SpaceXAI is set to enhance its use of NVIDIA’s infrastructure for developing agent systems and Grok. The company will implement Vera processors and aims to scale the Vera Rubin computing platform to gigawatt capacities for space deployment.
https://twitter.com/nvidia/status/2091920680317046847
Vera Processors for AI Agents
NVIDIA has announced that SpaceXAI will utilize the Vera CPU for its next-generation agent AI systems. Unlike graphics processing units (GPUs), which mainly handle model computations, these processors are designed to accelerate tasks related to inference: executing code, managing tools, processing data, orchestrating agents, and running simulations.
The Vera processor features 88 custom Olympus cores and LPDDR5X memory with a bandwidth of up to 1.2 TB/s. According to NVIDIA, in various agent AI workloads, reinforcement learning, and data processing, the processor delivers task performance that is 1.8 times faster than comparable x86 chips.
This capability is crucial for agent systems: while GPUs focus on the model's direct operations, the CPU must continually coordinate the agent's actions—launching tools and code, transferring data, and relaying new tasks back to accelerators. NVIDIA aims to enhance GPU utilization and overall efficiency of AI infrastructure through this approach.
SpaceXAI Expands Grok
At the same time, SpaceXAI plans to boost its infrastructure for Grok using the Vera Rubin platform, targeting computational power scaling to a gigawatt level.
This partnership extends beyond traditional data centers. SpaceXAI is developing the first Starmind AI satellite, which will feature a computational system built on an optimized version of NVIDIA Vera Rubin NVL72. The companies intend to adapt the platform for space operations considering constraints related to power consumption, cooling, bandwidth, and reliability.
Elon Musk previously stated that SpaceX would exclusively use NVIDIA for its AI infrastructure, describing the Vera Rubin architecture as the best available option.
“Our design is significantly simpler, cheaper, denser, and lighter than traditional racks,”Musk wrote.
By the end of 2026, the company aims to achieve over 2 GW of computational capacity and approach 10 GW by the end of 2027.
From Data Centers to Orbital Computing
NVIDIA and SpaceXAI are effectively constructing a unified technological ecosystem: the Vera CPU will handle agent logic and data processing, Vera Rubin will support the core AI infrastructure for Grok and model training/inference, while the optimized NVL72 will serve as the foundation for future orbital computing.
The first Starmind AI satellite is scheduled for launch in 2027. Additionally, SpaceX is rapidly expanding its terrestrial computing infrastructure for Grok.
As a result of this partnership expansion, SpaceX is becoming another significant client for NVIDIA’s entire AI architecture, encompassing not just GPUs but also proprietary CPUs, networking equipment, and software platforms.
In August, the SpaceXAI team introduced Grok 4.6, a new flagship model focused on long-term agent tasks.