SpaceXAI is deploying NVIDIA Vera CPUs in order to power its agentic AI application’s next-generation CPU intensive tasks. This will expand an AI infrastructure that SpaceXAI says has been built behind Grok model, and which the company claims scales up toward a gigawatt computing capacity. NVIDIA announced on August 24, 2026.
This deal goes far beyond the data center. SpaceXAI is planning to build its Starmind AI satellite, the first generation of which will be based on an optimized platform. NVIDIA Vera Rubin NVL72 Rack-scale system that uses the same AI factory architecture as its on-ground AI factories to power it into space.
This announcement is about CPUs, making it unusual for an entire industry which measures all things in GPUs. Agentic AI workloads — systems that take actions rather than simply generate answers — lean heavily on conventional processors to orchestrate tools, execute code, process data, and run simulations between model calls. These GPUs are idle if the CPUs get slow. Vera was designed to target this bottleneck.
“Agentic AI requires a new kind of computing system — one built not only to generate answers, but to take action,” Ian Buck said, NVIDIA Vice President of Hyperscale and High-Performance Computing. “SpaceXAI is taking this architecture from massive AI factories to the next frontier of computing in orbit.”
“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,” Mike Nicolls said in a press release that he is the president of SpaceXAI.
The Vera Rack
Vera is NVIDIA’s very first processor designed specifically for AI agents, rather than adapting general-purpose servers. This CPU is equipped with 88 Olympus Cores and NVIDIA Spatial Multithreading Technology, which allows for 176 threads to be created with partitioned core resource. NVIDIA claims that this combination can complete tasks up to 1.8x faster than x86 processors in agentic AI workloads such as reinforcement learning and data processing, but also up to 80percent faster when used on sandboxes. according to the company’s product documentation.
Memory subsystems are where the architecture differs most from the x86 standard. Vera’s memory subsystem is where the architecture diverges the most from the x86 status quo. It supports up to 1,5 terabytes in memory per socket. Second-generation fabric on the die connects 88 cores at 3.4TB/s bisectional speed, eliminating cross-chiplet latencies that are common with multi-die servers CPUs. NVLink C2C provides up to 1.8 TB/s coherent bandwidth for Vera GPUs.
Vera’s CPU rack is an independent platform which integrates 256 Veras CPUs within a liquid-cooled, dense chassis. This allows for more than 22,500 concurrent sandboxes, or isolated software environments, where agents execute code, invoke tools, and run through iterative evaluation loops. Each reasoning cycle in agentic systems can generate thousands of environments. The CPU fleet hosting them sets the ceiling on the AI factory’s throughput.
Starmind Extension
SpaceXAI’s ambitions in orbit give the deployment of AI a second level. It is building AI computing infrastructure in orbit where constraints such as power, thermal, bandwidth, reliability and physical integration are not present at all. Starmind’s proposed first-generation AI satellite is built on a Vera Rubin NVL72 optimized system. This rack-scale platform combines 72 Rubin CPUs with 36 Vera GPUs, ConnectX-9SuperNICs, BlueField-4 Data Processing Units linked by 6th-generation NVLink Switches.
NVIDIA said that it is working with SpaceXAI to adapt the platform for orbital needs while maintaining a common software and architecture ecosystem between ground-based and space-based deployments.
One architecture, Gigawatts Capacity
Vera is a great way for SpaceXAI to consolidate its computing stack as it grows. SpaceXAI expands its AI Infrastructure behind Grok, on the NVIDIA vera rubin platform. This is as the company scales to gigawatts in computing power. Vera Rubin is a platform that provides a solid foundation for training, reasoning, inference and networking.
NVIDIA’s flagship product is a launch of a line launched to address the CPU-side of agentic AI. Vera, according to the company, is the first AI CPU. SpaceXAI also joins Anthropic OpenAI Oracle Cloud Infrastructure and other early deployments.
Economical considerations are key. NVIDIA touts to Vera Rubin that it can deliver more tokens per Watt and lower costs per Token than Blackwell. So faster CPU orchestration, which keeps GPUs occupied, translates into more outputs per watt. The efficiency of SpaceXAI will increase as it expands to gigawatts. NVIDIA has positioned the Vera Rubin Platform on this basis, with more tokens per Watt and lower cost for each token than Blackwell.

