NVIDIA Puts Groq 3 LPX Into Full Production for Agentic AI Inference

What NVIDIA announced

On August 24, 2026, NVIDIA announced that Groq 3 LPX is in full production. The accelerator extends the NVIDIA Vera Rubin platform and is designed specifically for the token-generation side of latency-sensitive agentic inference. NVIDIA says Nebius is the first AI cloud provider adopting the system.

Why LPX exists

Agentic workloads have two different compute pressures: processing large contexts and generating each next token quickly enough for an agent to remain responsive while it plans, calls tools, reads results and iterates. NVIDIA positions Groq 3 LPX as a specialized complement to Vera Rubin NVL72 for the second problem, with the platform able to split or co-execute inference work across Rubin GPUs and LPX accelerators.

Performance claim and architecture

NVIDIA reports that Artificial Analysis measured 3,431 output tokens per second on a 100K-context benchmark using Gemma 4 31B. NVIDIA's technical description attributes the low latency to deterministic compiler scheduling, fine-grained overlap of communication and compute, SRAM-based execution and tightly planned chip-to-chip transfers. The company also describes configurations for prefill/decode disaggregation, attention/FFN disaggregation and speculative external-drafter decoding.

Why it matters

The launch reflects a broader shift in AI infrastructure from training-only optimization toward inference systems tuned for interactive agents. If specialized decode accelerators can keep long-running agents responsive while GPUs handle large-context and high-throughput work, AI clouds may increasingly deploy heterogeneous inference stacks instead of relying on one accelerator type for every stage.

Availability

NVIDIA describes Groq 3 LPX as in full production, not a future preview. Actual service availability, regional rollout and cloud capacity still depend on NVIDIA systems and adoption by infrastructure providers such as Nebius.