The orbit angle is what makes this deal different. SpaceXAI is putting NVIDIA’s Vera CPUs at the center of its agentic AI push, and the first Starmind satellite will carry the same rack architecture found in its ground data centers.
The news arrived Monday from NVIDIA, which described SpaceXAI’s compute expansion behind its Grok models as scaling toward gigawatts. For orbit, the firms are adapting a Vera Rubin NVL72 rack for Starmind’s first generation, keeping a common software base across ground and space.
CPUs seldom headline infrastructure deals, yet agentic workloads lean on them heavily. Orchestration, code execution, tool calls and simulation all run on conventional processors between model passes, and when that work stalls, the GPUs feeding on it sit idle. The silicon itself is new ground for NVIDIA. Vera is its first processor built specifically for agents, carrying 88 Olympus cores. A technique called Spatial Multithreading doubles the thread count to 176, while LPDDR5X memory moves data at up to 1.2 TB/s using roughly half the power of leading x86 chips.
A 256-CPU Vera rack supports more than 22,500 concurrent sandboxes, the isolated containers where agents run code and iterate. SpaceXAI president Mike Nicolls says the processor delivers the orchestration muscle needed to keep GPUs busy.
Anthropic, OpenAI and Oracle Cloud Infrastructure round out the early production deployments NVIDIA has named for the part.
