NVIDIA named the first Starmind rack. SpaceX named the satellite specs and a late-2027 factory target. Neither document is a launch. Neither is a customer SKU you can buy today.
NVIDIA and SpaceXAI put Vera Rubin NVL72 on gen-1 Starmind
On 24 August 2026 NVIDIA posted that SpaceXAI will deploy Vera CPUs for agentic workloads and expand Grok infrastructure on the Vera Rubin platform. The same release says SpaceXAI's planned first-generation Starmind AI satellite will be based on an optimized NVIDIA Vera Rubin NVL72 rack-scale system. NVIDIA and SpaceXAI are adapting that foundation for orbit while keeping a common NVIDIA architecture and software stack across Earth and space.
SpaceX's Starmind page describes AI1 as an orbital satellite with localized compute that radiates heat into vacuum and beams results back through high-bandwidth lasers into the Starlink constellation. Published AI1 figures: deployed height 30 m, wingspan 75 m, compute payload up to 250 kW peak / 175 kW average, vehicle efficiency 75 kW per ton. The page lists three limiting factors for gigawatt-scale space compute: mass to orbit, power generation, and AI chips. Starship is named for mass. Sun-synchronous orbit is named for continuous solar. A joint SpaceX-Tesla Terafab is named for future chip production.
The same page says SpaceX is building a Gigasat Factory in Bastrop to manufacture AI satellites at scale, planned to enable rapid production and deployment of thousands of AI satellites starting as soon as late 2027.
SpaceX's S-1 defines orbital AI compute as satellite constellations that act as orbital data centers and states an expectation to begin deploying those satellites as early as 2028. In risk factors, SpaceX says initiatives to develop orbital AI compute at scale involve significant technical complexity and unproven technologies, and that such initiatives may not achieve commercial viability.
SpaceXAI, NVIDIA, and who is not buying capacity yet
SpaceXAI and NVIDIA authored the co-design claim for gen-1 Starmind hardware. SpaceX owns the AI1 vehicle design, Starlink laser return path, Gigasat Factory plan, and Starship mass-to-orbit dependency. Buyers and operators are not bound to a service tariff. There is no public Starmind capacity product on the page.
Hyperscale and AI infrastructure teams tracking orbital compute should treat this as a named rack target and a factory date, not as delivered FLOPS. Teams already watching Google's Project Suncatcher TPU prototype in orbit have a separate flight-hardware story; Starmind is still pre-launch design and production planning.
A named Earth rack SKU plus a Bastrop factory date
The NVIDIA release puts a specific Earth AI-factory SKU (Vera Rubin NVL72, optimized) on gen-1 Starmind, instead of a vague "space GPU" line. SpaceX's page puts hard AI1 power and size numbers next to a Bastrop factory timeline. That pairing is the news. It is not a completed orbital AI factory.
Dates, vendor wording, and the S-1 viability caveat
Date: NVIDIA post 24 August 2026; SpaceX Gigasat "as soon as late 2027"; S-1 deployment "as early as 2028." Hardware: optimized Vera Rubin NVL72 for gen-1 per NVIDIA; SpaceX's page also says the wider architecture is "AI chip vendor agnostic" and can take modules from any provider. Those statements sit next to each other. Price: not disclosed. Region / orbit: sun-synchronous orbit on the SpaceX page; FCC authorization status is not settled on these pages. Commercial service: not offered. Whether late-2027 volume production and 2028 deployment survive Starship cadence, radiation, thermal, and chip-refresh constraints remains unsettled. SpaceX's own S-1 says commercial viability is not assured.
Diary late 2027; do not book orbital FLOPS
Infrastructure leads should diary late 2027 / 2028 as company targets, not as booked capacity. Do not replace terrestrial power procurement with a Starmind assumption. If you already filed Project Suncatcher as the first public TPU-in-orbit test, keep Starmind on a separate track: announced NVL72 co-design and factory plan, no satellite on orbit yet. Re-read SpaceX's orbital-AI risk language before treating IPO or partnership decks as a delivery schedule.