Nvidia AI Leap: Groq 3 LPX Powers Future Inference

Nvidia announced that its new Groq 3 LPX, an advanced artificial intelligence inference system designed for low latency, has entered full production. This system is slated to become operational before the conclusion of 2026. Nebius is set to be the inaugural AI cloud to implement Groq 3 LPX via its Nebius Token Factory platform.

This production milestone follows an significant $20 billion licensing agreement between Nvidia and the chip designer Groq Inc., which was finalized on December 24, 2025. Nvidia also onboarded Groq's founder, CEO Jonathan Ross, president Sunny Madra, and a substantial portion of their engineering team, though Groq remains an independent entity.

The Groq 3 LPX system, featured with 256 language processing units (LPUs) and liquid cooling, is integrated into Nvidia's Vera Rubin platform. While Nvidia's GPUs handle AI training, the LPX is optimized to accelerate inference workloads where speed is paramount. Its design prioritizes on-chip memory for inference efficiency.

When paired with Nvidia's Vera Rubin chips, the Groq 3 LPX system is projected to deliver up to 35 times greater inference throughput per megawatt of power. This enhanced efficiency is crucial given the current industry-wide constraints on electricity access. Crucially, customers can utilize the LPX alongside Vera Rubin chips without altering their existing CUDA workflows, a software layer fundamental to Nvidia's dominance in the AI sector.