Latitude
—
A 60 kg box the size of a bar fridge, carrying the first data-centre-class GPU ever flown. It has been in orbit since 2 November 2025, and it has already trained a language model up there.
| Date | Rise | Peak | Set | Max elev. | Duration | Sunlit at peak |
|---|---|---|---|---|---|---|
| Computing… | ||||||
Passes above 10° elevation over the next 5 days. Starcloud-1 is small and low — at magnitude terms it is not an easy naked-eye object, but a sunlit pass at high elevation is worth a look just after dusk.
Starcloud's entire commercial argument is sunlight. The company's FCC application for 88,000 satellites puts them in sun-synchronous orbit at 600–850 km, where a satellite stays in sunlight almost permanently and the solar panel does roughly eight times the work it would on the ground.
Starcloud-1 is not in that orbit. It flew on Bandwagon-4 — SpaceX's mid-inclination rideshare series, which exists precisely because it is not the sun-synchronous Transporter line. The Space-Track element set puts it at 45.40° inclination, 501 km, circular, 94.7 minutes per revolution.
Propagate that element set through one revolution and the satellite is in sunlight roughly 62% of the time. A dawn-dusk sun-synchronous orbit — the one in the application — sits above 99%. That is the orange and blue on the map above: it is not decoration, it is the gap between the pitch and the hardware. Which is fine: it is a technology demonstrator, and it proved the chip survives. But it did not demonstrate the power thesis, because it is not in the orbit the power thesis needs.
Gunter's Space Page and Data Center Dynamics both give Starcloud-1's altitude as 325 km. The element set and the Bandwagon-4 target orbit both say roughly 504–510 km. Coverage describing Starcloud as flying sun-synchronous orbits is describing the 88,000-satellite application, not the satellite that is actually up there.
Starcloud-1 was pitched with an eleven-month operational life, after which the spacecraft performs a controlled deorbit and disintegrates in the atmosphere. It launched on 2 November 2025. Eleven months puts the end of mission at roughly 2 October 2026.
Compare two element sets and you can measure the decay directly. Between the epoch of 1 June 2026 and 15 July 2026 — 44.5 days — mean altitude fell from 502.5 km to 501.1 km. That is 30 metres per day.
At that rate Starcloud-1 does not come down in eleven months. It does not come down in eleven years. Atmospheric drag climbs steeply as a satellite loses height, so a linear projection overstates the tail badly — but not by the order of magnitude required to turn three decades into three months. At 501 km, natural decay is a decade-plus proposition.
Which means the deorbit is not a matter of letting physics take its course. It depends entirely on propulsion firing, on schedule, on a 60 kg demonstrator. If it fires, the commitment is met and this paragraph is a footnote. If it does not, a dead GPU stays in low Earth orbit long past the FCC’s five-year post-mission disposal rule, and there is no passive backstop.
That matters more than one satellite. Starcloud filed with the FCC in February 2026 for a constellation of up to 88,000 satellites. The disposal case for a fleet that size rests on deorbit working as designed, every time. The first demonstrator is the only evidence anyone has. It is worth watching whether it comes down on 2 October — and this page will keep propagating the element set either way. The follow-up, Starcloud-2, launches late 2026 carrying Bitcoin-mining hardware, and inherits every one of these questions at a hundred times the power.
Downlink details from SatNOGS, flagged unconfirmed. Note the operator was renamed from Lumen Orbit to Starcloud in early 2025 after a challenge from Lumen Technologies — some catalogues still carry the old name.
November 2025 — Reached orbit and made contact. First data-centre-class GPU in space.
December 2025 — Ran Google's Gemma model on the H100: the first large language model operated on a high-powered GPU aboard a spacecraft. Then trained Andrej Karpathy's nanoGPT in orbit — the first in-orbit training run.
Ongoing — Running inference on synthetic-aperture radar data from Capella Space, drawing conclusions on orbit without downlinking the raw imagery first. This is the part that matters commercially: the value is not the compute, it is not moving the data.
From early 2027 — Crusoe plans to offer limited commercial access to orbital GPU capacity under an agreement with Starcloud.
—