Grok compute heads to orbit on a Vera Rubin NVL72

SpaceXAI plans to put a rack-scale NVIDIA system into orbit: the first-generation Starmind AI satellite is based on a Vera Rubin NVL72 optimized for space, designed jointly by SpaceX and NVIDIA. The plan was laid out in an announcement published on Threads.

At a glance

• The orbital variant reworks a rack built for data centre floors into a satellite payload, and according to Crypto Briefing the design comes out simpler, cheaper, denser and lighter than a traditional rack.
• On the ground, SpaceXAI is expanding the AI infrastructure behind Grok with the NVIDIA Vera Rubin platform, and says it is scaling toward gigawatts of computing capacity.
• The same platform sits on both sides of the announcement, an orbital unit for Starmind and a terrestrial expansion for Grok, rather than separate hardware lines for the space and ground workloads.

Moving an NVL72 off the data centre floor reads as an attempt to treat orbit as another deployment site for standard accelerated computing rather than a niche for bespoke space hardware. The constraints differ on each side: ground capacity is limited by power, which is likely why SpaceXAI counts its build-out in gigawatts, while a satellite is limited by mass and volume, the two properties the reworked design is said to improve.

The orbital unit is a rework of hardware designed for terrestrial data centres. According to Crypto Briefing, the space version comes out simpler, cheaper, denser and lighter than a traditional rack. Starmind is the satellite it is meant to fly on, described as a first-generation system.

Vera Rubin is NVIDIA's platform generation following Blackwell, and NVL72 is the rack-scale configuration that pairs Vera CPUs with Rubin accelerators in a single unit. The same designation now covers both the terrestrial systems SpaceXAI is installing and the space-optimized variant built for orbit.

On the ground, SpaceXAI says it is expanding the AI infrastructure behind Grok with the NVIDIA Vera Rubin platform as it scales toward gigawatts of computing capacity. The CPU half of that expansion is stated directly in the announcement's own wording:

SpaceXAI will deploy NVIDIA Vera CPUs to accelerate the work behind its next generation of agentic AI workloads.

When Starmind is due to fly

The announcement pointed to an orbital launch later this year; the launch is planned for Q4 2027. No mass, power or configuration figures have been given for the flight hardware, and no timeline was attached to the gigawatt-scale ground capacity or to the sites that would host it. Starmind is described only as a first-generation system, with no follow-on units named.