SpaceXAI Adopts NVIDIA Vera CPUs for Gigawatt AI

SpaceXAI is integrating NVIDIA's newly announced Vera CPUs into its rapidly expanding computing infrastructure, which is scaling toward gigawatts of capacity to support its Grok models. Beyond terrestrial data centers, the company plans to launch its first-generation Starmind AI satellite equipped with an optimized NVIDIA Vera Rubin NVL72 rack-scale system. This orbital platform will combine 72 Rubin GPUs with 36 Vera CPUs, ConnectX-9 SuperNICs, and BlueField-4 data processing units, all linked by sixth-generation NVLink switches.

The decision to focus on CPUs highlights the unique demands of agentic AI, which requires heavy conventional processing to manage tools, run simulations, and execute code between GPU calls. The Vera CPU features 88 custom Olympus cores with Spatial Multithreading to deliver 176 threads. It utilizes detachable LPDDR5X memory on SOCAMM modules, providing up to 1.2 terabytes per second of bandwidth. According to NVIDIA, this architecture completes tasks up to 1.8 times faster than traditional x86 CPUs for agentic workloads and up to 80 percent faster in sandbox environments, while drawing about half the power.

For hardware practitioners, the chip offers up to 1.5 terabytes of memory per socket and a second-generation on-die fabric with 3.4 TB/s of bisectional bandwidth to eliminate latency. It also features an NVLink-C2C interface providing 1.8 TB/s of coherent bandwidth to Rubin GPUs. A standalone rack can house up to 256 liquid-cooled Vera CPUs, supporting more than 22,500 concurrent sandbox environments. By adopting this unified architecture, SpaceXAI, alongside other early adopters like OpenAI, Anthropic, and Oracle Cloud Infrastructure, aims to achieve more tokens per watt and a lower cost per token than NVIDIA's previous Blackwell generation.