Two companies people talked about as rivals just a year ago showed off a shared rack this week. Groq 3 LPX was designed together with NVIDIA's new platform, and it is no longer a demo. It's a product.
Fact
The facts: Groq 3 LPX is a rack-level inference accelerator, co-designed with NVIDIA's Vera Rubin NVL72 platform. It combines GPUs with dedicated fast-response modules and handles token generation for agentic systems. By NVIDIA's own numbers it delivers 3,400 output tokens per second on tasks with a 100,000-token context on the Gemma 4 31B model, 4x faster than the closest alternative platform. It is already in full production and deploying at Nebius, CoreWeave and SpaceXAI. Source: the NVIDIA blog.
Who is who here
The page stays quiet on this. It does not say what the ownership looks like, or what the deal between the two companies is. I am not guessing. I am flagging it, because on a jointly designed rack that is the first question anyone buying capacity would ask.
One side does the heavy math, the other lines up the words fast. Together they build a machine that never stops writing while the agent thinks.
The division of labor is the whole story. The heavy computing still sits with the GPUs, while fast token output runs through the dedicated modules. If you are renting capacity this fall from any of the three providers, ask directly which chip does which part. The answer changes both the price and the speed you get.
Official primary sources
NVIDIA Blog: Vera Rubin LPX, Spectrum-X, NVLink FusionThe visual is generated code-art, no third-party images.