It was a single word — just “Yes” — but it confirmed what the space and tech worlds had been speculating about for months. Late Sunday night, Elon Musk replied to a user on X who asked whether the newly trademarked name “Starmind” would apply to SpaceX’s orbital AI compute network. That one-word answer locked it in. The constellation has a name.

If Starlink was SpaceX‘s answer to connecting the planet, Starmind is its bid to compute for it. The distinction matters. Starlink moves data. Starmind processes it. The idea, laid out in an FCC filing earlier this year, is a constellation of up to one million satellites — each one essentially a data center node in low Earth orbit, running AI inference workloads on board, powered by massive solar arrays, and linked together by optical lasers.

The filing, submitted January 30, describes what SpaceX calls the “Orbital Data Center System.” Satellites would operate in narrow orbital shells about 50 kilometers thick, at altitudes between 500 and 2,000 kilometers. They’d use sun-synchronous orbits to stay in near-constant sunlight, drawing power directly from the sun and radiating heat into the vacuum of space — no chillers, no water cooling, no grid connection required.

The first hardware has a name too: AI1. Sometimes called “AI Sat Mini” in early chatter, it’s anything but mini. SpaceX unveiled the design in early June via a video on X. The satellite spans 70 meters — about 230 feet — from tip to tip when its solar arrays and radiators are fully deployed. It stands 20 meters tall. Average compute output: 120 kilowatts. Peak: 150 kilowatts. Thermal management comes through a liquid-radiator system with redundant pumping loops and micrometeoroid shielding.

Musk has said the design leverages tech already developed for Starlink V3. The compute payload is modular — NVIDIA Rubin chips, GB300s, TPUs, whatever the customer needs. The satellite bus handles power, cooling, and connectivity.

Two AI1 prototypes are targeting an early 2027 launch. A new manufacturing facility — Gigasat, in Bastrop, Texas — aims to be operational by year-end 2027. Volume production would follow.

None of this happens without Starship. The economics only work if launch costs drop by an order of magnitude. SpaceX’s FCC filing explicitly requests a waiver of the usual milestone rules — half the constellation in six years, full deployment in nine — because the timeline depends entirely on Starship achieving full, rapid reusability. The company has told investors it could add 100 gigawatts of AI compute capacity annually once Starship is flying regularly, launching a million tonnes of satellite hardware per year at 100 kilowatts per tonne.

The trademark for Starmind was filed by xAI, Musk’s artificial intelligence venture, founded in 2023. In February 2026, xAI merged into SpaceX in a transaction valuing the combined entity at $1.25 trillion. That merger — still not fully public in its details — brought Grok, X (formerly Twitter), and xAI’s compute ambitions under the same roof as the world’s most prolific launch provider.

The strategic logic is straightforward. Terrestrial AI data centers are hitting hard limits: power, cooling, land, permitting, water. The biggest clusters now draw hundreds of megawatts. xAI’s Colossus facility in Memphis has drawn scrutiny over air pollution. Google, Microsoft, Amazon — they’re all hunting for gigawatt-scale sites. Space offers effectively unlimited solar energy, free vacuum cooling, and no neighbors to complain to.

SpaceX isn’t alone in seeing this. Blue Origin filed for TeraWave, a constellation of thousands of optical-communication satellites targeting 6 terabits per second links. Google’s Project Suncatcher envisions TPU clusters in orbit, potentially scaling to terawatts. NTT, Axiom, Ramon.Space, Aetherflux, Sophia Space — all have orbital compute plans. Starcloud (formerly Lumen Orbit) already flew an H100 GPU as a demo. Jeff Bezos has said gigawatt orbital data centers are a decade out. Eric Schmidt bought a rocket company to pursue the same.

But SpaceX has two things the others don’t: a working megaconstellation already in orbit (Starlink, with over 9,500 launched and 8,000 active), and the only fully reusable super-heavy lift vehicle in development. The Starlink laser mesh — currently 200Gbps per link, heading to 1Tbps — provides the backhaul. Starmind satellites would plug into that mesh, compute on orbit, and send results down through Starlink’s ground stations.

The FCC filing is light on hardware specifics — no mass figures, no detailed power budgets, no chip manifests. That’s deliberate. The architecture is still fluid. But the scale is not. One million satellites would dwarf every other object in orbit combined. Today, roughly 14,000 active satellites circle Earth. Starlink alone accounts for most of them. China has filed for 200,000 across two constellations, though analysts suspect those are partly orbital-plane reservations.

Regulatory approval is far from guaranteed. The FCC has never evaluated a constellation this large. Debris risk, spectrum coordination, astronomical interference — every review will be precedent-setting. SpaceX has asked for milestone waivers, effectively arguing that the usual deployment schedules don’t apply to a system dependent on a rocket that hasn’t yet demonstrated full reusability. Starship’s next test flight is expected by March 2026; the last two tests of 2025 reached orbit but didn’t recover either stage.

Meanwhile, SpaceX went public on June 12, 2026 — the largest IPO in history at $75 billion raised, valuing the company at over $1.75 trillion. The prospectus mentions orbital data centers as a long-term growth vector, with revenue-generating deployment “as early as 2028.” Investor briefings suggest demo missions could fly by late 2027.

The name Starmind fits a pattern. Starlink. Starshield (the government/military variant). Now Starmind — the compute layer. Each builds on the last. The laser mesh, the bus architecture, the manufacturing line, the launch vehicle — all shared. The marginal cost of adding compute to an existing satellite platform is, in theory, far lower than building from scratch.

Whether a million satellites ever fly is an open question. Musk’s timelines have a habit of stretching. But the prototype hardware is real. The factory is being built. The FCC application is on the record. And the economic pressure — AI compute demand doubling every few months, terrestrial infrastructure straining — isn’t easing.

If Starmind flies, it won’t look like a traditional satellite constellation. It’ll look like a distributed supercomputer wrapped around the planet, running inference for billions of users, powered by sunlight, cooled by the void. The name, at least, is settled.