NVIDIA's Groq 3 LPX enters full production as Nebius becomes first cloud to deploy

NVIDIA has moved its Groq 3 LPX interactive AI inference accelerator into full production, with racks going operational later this year, and Nebius became the first AI cloud to deploy it. Artificial Analysis measured roughly 3,431 tokens/s running Gemma 4 31B, a new output-speed record. Positioned as an extension of the Vera Rubin NVL72 system for long-context, low-latency agentic workloads, LPX is the first fruit of last year's roughly $20 billion Groq licensing-and-talent deal and could reshape the ultra-low-latency inference market.

Confirmed

• Groq 3 LPX is in full production as an extension of the Vera Rubin NVL72 platform, delivering ultra-fast token generation for latency-sensitive agentic workloads with a 100K context window (@zephyrz9, @ctnzr, @heypearlai, @badumtsssst).
• On an early private deployment granted by NVIDIA, Artificial Analysis measured about 3,431 tokens/s running Gemma 4 31B at both 10K and 100K input lengths, a record well ahead of existing public endpoints (@ArtificialAnlys); @ctnzr also reports the platform makes Vera Rubin inference 4x faster.
• NVIDIA's LPX team blog post adds that Gemma 4 reaches 10,996 OTSU in high-quality settings and a median 3,431 OTSU at 100K input context, with additional results on 2T-parameter-class models (@ricklamers, @AccBalanced).
• Nebius announced it is the first to bring Groq 3 LPX to its own AI cloud, integrating ultra-low-latency inference with production runs of leading open models; it cites up to 30x agent throughput per megawatt on Rubin systems (@demianai, @BenBajarin, @IanAndrewsDC).
• Groq also announced the LPX system, designed to work with NVIDIA's next-generation GPUs to break the speed-versus-cost tradeoff, arguing inference rather than training is the key bottleneck and that it is building hundreds of megawatts of inference capacity (@goyalshaliniuk).
• Ian Andrews notes this is the first outcome of last year's Groq–NVIDIA licensing agreement; @eyishazyer adds the deal was worth about $20 billion and LPX reached full production just 8 months later (@IanAndrewsDC).
• Per WSJ (relayed by @rohanpaulai), agentic AI poses two compute challenges—huge context handling and extremely low-latency token generation—with single agent tasks requiring hundreds of sequential inference steps where millisecond latencies accumulate; Groq 3 LPX targets exactly this.

Unconfirmed

• @eyishazyer mentions SpaceXAI deploying NVIDIA's new Vera 'agent-specific CPU'; this appears in only one post and lacks corroboration.

Why it matters

• Nebius frames it in 'AI factory' terms: power is the scarce resource, and the metric is inference throughput per unit of energy rather than peak FLOPS (@demianai).
• Combining ultra-low-latency inference with production open-model serving, plus Groq's push for low-cost inference at scale, could shift competitive dynamics among AI clouds in the agent inference market.
• Going from a mega licensing deal to full chip production in roughly 8 months also signals NVIDIA's accelerating system-level iteration cadence.

2026-08-24 ~ 2026-08-26 · 18 related posts

• [source] NVIDIA's Groq 3 LPX enters full production, hits 3,400 tokens/sec in tests — zephyr_z9 · 2026-08-24
• Nebius Adopts NVIDIA Groq 3 LPX for Fastest Inference — BenBajarin · 2026-08-25
• [source] NVIDIA Confirms Groq LPX Production; Nebius First to Deploy in Cloud — IanAndrewsDC · 2026-08-25
• NVIDIA blog: Gemma 4 hits 10,996 OTSU with Vera Rubin optimizations — ricklamers · 2026-08-25
• NVIDIA Groq 3 LPX Hits 3,431 tok/s on Gemma 4 Long Context Benchmark — ricklamers · 2026-08-25
• Nebius brings NVIDIA Groq 3 LPX online; Rubin NVL72 hits up to 30x throughput per MW on agentic coding — demian_ai · 2026-08-25
• NVIDIA Groq 3 LPX Enters Production, Delivering 4x Faster Inference on Vera Rubin — ctnzr · 2026-08-25
• Groq Launches LPX for Fast, Affordable Inference Alongside NVIDIA GPUs — goyalshaliniuk · 2026-08-25
• Groq 3 LPX Hits Full Production; Nvidia's Agent-Orchestration Vera CPU Lands — eyishazyer · 2026-08-25
• NVIDIA Groq 3 LPX targets low latency for long Agentic AI workflows — rohanpaul_ai · 2026-08-26
• NVIDIA Groq 3 LPX Enters Production for Agentic AI Speed — badumtsssst · 2026-08-26

7 near-duplicate retellings: nvidia · heypearlai · haider1 · AccBalanced · ArtificialAnlys · IanAndrewsDC · tomjaguarpaw