Samsung’s foundry unit has cleared a milestone that Nvidia hopes changes the economics of AI inference: the Groq 3 LPX accelerator it builds on 4nm has moved into full production.
The August 24 announcement at Hot Chips marks the payoff of a relationship Nvidia CEO Jensen Huang made public at GTC in March, when he named Samsung as the manufacturer for the chip Nvidia acquired through its $20B purchase of Groq.
Groq 3 LPX is built for the token-heavy end of AI, where agents chain hundreds of inference steps. Nvidia says it reached 3,400 output tokens per second in Artificial Analysis benchmarks, roughly 4x faster response for latency-sensitive agents, and compresses coding work that spanned hours into minutes.
The accelerator slots beside the Vera Rubin NVL72 platform, Nvidia’s training and general-purpose workhorse, with the LPX handling the interactive layer of AI factories.
Huang framed inference as the growth engine of AI, and the ramp gives Samsung foundry a marquee win as it fights for AI-era customers against TSMC.
Racks built around the chip should come online this year, with Nebius among the first to deploy them, according to reports.
