Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
SpaceXAI's Vera CPU Is a Factory Plan, Not a Grok Bench
NVIDIA's Aug 24 newsroom post: SpaceXAI will deploy NVIDIA Vera CPUs for agentic workloads around Grok.
Key takeaways
- NVIDIA's Aug 24 newsroom post: SpaceXAI will deploy NVIDIA Vera CPUs for agentic workloads around Grok.
- The same note says SpaceXAI is expanding Grok infrastructure on Vera Rubin toward gigawatts, and plans a first-generation Starmind satellite on an optimized Vera Rubin NVL72.
- SpaceXAI has not posted this on its own news page. NVIDIA's 1.8x vs x86 is their number. I do not have a Vera box.
Local LLMs on NVIDIA Spark / ASUS GX10
NVIDIA posted it Aug 24. SpaceXAI will deploy Vera CPUs. Grok infrastructure expands on Vera Rubin, toward gigawatts. A first-generation Starmind satellite is planned on an optimized Vera Rubin NVL72.
I checked SpaceXAI's news page. The Vera note is not there. The quote in NVIDIA's release is from Mike Nicolls, president of SpaceXAI. That is still NVIDIA's byline.
Read the verbs. Will deploy. Plans to expand. Planned satellite. The instrument is a factory roadmap, not a Grok score.
NVIDIA's own spec sheet in that post: 88 Olympus cores, Spatial Multithreading, LPDDR5X, up to 1.2TB/s. They say Vera is up to 1.8x faster than x86 on agentic AI, reinforcement learning, and data processing. I am not going to launder that into a bench I did not run.
Same habit as the Ohio 8-K. A big NVIDIA number is an instrument. This one is a CPU SKU and a gigawatt sentence. It does not tell me whether Grok 4.6 is Sonnet-class. It does not replace the GX10.
I run Grok when it earns it, and a GX10 when I want the weights in the room. Vera is the cloud side of that split. If SpaceXAI posts the same deal on their own site, or I get a week of coding receipts, that note comes next. Not this press release.
Frequently asked questions
NVIDIA did, dated Aug 24, 2026. Mike Nicolls, president of SpaceXAI, is quoted in that release. As of this note, x.ai/news has no matching post.
NVIDIA calls it the first CPU built for AI agents. The card they print: 88 Olympus cores, Spatial Multithreading, LPDDR5X, up to 1.2TB/s of memory bandwidth. They claim up to 1.8x faster task completion than x86 on agentic AI, reinforcement learning, and data processing. That sentence belongs to them.
No. NVIDIA says SpaceXAI's planned first-generation Starmind AI satellite will be based on an optimized Vera Rubin NVL72. Planned is the word.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.