Nvidia begins mass production of Groq 3 LPX AI inference chips

Aug 24, 2026

Nvidia has commenced mass production of its Groq 3 LPX AI inference accelerator racks, which are built with 256 language processing units manufactured on Samsung's 4-nanometer process. Designed to resolve compounding decode latency in agentic AI, the SRAM-based system achieved a record 3,400 tokens per second in independent benchmarks. Neocloud provider Nebius will be the first to deploy the hardware before December 31, 2026, as Nvidia competes against low-latency offerings from AMD and Cerebras.

Groq 3 LPX production launch

▪Amsterdam-headquartered neocloud Nebius will be the first AI cloud to deploy the Groq 3 LPX hardware through its Token Factory platform before December 31, 2026.
▪The Groq 3 LPX rack integrates 256 language processing units in a liquid-cooled chassis, delivering 128 gigabytes of aggregate on-chip SRAM.
▪Nvidia announced at the Hot Chips 2026 conference on August 24, 2026, that its Groq 3 LPX inference accelerator rack has entered full production.

SRAM decode architecture advantages

▪The Groq 3 LPX system achieved 3,400 output tokens per second running Gemma 4 31B in benchmarks conducted by Artificial Analysis.
▪Each Groq 3 LPU carries 500 megabytes of on-chip SRAM, which is approximately 576 times less than the HBM4 capacity of a single Nvidia Vera Rubin GPU.
▪The Groq 3 LPU stores model weights in on-chip static random-access memory (SRAM) to deliver 150 terabytes per second of bandwidth per chip.

Agentic AI inference bottlenecks

▪The decode phase of large language model inference is memory-bandwidth-bound because the model must load its full weight tensor from memory for every token generated.
▪Nvidia claims the Groq 3 LPX delivers a 4x reduction in decode latency compared to the nearest GPU-based alternative.

Samsung foundry manufacturing partnership

▪Samsung Electronics is the sole manufacturer of the Groq 3 LPX's language processing units, producing them on its 4-nanometer foundry lines.
▪KB Securities projects that the ramp-up of 4-nanometer LPU production and a 15 percent price increase could return Samsung's foundry business to profitability in Q3 2026.
▪Nvidia acquired the assets and technology of chip startup Groq in December 2025 for a reported $20 billion licensing agreement.

AI inference market competition

▪Cerebras Systems announced its CS-4 system on August 18, 2026, claiming it delivers up to 30 times faster inference than GPU systems on trillion-parameter models.
▪Advanced Micro Devices announced plans to integrate its rack-scale systems with chips from Cerebras to focus on low-latency AI inference.