NVIDIA announced on the 24th (local time) that SpaceXAI, which operates Grok, the AI chatbot for Elon Musk’s camp, will adopt its first CPU designed for AI agents, called “Vera.”
Vera features 88 in-house designed cores and can process agent tasks such as code execution and data handling up to 1.8 times faster than conventional x86 CPUs.
SpaceXAI will scale the AI infrastructure behind Grok to gigawatt levels using NVIDIA’s “Vera Rubin” platform. One gigawatt is roughly equivalent to the electricity generation capacity of a single nuclear power plant.
The company also unveiled plans to bring the same system into space. The first-generation AI satellite, “Starmind,” will carry a Vera Rubin server rack modified for the space environment.
As ground-based data centers face power shortages and regulatory hurdles, orbital data centers powered by solar energy are drawing attention as a possible alternative.
AI computing in Elon Musk’s camp is moving beyond Earth and into space.
NVIDIA said on the 24th (local time) that SpaceXAI, which operates Grok, the AI chatbot, will adopt its “Vera” CPU for next-generation AI agent services. The plan is to expand the Grok infrastructure with NVIDIA’s latest “Vera Rubin” platform and pursue a first-generation AI satellite, “Starwind,” equipped with the same system.
◆ In the age of agents, the CPU takes center stage again
Until now, the main star of AI computing has been the graphics processing unit, or GPU. That is because GPUs handle the core computations that allow AI to generate answers. But the rise of AI agents that receive instructions and act on their own is changing the situation.
Agents do not merely produce answers; they must constantly handle chores such as calling tools, running code, and organizing information. The semiconductor that handles those chores is the CPU. If the CPU is slow, expensive GPUs are left idle, waiting for work.
Vera is the first CPU designed to address that bottleneck for agents. It has 88 Olympus cores designed in-house by NVIDIA and a memory bandwidth of up to 1.2 terabytes per second.
NVIDIA says it is up to 1.8 times faster than conventional x86 CPUs at completing tasks in agentic AI, reinforcement learning, and data-processing workloads. Mike Nichols, president of SpaceXAI, said, “With the same amount of power, it will get more useful work done.”
SpaceXAI will grow its AI infrastructure to gigawatt scale using the Vera Rubin platform, including Vera. One gigawatt is a power scale comparable to the output of a single nuclear power plant.
◆ Why send data centers into space?
The most eye-catching part of the announcement is space. Starwind, the first-generation AI satellite being prepared by SpaceXAI, will carry a Vera Rubin NVL72 server rack, optimized for the space environment, that is used in ground-based AI data centers.
In other words, it is a structure similar to the computer running Grok on Earth, placed into orbit. Ian Buck, NVIDIA vice president, said it was “bringing the architecture used in massive AI factories to orbit, the next frontier of computing.”
The reason space data centers are being discussed is the limitations on Earth. Around the world, data centers are running into power shortages, water scarcity, local opposition, and construction regulations.
In space, electricity can be obtained from solar power, and there are no land-use or local community conflicts. On the other hand, challenges such as launch costs, thermal management, communication bandwidth, and repairs in the event of malfunctions are significant, so some point out that time is needed before economic viability can be proven.
As the battleground shifts from chips to electricity and then to location, Musk’s camp, which has rockets at its disposal, has chosen orbit as its first card. If Starwind proves its performance in actual orbit, the rules of the data-center location race could be rewritten.