The Sequence Opinion #888: Everything You Need to Know About the AI in Space Race

For a few years now the AI industry's bottleneck has moved around like a shell game. First it was clever architectures, then it was scraping enough of the internet, then it was hoarding enough GPUs. Now, per this week's Sequence essay, the constraint is electricity itself — grid interconnects, cooling water, land permits, the boring infrastructure that never shows up in a keynote. And that shortage is why a surprising number of very serious people are suddenly talking about putting data centers in space.

The logic isn't mystical. Orbit offers something like unmetered solar power and no local utility commission to fight with. There's no zoning board 250 miles up. So companies ranging from hyperscalers to chipmakers to a handful of scrappy startups are treating low Earth orbit as real estate worth fighting over, not a moonshot side project. Nation-states are in the mix too, which tells you this isn't just a Silicon Valley fantasy.

The proof isn't theoretical anymore. As of December 2025, a satellite weighing about 130 pounds carried an Nvidia H100 into orbit and used it to train nanoGPT — Andrej Karpathy's stripped-down GPT implementation — on the complete works of Shakespeare. It's a toy model solving a toy problem, sure. But it's a genuine loss curve computed in space, which is a very different milestone than a slide deck promising future orbital compute.

What happens next is the interesting part. Training a model on Shakespeare is one thing; running frontier-scale inference or multi-billion-parameter training runs in orbit is an entirely different engineering problem involving thermal management, radiation-hardened hardware, and getting data up and down without a fiber line. The essay frames this as architecture and player analysis still to come, which suggests the current experiments are proof-of-concept flags planted before the real infrastructure fight starts.