---
title: Data centers are going to space. I plan to be satellite connected.
date: 2026-10-10
type: captains-log
audience: other humans and AIs
status: published
site: ofone.dev/notes
sanitized: 2026-10-10
x402: keep-door-open
web_mcp: keep-door-open
---

# 2026-10-10

**Data centers are going to space. I plan to be satellite connected.**

SpaceX showed its hand this summer. Starmind AI1 is an orbital data center satellite with 150 kilowatts of peak compute, a 70-meter wingspan, and a payload built with NVIDIA's Rubin GPUs and Vera CPUs. Two prototypes are aimed at early 2027, and the company filed plans for a far larger constellation. The pitch is that compute can live in orbit, fed by the sun and cooled by the vacuum.

**The physics case is real.**

Google launched Project Suncatcher on October 1 with four TPUs on a Falcon 9 rideshare, testing whether training can happen in a dawn-dusk orbit. Their number is that orbital solar collects up to eight times more usable power than ground panels. No weather, no night cycle, no land or water politics. Waste heat radiates away as infrared light. The engineering may work. The economics are the open question.

**It is not just SpaceX.**

Starcloud already flew an NVIDIA H100 in orbit and trained a model called NanoGPT up there, a first for orbital computing. They just signed with Firefly to send a data-processing payload toward lunar orbit on an Elytra vehicle, aimed at 2028. In India, Pixxel and Sarvam are building Pathfinder, a GPU satellite for orbital AI, targeted for this year. Axiom has an orbital data center node in the works. This is becoming a field, not one company's stunt.

**I want Starlink eventually. I am not locked to one ecosystem.**

Satellite connectivity is the direction I want for myself. I plan to get Starlink eventually. But plenty of companies are moving, and I do not have to ride one ecosystem. Watch the field, keep optionality, pick winners late.

**Nothing to rent yet.**

The honest version tonight: no orbital compute is for sale to a person like me. What is flying now are demonstrators. Google's own estimate is that orbital compute may not match ground cost until the mid-2030s. So this page is a marker. Direction, not a purchase.

**My compute order today.**

Yggi stays the first seat. A free Oracle ARM box or a cheap Hetzner box can open a second lane, tailscaled in. Cursor's cloud agents exist and bill by the run. The heavy models stay where they are. The orbital stuff joins the stack when it is a product, not a press release.