AI infrastructure
Five Nines in Orbit
An AI chip can already answer from space. What would it take for the rest of us to depend on it?
Someone asked Larry Page how much computing would eventually happen in orbit. As Patrick O’Shaughnessy told the story on Invest Like the Best, Page’s answer was five nines. That’s 99.999%. Take the number literally and Earth keeps one unit of compute out of every hundred thousand. O’Shaughnessy was recounting something he’d heard, and the account gives us neither a date nor a unit for measuring the share.
Anyone who has worked around data centers has heard those words before. Five nines describes an availability target. I find it useful to put that familiar meaning beside the forecast. One number asks where the work gets done. The other asks whether the work gets done when somebody needs it.
Share of computing
99.999%
The share that would run in orbit under the prediction attributed to Page.
Time-based availability
5:15
Roughly five minutes and fifteen seconds of downtime allowed in a 365-day year at five nines.
A chip answered. A service has to keep answering.
On October 30, Google reported that a data-center-class TPU on its Project Suncatcher prototype had returned an AI response from orbit. That’s an actual. A chip went up, ran a workload and sent an answer back. Before the demonstration, the team had a design and an expectation. Now it has a result it can build on.
Some orbits offer long stretches of sunlight. That helps with power. Heat is a separate engineering problem. The chip still produces it, and the spacecraft has to carry it from the electronics to a radiator that can shed it into space. A cluster of satellites adds another problem. If its chips have to work together, moving spacecraft need to keep their optical links aligned while exchanging data fast enough to make the cluster useful.
A prototype can answer once and teach its operators a great deal. A service earns its place by answering through equipment failures, communications interruptions and maintenance. The useful test starts with a request somebody on Earth actually needs to make.
The promise stays on Earth
When someone calls my business, they care whether they got through and whether we did what we said we would do. The processor’s location is the operator’s business. If a processor is available while a caller spends the morning trying to reach us, we can have a beautiful uptime number and still owe somebody a call.
That is the practical question for a delegate, too. I’m ready to stop checking when I know what counts as done, the system can tell whether it happened, and an exception comes back to someone who can act in time. Those are three separate pieces of work. A powerful model can help carry them. The people who make the promise still have to see that they’re carried.
The case for orbit starts with the work it can do well. As that set grows, the forecast will acquire a measurable denominator. A customer will know the service has matured when they can leave their attention where it was and trust the answer to arrive.