SpaceXAI to Deploy Nvidia's Standalone Vera CPU for Grok Agents, Both in Data Centers and in Orbit

The deployment, announced by Nvidia on Monday, makes SpaceXAI the second hyperscaler after Meta to commit to Vera outside of full Vera Rubin racks. The decision underscores a growing industry trend: as AI agents become more prevalent, the bottleneck in training and inference is shifting from GPU compute to the CPU orchestration layer.

Agents spend much of their runtime off the GPU. A model that writes code, for example, calls tools, queries databases, and parses results by leaning on the host CPU between every inference pass; idle GPUs waiting on that work represent wasted capital. "Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code, and data processing while keeping GPUs doing what they do best," said Mike Nicolls, president of SpaceXAI, in the release.

SpaceXAI, which operates the Grok large language model developed by Elon Musk's xAI, will use the standalone Vera CPUs to accelerate agentic workloads in its data centers. The company also plans to launch a space-optimized Vera Rubin NVL72 system to orbit in Q4 2027, with significant scale expected in 2028. The orbiting system, called Starmind, would bring Nvidia's AI hardware into space for the first time, potentially enabling low-latency inference for satellite communications and on-orbit data processing.

Meta was the first company to announce a large-scale standalone Vera deployment, confirming in February that it would run Grace-only servers in production with Vera to follow as soon as 2027. Nvidia has since disclosed shipping hundreds of thousands of standalone Grace servers and more than 2.5 million Grace CPUs in total, signaling a pivot from the company's traditional GPU-centric messaging.

Critics note that Nvidia's performance claims of 1.8x improvement over x86 have not been independently verified, and the challenges of operating high-performance computing hardware in space—radiation, thermal management, and power constraints—remain significant. The Starmind satellite's design and capabilities have not been detailed beyond the planned NVL72 system.

Why This Matters

The deployment signals a shift in AI infrastructure: as agentic AI grows, the CPU becomes a critical bottleneck, and Nvidia's standalone CPUs are vying for a role previously dominated by x86 server chips.
Space-based AI computing could enable new applications, such as real-time data processing for satellite constellations, potentially reducing latency and bandwidth needs for Earth observation and communications.
The partnership between SpaceXAI and Nvidia ties together two of the most capital-intensive trends in technology: AI infrastructure and space commercialization.

Background

Nvidia's Vera CPU, based on the 88-core Grace architecture, was originally designed as the host processor for the company's Grace Hopper and Grace Blackwell superchips. However, as AI workloads evolved from pure training to agentic loops that require constant CPU orchestration, Nvidia began marketing Vera as a standalone server CPU. Meta was the first hyperscaler to announce standalone Vera deployment in February 2026, with production expected in 2027. SpaceXAI, formed in 2024 as a joint venture between SpaceX and xAI, operates Grok and develops AI infrastructure for Musk's ventures. The Starmind satellite program aims to put AI compute in orbit, a concept that has been discussed for years but never realized at scale.

Key Perspectives

SpaceXAI: The company sees Vera as the key to scaling Grok's agentic capabilities efficiently. By putting AI hardware in orbit, it could offer unique low-latency AI services for satellite-based applications. Mike Nicolls emphasized the importance of CPU performance for keeping GPUs fully utilized. Nvidia: The company is expanding its CPU business beyond the traditional GPU-accelerated role, positioning Vera as a direct competitor to AMD and Intel server processors. The SpaceXAI deal is a high-profile validation of this strategy. Critics/Skeptics: The claimed 1.8x performance improvement over x86 is unverified, and Nvidia's CPU architecture has limited software ecosystem compared to x86. Operating such hardware in space introduces severe reliability challenges—radiation hardening, thermal management, and power constraints are all non-trivial. The Starmind satellite's cost and timeline remain uncertain.

What to Watch

Independent benchmarks of Vera CPU performance against x86 competitors (expected in late 2026 or early 2027).
Launch of the Starmind satellite, currently scheduled for Q4 2027, and any delays or design changes.
Adoption of standalone Vera by other hyperscalers; if Meta and SpaceXAI are followed by others, it could signal a real shift in data center CPU architecture.

ZOTPAPER