Samsung Foundry begins full mass production of Nvidia's Groq 3 LPX chip

It was announced five months ago that Samsung Foundry had started making Groq 3 LPX chips for Nvidia. Now, Nvidia has announced that full-scale mass production of those chips has begun, giving Samsung a major boost in confidence and revenue.

Nvidia's Groq 3 LPX is a specialized AI inference chip built around the Language Processing Unit (LPU). While traditional GPUs, like Nvidia's own GB200, are great at heavy parallel computing and model training, they can face bottlenecks during the inference phase, the stage in which a trained AI model generates responses and decodes tokens. This is where an LPU architecture comes in to solve memory-fetching bottlenecks and cache stalls.

Groq is an AI startup founded in 2016 by a group of former Google engineers and led by Jonathan Ross, who invented Google's famous Tensor Processing Unit (TPU). The company built chips that accelerated low-latency token generation.

Nvidia paid Groq $20 billion in December 2025 for a non-exclusive intellectual property licensing agreement and most of its key engineering talent. Nvidia then used Groq's LP30 chip, made on Samsung Foundry's 4nm process node, in its latest Vera Rubin platform.

This Groq chip is now in full-scale production. It handles specialized AI inference systems designed to drive AI agents and enhance response speeds. Each rack-scale deployment of Vera Rubin uses 256 LPUs, which are manufactured at Samsung Foundry's Pyeongtaek campus in South Korea.

In recent Artificial Analysis benchmarks using the open-source Gemma 4 31B model with a 100,000-token context window, the Groq 3 LPX platform hit 3,400 output tokens per second, making it four times faster than alternative platforms for long-context agentic workloads.

AI cloud operator Nebius is the first client to use the Groq 3 LPX platform to boost token generation speeds. Moreover, CoreWeave has deployed Spectrum-X Multiplane in production to connect Vera Rubin racks using parallel switches, while Elon Musk's SpaceXAI will adopt Nvidia's Vera CPUs on a large scale for agentic AI systems in its data centers and satellites.

Over the past year, Samsung Foundry, which once faced massive losses, has secured a wave of new clients. It all started with a $16.5 billion deal to make Tesla's AI6 chip on its 2nm process node, followed by additional production for Tesla's AI5 chips.

Samsung is also reportedly in talks to make AI chips for Anthropic, BYD, ByteDance (TikTok), Meta, and Neuralink. Reports indicate that Samsung Foundry could swing back to full-year profit in 2027. So, it appears that the South Korean firm has clawed back to a good position to challenge its primary rival TSMC in the future.