[REVISION HISTORY]

NVIDIA Groq 3 LPX production and manufacturing

Updated 1 time since CLSTR started tracking revisions of this situation.

What changed

2026-08-26 14:35 UTC → 2026-09-14 16:52 UTC · added ~~removed~~

NVIDIA has entered full production of its Groq 3 LPX AI inference racks, ~~a move~~ following its December 2025 acquisition of Groq assets. The system is designed to complement NVIDIA’s Vera Rubin platform by specializing in the low-latency “decode” phase of AI model generation. Each rack integrates 256 Language Processing Units (LPUs) and utilizes 500 megabytes of on-chip SRAM to reduce memory bottlenecks. Samsung Electronics is the exclusive manufacturer of the hardware, utilizing a 4-nanometer fabrication process. ~~The~~ At the Hot Chips 2026 conference, it was noted that this mass production ~~of these chips is expected to potentially~~ could return Samsung’s foundry unit to ~~profitability,~~ profitability as early as the third quarter, serving as a financial inflection point for its contract chipmaking business. Nebius has been identified as the first AI cloud customer for the Groq 3 LPX system. To further expand its inference capabilities, NVIDIA has entered a multi-year collaboration with d-Matrix to integrate Raptor inference XPUs into NVIDIA MGX racks via NVLink Fusion. This partnership aims to provide a heterogeneous architecture that splits workloads between GPUs and XPUs, utilizing a 3D-DRAM architecture that offers a different technical approach to inference compared to the SRAM-based Groq 3 LPX. Additionally, SpaceXAI has committed to deploying NVIDIA Vera CPUs as part of its next-generation AI architecture, with plans to extend capabilities to orbital data centers.

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.