is to go international or to go into space.
That reframes the whole claim. Five nines is not a bet that orbital mechanics got easier. It is a bet that the ground got harder — that the binding constraint on AI capacity is a planning committee, not a power plant.
What is actually up there
The gap between the claim and the hardware is worth stating plainly.
Google's Project Suncatcher launched its first prototype on 1 October, on a SpaceX Transporter-18 rideshare, built with Planet. The satellite is about the size of a refrigerator and carries four TPUs drawing roughly one kilowatt — one server, in orbit, in a sun-synchronous orbit where the panels see sunlight almost continuously and can generate up to eight times what the same panel makes on Earth. On 30 October, Sundar Pichai reported the first complete AI response computed from orbit.
Others are further along in raw silicon. Starcloud put an NVIDIA H100 in orbit in November 2025, raised 1.1 billion valuation in March 2026, and plans a full GPU cluster on its second satellite in 2027. Axiom Space placed two orbital data-centre nodes in January 2026. NVIDIA announced a space-rated Vera Rubin module at GTC this year. Google's own plan is a constellation of solar-powered satellites linked by free-space optical lasers.
The intentions are enormous. SpaceX has filed for a constellation of up to one million AI satellites under the name Starmind, with deployment from the fourth quarter of 2027 and scale-up in 2028, at a stated 100 kW of compute per tonne. Blue Origin has described a platform of nearly 52,000 satellites.
What has to be true first
A hyperscale campus runs at a hundred megawatts or more. The prototype in orbit runs at one kilowatt — five orders of magnitude apart, which is, by coincidence, the same ratio as the prediction itself.
against the top of US data-centre power spend, in one analyst’s arithmetic
the baseline it is measured against
Four things have to stop being problems, in order. Launch has to be cheap enough per kilowatt-year that a satellite beats a substation. Hardware has to survive radiation and thermal cycling for years, not months. Cooling has to work in a vacuum, where the only way to shed heat is to radiate it. And the constellation has to be serviceable — because a failed accelerator in orbit is not a ticket, it is a dead satellite.
Permit
The step that is failing on the ground: hundreds of proposed bills, and operators saying they cannot break ground.
Launch
Cost per kW-year has to beat grid power. One analysis puts it at roughly five times today.
Survive
Radiation, vibration and thermal cycling. A year of operation is the first real test.
Link
Lasers between satellites, so the constellation behaves like one machine rather than a thousand.
Why the endorsement is not evidence
Musk backing the number is meaningful in one narrow way: he controls the launch capacity that any orbital compute plan depends on, and SpaceX is already the most aggressive builder of the thing Page is describing. When the person with the rockets says the prediction is right, that is a statement about his own roadmap as much as about physics.
It is not evidence that five nines happens. The claim is secondhand, the count of 500 bills is Casado's own tally rather than an official one, and the orbital compute installed today is a rounding error against a single data hall — four chips, a few nodes, one H100.
But the shape of the argument is what matters. For a decade the question about AI capacity was where the electricity comes from. The interesting thing about "five nines" is that it is not an energy claim at all. It is a claim that the constraint has moved to the planning system, and that capital will route around a permit the same way it once routed around a tariff.
If the bills pass and the ground keeps getting harder, the number stops sounding like a joke. If they do not, it stays a line on a podcast. Either way, the sentence to remember is not the one about space. It is Casado's: it's just basically regulatory.
Source: Invest Like the Best, episode with Martin Casado, 4 November 2026; Patrick O'Shaughnessy's account; Elon Musk on X, 8 November 2026; Google's Project Suncatcher posts (September–October 2026); Starcloud, Axiom Space, SpaceX Starmind and Blue Origin filings as reported; Futurum Group's orbital computing estimate.
Contact: [email protected]