Nvidia Releases Groq 3 LPX Rack, Boosting AI Inference Throughput by 35x

Nvidia launched the Groq 3 LPX rack at GTC 2026, designed for low-latency and long-context needs of intelligent agent systems. Positioned as Vera Rubin's AI inference accelerator, it is expected to launch in the second half of this year.

The rack uses full liquid cooling and MGX infrastructure, with 32 1U compute trays each containing 8 Groq3 (LP30) chips—totaling 256 chips and 128GB on-chip SRAM, delivering 40PB/s bandwidth to speed up decoding and reduce latency.

Each LP30 chip connects to the system via 96 112Gbps C2C links, and a single tray can expand to 384GB DRAM for backup when SRAM is insufficient.

Nvidia states that Groq3 LPX increases Vera Rubin's per-megawatt inference throughput by 35x and creates 10x revenue opportunities for trillion-parameter models.