Groq 3 LPX has entered full production, targeting high-speed inference for Agentic AI applications.

NVIDIA has advanced its Groq 3 LPX inference accelerator into full production, engineered to meet the intensive computational demands of agentic AI. Designed as an expansion of the Vera Rubin platform, the specialized silicon achieves 3,400 output tokens per second in benchmark testing. Nebius has been secured as the first cloud provider to deploy the architecture.

Agentic systems rely on thousands of rapid reasoning steps, making raw token generation throughput critical for real-time operations. In benchmarks conducted by Artificial Analysis using the open-source Gemma 4 31B model with a 100,000-token context window, the Groq 3 LPX established a new performance ceiling. NVIDIA claims the accelerator provides 4x faster responsiveness compared to the closest competing systems, cutting long coding tasks down from hours to minutes.

The chip functions as an interactive complement to Vera Rubin NVL72 rack-scale systems. While conventional architectures handle general-purpose processing, the LPX is optimized exclusively for generation throughput, enabling software agents to verify code and execute tools without delay. NVIDIA CEO Jensen Huang addressed this shift in workload distribution during his presentation at Hot Chips:

Nebius is integrating the hardware directly into its Token Factory platform, allowing engineers to access accelerated performance through existing software interfaces without restructuring their code. Purpose-built provider Groq will similarly rank among the earliest operators to integrate the units into its server fleet.

The broader hardware rollout spans 7 distinct chips across 5 dedicated rack configurations. These architectures incorporate BlueField 4 data processing units, Vera CPU modules, Vera BlueField 4 STX storage arrays, and Spectrum 6 SPX Ethernet networking to maintain low latency across multi-agent environments.