SpaceXAI Picks NVIDIA Vera Rubin for Starmind—and Targets Q4 2027 Orbital Launch

SpaceXAI has chosen the computing architecture it wants to carry from Earth’s largest AI facilities into orbit.

NVIDIA announced Monday that SpaceXAI will deploy its Vera CPUs for the control-heavy work behind next-generation AI agents. The same partnership reaches much farther: an optimized Vera Rubin NVL72 system is being adapted for SpaceX’s first-generation Starmind satellite.

Elon Musk says the space-qualified system is targeted to launch in the fourth quarter of 2027, with significant scaling planned for 2028.

This is more than another chip-supply announcement. SpaceXAI is trying to build one computing foundation for Grok on the ground and a new class of high-power AI satellites above it.

NVIDIA laid out the partnership and the division of labor in its official announcement.

The first CPU built for agents is going to work at scale. @SpaceX is deploying NVIDIA Vera to accelerate the orchestration, code execution, and data processing that powers its next generation of agentic AI — keeping GPUs fed and agents acting fast.

From gigawatt AI factories… pic.twitter.com/xJrxAw9EtQ

— NVIDIA (@nvidia) August 24, 2026

NVIDIA says Vera will handle the work that piles up around the GPU: tool coordination, code execution, data processing and simulations between model calls. That is the less glamorous part of an AI system, but it can become the bottleneck when thousands of agents are trying to act at once.

Vera carries 88 NVIDIA-designed Olympus cores and uses Spatial Multithreading to run 176 threads. NVIDIA lists up to 1.2 terabytes per second of LPDDR5X memory bandwidth.

The company says the design can complete certain agentic AI, reinforcement-learning and data-processing tasks up to 1.8 times faster than x86 CPUs. Those are NVIDIA’s own performance claims, not independent benchmark results.

SpaceXAI plans to use Vera Rubin as it expands the infrastructure behind Grok toward gigawatt scale. NVIDIA describes a common architecture spanning CPUs, GPUs, networking, data processing and software, with the goal of keeping the expensive accelerators busy instead of waiting on control and data work.

The orbital piece is where the announcement becomes unmistakably SpaceX.

The first Starmind satellite will not carry a conventional terrestrial rack untouched. NVIDIA and SpaceXAI say they are adapting an optimized Vera Rubin NVL72 system around the realities of orbit, where power generation, heat rejection, bandwidth, radiation exposure, reliability and physical integration all become spacecraft problems.

Musk put a date on that plan and pushed the story beyond an open-ended research project.

SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028 https://t.co/qdDq8YBkzl

— Elon Musk (@elonmusk) August 24, 2026

SpaceX describes AI1 as a very large solar-powered computing satellite. The published design stands 20 meters, or 65 feet, tall when deployed and stretches 70 meters, or 229 feet, across.

Its listed compute payload draws 120 kilowatts on average and peaks at 150 kilowatts.

The company says the satellites would operate in sun-synchronous orbit, where their solar arrays can capture steady sunlight without clouds or atmosphere in the way. Heat would radiate into space rather than being removed by the chillers, cooling towers and fans used by terrestrial data centers.

Results would return to Earth through high-bandwidth laser links connected to Starlink. That makes Starmind less like a data center lifted into space and more like a new layer built into SpaceX’s existing launch, satellite-manufacturing and communications network.

There is still a huge engineering gap between a published architecture and a dependable orbital computing service. A machine designed for a controlled data center has to survive launch vibration, radiation, thermal cycling and a maintenance environment where a technician cannot simply swap a failed board.

The scale SpaceX is describing also depends on Starship. Its Starmind page calls the fully reusable rocket’s payload capacity critical for deploying the large, heavy spacecraft.

Manufacturing is the other half of the plan. SpaceX says its proposed Gigasat Factory in Bastrop, Texas, is intended to support rapid production and deployment of thousands of AI satellites beginning as soon as late 2027.

NVIDIA’s Vera overview explains why the CPU sits at the center of this architecture even in a world dominated by GPU headlines. AI agents call tools, run code, inspect data, manage state and coordinate long chains of actions.

All of that work places pressure on the host processor before and after the GPU handles a model request.

Vera is built to keep that surrounding work moving while Rubin GPUs handle the heavy model computation. NVIDIA connects the two through its NVLink-C2C system, which the company rates at up to 1.8 terabytes per second of coherent bandwidth between CPU and GPU.

Using the same underlying platform on Earth and in orbit could simplify software development and workload movement. SpaceXAI still has to solve the hardware challenge of operating a high-power computer in space, but it now has a defined starting point instead of a collection of disconnected experiments.

The timeline remains a target. Q4 2027 is not a completed launch contract, and “significant scale” in 2028 leaves major questions about satellite count, launch cadence, cost and customer demand.

But the architecture is now named, the partners are public and the first flight window is on the calendar.

SpaceXAI is betting that the next great expansion of AI infrastructure will not stop at the edge of Earth’s electrical grid. With NVIDIA Vera Rubin inside Starmind, it intends to test that bet in orbit.