Nvidia Starts Groq 3 LPX Production With Samsung-Made Chips
Competition in generative artificial intelligence is shifting from model training toward faster, more economical inference. That transition is particularly important for AI agents, which must execute multi-step tasks while producing responses with minimal delay. Nvidia’s Groq 3 LPX system uses Language Processing Units, or LPUs, optimized for rapid token generation, aiming to improve the speed and efficiency of running large language models.
Nvidia said the Groq 3 LPX rack has entered mass production, with all of its chips manufactured by Samsung Electronics using a 4-nanometer process. Each system contains 256 LPUs and is designed for high-speed AI inference workloads. AI cloud provider Nebius will become the first customer to deploy the racks on its inference platform. Nvidia did not disclose the contract value or provide a specific delivery date.
All Coverage
1 original reportsThe Backstory
The history behind this eventNVIDIA Unveils New Groq-Powered Inference Chip to Be Manufactured by Samsung
The generative AI market is shifting from model training toward inference at scale, making the ability to generate tokens with low latency and high throughput critical for cloud providers seeking to control costs. By integrating Groq's LPU architecture, NVIDIA aims to strengthen its inference lineup beyond GPUs and compete with custom ASICs developed by cloud service providers.
Jensen Huang unveiled the NVIDIA Groq 3 LPU at GTC 2026, the company's first inference chip to integrate Groq technology. Samsung will manufacture the chip, with shipments expected in the third quarter. Reports valued the technology licensing deal at $20 billion. Huang also predicted that large language models would adopt tiered pricing based on token quality, with AI factories dynamically allocating training and inference workloads.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.
If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →