Nvidia has shifted its Groq 3 LPX inference accelerator for AI agents into mass production. The chip extends Nvidia\'s Vera Rubin data center platform and targets high-speed token generation to reduce decode latency in agent workflows. In an Artificial Analysis benchmark, it produced 3,400 tokens per second running the open-source Gemma 4 31B model with a 100,000-token context window. Nebius Group is among early adopters, and SpaceX was named a new flagship customer.

Nvidia unveiled new AI inference accelerators and network infrastructure products at Hot Chips 2026 in the United States on Aug. 25, a day before its second-quarter earnings. The company said interactive inference accelerator Nvidia Grok 3 LPX has entered mass production, alongside Spectrum-X Multiplane Ethernet scaling and ScaleIn infrastructure acceleration technology. The announcements focused on inference, particularly token generation speed, and also highlighted inference speed, network scalability and agent-focused CPUs within the Vera Rubin platform.