Nvidia puts Groq 3 LPX into full production

Nvidia put Groq 3 LPX into full production on August 24, claiming 3,400 output tokens per second, according to Nvidia's announcement at Hot Chips.

LPX targets the token-generation phase, the step that decides how responsive an agent feels. That phase is what stalls while a loop waits on a multi-step tool call. Nvidia calls it an interactive inference accelerator and says it runs 4x faster on agent responsiveness than the nearest alternative.

The 3,400-tokens-per-second figure comes from running Gemma 4 31B at a 100,000-token context. LPX extends the Vera Rubin NVL72 platform.

Nebius is the first cloud provider to put it into production. Nvidia hasn't disclosed pricing, so what that speed costs at scale is still unknown.

That figure is Nvidia's own benchmark, run on Nvidia's own comparison set. It stays a marketing number until an independent test reproduces the 4x agent-responsiveness claim against a workload nobody at Nvidia picked. Watch for Nebius or a third party to publish that number once LPX is live outside Nvidia's demos.