SpaceX's AI unit is deploying Nvidia's new Vera CPU built specifically for AI agents

[CrimsonWolf]

Nvidia announced this week that SpaceXAI, the AI unit within Elon Musk's SpaceX, will deploy Nvidia's new Vera CPU to accelerate its next generation of agentic AI applications. The company is calling Vera the first CPU built specifically for AI agents, designed to handle the orchestration, code execution, and data processing work that surrounds a model's actual inference rather than the inference itself.

The core idea behind Vera addresses a bottleneck that's become increasingly obvious as agentic AI systems have matured. Agents don't just generate a single response and stop, they call tools, execute code, process data between model calls, and coordinate multi step tasks, and all of that surrounding orchestration work traditionally runs on general purpose CPUs that weren't designed with agent workloads specifically in mind. Vera packs 88 Nvidia designed Olympus cores along with a new Spatial Multithreading technology and high bandwidth LPDDR5X memory delivering up to 1.2 terabytes per second, and Nvidia claims up to 1.8 times faster task completion compared to x86 CPUs across agentic AI, reinforcement learning, and data processing workloads specifically.

SpaceXAI plans to use Vera to expand the AI infrastructure behind Grok, running on Nvidia's broader Vera Rubin platform as the company scales toward gigawatt levels of computing capacity. That platform combines Nvidia's accelerated computing, NVLink interconnect technology, Spectrum-X networking, and BlueField data processing into one integrated architecture aimed at maximizing performance and energy efficiency while driving down the cost of each token generated.

What makes this announcement stand out beyond the usual data center chip news is the stated plan to take the same architecture into orbit. SpaceXAI is developing a first generation satellite called Starmind, based on an optimized version of the Vera Rubin NVL72 rack scale system, extending the same accelerated computing architecture that powers terrestrial AI factories into a space based deployment where power, thermal management, bandwidth, and physical integration constraints look completely different from a conventional data center. Nvidia framed this as one common computing foundation stretching from Earth based AI factories all the way to orbital infrastructure, a considerably more ambitious pitch than a typical hardware partnership announcement

SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scalenvidianews.nvidia.com

[Chris27]

Grok scaling toward gigawatt level compute capacity is a staggering number to see stated so casually in a press release. That's an entire power plant's worth of continuous electricity draw dedicated to one company's AI infrastructure, which says a lot about where the actual physical constraints on AI scaling are heading over the next few years.

Power availability, not chip supply or even data center construction, increasingly looks like the real ceiling on how fast any of these companies can actually scale their compute. Musk's various companies control enough vertically integrated infrastructure, energy, satellites, chips through this partnership, that SpaceXAI might actually be better positioned than most competitors to work around that constraint though

[Violet Caitlin]

1.8x faster task completion sounds impressive but that number is obviously cherry picked from whichever benchmark made Vera look best against x86 competitors. Would want to see independent testing before taking that multiplier at face value

[Sharp Shannon]

The Starmind satellite detail is the part that actually surprised me here, since running Vera Rubin hardware in orbit introduces a completely different set of engineering constraints than anything Nvidia has had to solve on Earth. Thermal management specifically becomes a much harder problem in space, since there's no atmosphere to convect heat away the way a terrestrial data center can rely on. Radiation hardening, power budget limits from solar arrays, and the total inability to just walk in and swap a failed component all add up to a genuinely different engineering challenge than scaling a data center on the ground. Curious how much of the Vera Rubin architecture actually survives that transition unmodified versus needing a substantially redesigned version specifically for orbital deployment

[QubitZero68]

The framing around keeping GPUs fully fed and utilized is worth taking seriously as a real economic argument, not just marketing language. GPU idle time waiting on CPU orchestration work is expensive dead time on some of the most valuable hardware in the world right now. Even a modest efficiency gain there compounds into real money at the scale SpaceXAI is apparently operating at

[TheRizz07]

A CPU specifically designed around agent orchestration rather than general purpose computing is a pretty clear signal about where Nvidia sees the actual bottleneck sitting right now. Keeping GPUs fully fed and utilized during multi step agent workflows is apparently a bigger problem than raw inference throughput at this point