Groq tech woven into NVIDIA racks
Reports show NVIDIA is integrating Groq racks into its Vera Rubin inference setups after a recent Groq deal — the new Groq 3 LPU reportedly ships with 500MB SRAM per chip and 150TB/s bandwidth, delivering huge inference throughput claims reported. That integration tightens NVIDIA’s end‑to‑end play and raises questions about portability for customers wanting non‑proprietary silicon paths.
NVIDIA unveiled the Vera Rubin platform and its associated LPX inference racks on March 16, 2026 at GTC 2026 ([nvidianews.nvidia.com)]. NVIDIA’s LPX inference rack is built around a 256‑accelerator tray of Groq 3 LPU units and is being integrated with the company’s MGX ETL rack architecture. (developer.nvidia.com) NVIDIA told customers the seven‑chip Vera Rubin family is now in full production and CRN reports the Groq‑based LPX racks will ship in the second half of 2026. (nvidianews.nvidia.com) NVIDIA completed a high‑profile purchase of Groq’s assets for roughly $20 billion, a deal first reported on December 24, 2025, that underpins the licensing and engineering path for the LPUs. (cnbc.com) NVIDIA’s Rubin GPUs list 22 TB/s of HBM4 memory bandwidth per chip, while the LPX architecture exposes 2.5 TB/s of scale‑up bandwidth for LPU interconnect inside the rack. (theregister.com) NVIDIA paired the hardware shift with software for agentic inference—announcing the NemoClaw stack for the OpenClaw agent platform and naming support from OpenAI, Anthropic and Meta on Rubin—positioning LPX as part of a software‑hardware bundle. (nvidianews.nvidia.com) The MGX modular rack spec is published and already adopted across OEM/ODM partners and rack vendors (Open Compute MGX spec, Gigabyte/MGXS partners, and Rittal’s MGX rack availability), creating a standardized chassis ecosystem for deploying Rubin+LPX at scale. (opencompute.org)
