Nvidia begins production of Groq AI racks after $20B purchase

Nvidia announced Monday that its Groq 3 LPX artificial intelligence racks have entered full production, marking the commercialization of technology acquired through the chipmaker's record $20 billion purchase of Groq assets in December. The specialized systems will be deployed alongside Nvidia's Vera central processing units and Rubin graphics processing units at cloud infrastructure provider Nebius and are expected to become operational later this year.

Each rack packages 256 Groq 3 chips and can generate approximately 3,400 tokens per second, according to a benchmark cited by the company. The Groq chips are manufactured by Samsung Electronics, while Taiwan Semiconductor Manufacturing Company produces Nvidia's graphics processors.

Technical specifications

The Groq 3 LPX is designed specifically for low-latency inference — the process through which trained AI models generate responses. The architecture places 500 megabytes of high-speed static random-access memory directly on each chip to eliminate memory-related bottlenecks, addressing delays that can degrade user experience in AI agents and coding assistants.

Nvidia stated that these systems are intended to complement rather than replace GPUs. While graphics processors can perform both AI model training and inference, Groq chips primarily target the latency-sensitive "decode" phase of running models.

Strategic positioning

Nvidia CEO Jensen Huang said in March that the company expects cumulative sales from its Blackwell and Vera Rubin platforms to reach $1 trillion through 2027. He also noted that a quarter of the data center capacity allocated to coding applications would use Groq chips.

Competition in specialized inference hardware has intensified as technology companies seek to make AI services faster and more economical. Nvidia rival Advanced Micro Devices has announced plans to integrate rack-scale systems with chips produced by Cerebras.