AnswersPublished22 sources
Nvidia’s Groq 3 LPX Enters Full Production for Faster AI Inference
Nvidia announced on August 24 that Groq 3 LPX had entered full production. In Artificial Analysis testing, it generated 3,400 output tokens per second on Gemma 4 31B with a 100,000 token context, while Nvidia claims a...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- Nvidia announced on August 24 that Groq 3 LPX had entered full production. In Artificial Analysis testing, it generated 3,400 output tokens per second on Gemma 4 31B with a 100,000 token context, while Nvidia claims a...
- LPX is not a replacement for Nvidia GPUs: it is a dedicated accelerator for the decode phase of inference, where models generate responses token by token.
- Nebius is the first announced AI cloud adopter, with plans to integrate LPX into its Token Factory platform; Groq also plans to deploy it alongside Vera Rubin NVL72 systems.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Nvidia announce at Hot Chips 2026 about the full-production launch of its Groq 3 LPX dedicated AI-inference accelerator—acquired th. Article summary: At Hot Chips on August 24, Nvidia said its Groq 3 LPX rack-scale accelerator had entered full production as the low-latency, long-context inference component of the Vera Rubin platform. It is intended for rapid token gen. Topic tags: general, general web, user generated, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks,
Nvidia announced at Hot Chips on August 24 that its Groq 3 LPX rack-scale inference accelerator had entered full production. The system is designed to work alongside Vera Rubin NVL72, accelerating the token-generation stage that determines how quickly interactive AI systems respond. 26
In an Artificial Analysis benchmark, Groq 3 LPX produced 3,400 output tokens per second while running Google’s open-source Gemma 4 31B model with a 100,000-token context. Nvidia says that result was four times faster than the nearest alternative platform for the tested workload. 17
35
What Groq 3 LPX does
Groq 3 LPX targets low-latency inference for agentic AI, coding assistants and other applications that must respond quickly while processing long context. Nvidia describes it as an interactive inference accelerator for predictable, high-speed token generation. 19
26
The key distinction is between prefill and decode. Prefill processes the user’s prompt and its surrounding context. Decode generates the answer one token at a time. LPX is aimed primarily at making that second stage faster, while Vera Rubin NVL72 remains the more flexible system for training, broader inference and high-throughput context processing. 34
That division lets Nvidia present LPX as a complement to its GPU infrastructure rather than a wholesale replacement for it. GPUs continue to provide general-purpose accelerated computing, while the Groq-derived architecture is assigned to the latency-sensitive part of serving an AI model.
The benchmark—and its limits
The headline result was a median of 3,431 output tokens per second in Artificial Analysis’s 100K-context benchmark on Gemma 4 31B; Nvidia rounded the figure to 3,400 tokens per second in its announcement. 17
Nvidia’s comparison is a workload-specific claim, not a universal measure of AI performance. It applies to the tested model, context length and benchmark setup. The company characterized LPX as four times more responsive than the nearest alternative platform in that scenario. 35
That caveat matters because token-generation speed is only one part of an inference system’s performance. Model size, prompt length, concurrency, memory, networking and the time required to process the initial context can all affect the experience of a real application.
Where LPX will be deployed
Nebius is the first announced AI-cloud adopter. It plans to integrate Groq 3 LPX into Nebius Token Factory, its production-inference platform, before the end of 2026. 2
Groq also says it will be among the first adopters, deploying LPX alongside Vera Rubin NVL72 in its purpose-built AI-inference cloud. The intended combination is straightforward: Rubin systems provide the broader platform capabilities, while LPX increases token-generation speed for context-heavy workloads. 18
How the launch fits into Vera Rubin
Nvidia has positioned Groq 3 LPX as a seventh major component of the Vera Rubin platform. The broader platform also includes the Vera CPU, which Nvidia says has 88 custom-designed Olympus cores. 33
The Vera CPU addresses the work surrounding model inference, including orchestration, data processing, tool use and code execution. Nvidia’s announcement also said SpaceXAI plans to deploy Vera CPUs for next-generation agentic AI applications and use an optimized Vera Rubin NVL72 system for a first-generation Starmind AI satellite. 40
Together, those announcements illustrate Nvidia’s broader strategy: build AI factories from specialized components rather than treating every workload as a GPU problem. In that model, CPUs coordinate agent workloads, GPUs handle flexible large-scale computation, and LPX focuses on fast, deterministic response generation.
What the $20 billion Groq deal actually involved
The launch is the clearest product outcome so far from Nvidia’s December 2025 agreement with Groq, but calling it a straightforward acquisition of Groq is misleading. Reporting describes a non-exclusive license for Groq’s inference technology combined with Nvidia’s hiring of Groq founder and CEO Jonathan Ross, President Sunny Madra and other key employees. 4
8
Groq remained active after the transaction and later raised additional funding, which is consistent with a technology-and-talent deal rather than the purchase of the entire operating company. 3
That structure also helps explain the product name and positioning. Nvidia has incorporated Groq’s language-processing technology into its own rack-scale platform, while Groq continues operating as an independent company and plans to use the resulting Nvidia hardware in its cloud.
The strategic takeaway
The important announcement is not simply that Nvidia has produced another AI chip. It is that Nvidia is separating the stages of AI service delivery and assigning the most latency-sensitive one to a dedicated accelerator.
For interactive agents, faster decode can make multi-step tasks feel more responsive. Nvidia’s benchmark suggests substantial performance in one long-context configuration, but real-world gains will depend on the application and the rest of the serving stack. The full-production milestone, Nebius deployment and Groq’s planned adoption will provide the more meaningful test of whether LPX can turn that benchmark advantage into broadly available cloud performance.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Nvidia’s Groq 3 LPX Enters Full Production for Faster AI Inference"?
Nvidia announced on August 24 that Groq 3 LPX had entered full production. In Artificial Analysis testing, it generated 3,400 output tokens per second on Gemma 4 31B with a 100,000 token context, while Nvidia claims a...
What are the key points to validate first?
Nvidia announced on August 24 that Groq 3 LPX had entered full production. In Artificial Analysis testing, it generated 3,400 output tokens per second on Gemma 4 31B with a 100,000 token context, while Nvidia claims a... LPX is not a replacement for Nvidia GPUs: it is a dedicated accelerator for the decode phase of inference, where models generate responses token by token.
What should I do next in practice?
Nebius is the first announced AI cloud adopter, with plans to integrate LPX into its Token Factory platform; Groq also plans to deploy it alongside Vera Rubin NVL72 systems.