What did Nvidia announce at Hot Chips 2026 about the full production launch of its Groq 3 LPX dedicated AI inference accelerator—acquired through its $20 billion purchase of Groq, which brought founder Jonathan Ross and President Sunny Madra to Nvidia—including when and where the

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 generation in agentic AI, complementing Vera Rubin NVL72 systems rather than replacing GPU-based training and broader inference work. 36

What Nvidia announced

• Nvidia reported 3,400 output tokens per second on Google’s open-source Gemma 4 31B with a 100,000-token context in an Artificial Analysis benchmark; Nvidia’s more precise cited result is 3,431 tokens per second. 16

• Nvidia claimed this made LPX four times faster in responsiveness than the nearest alternative platform for these long-context, latency-sensitive workloads. 6

• The product’s core purpose is the decode/generation portion of inference—producing the next token quickly and predictably—while Vera Rubin systems handle the broader training, context processing, and AI-factory workload. 36

• Nvidia positioned LPX as a seventh major Vera Rubin component, alongside the Vera CPU and the platform’s other chips; the Vera CPU has 88 custom Olympus cores. 2

• Nebius is announced as the first AI-cloud adopter and plans to add Groq 3 LPX to its Token Factory production-inference platform. 6

• Groq itself also said it would be among the first adopters, deploying LPX alongside Vera Rubin NVL72 in its purpose-built AI-inference cloud. 4

Strategic meaning

Nvidia’s framing is that future AI factories should be workload-optimized: GPUs and Rubin systems for flexible, large-scale AI compute, with LPX dedicated to the interactive token-generation bottleneck that determines an agent’s perceived responsiveness. 36

Important corrections and limits

• The premise that Nvidia “purchased Groq” is not fully accurate. Available reporting describes a $20 billion deal to license Groq technology and hire key personnel, rather than an outright acquisition of the company; Groq subsequently remained active and raised new funding. 5

• The supplied evidence does not adequately substantiate the specific hardware claims of 256 chips per rack, 500 MB of on-die SRAM per chip, Samsung fabrication versus TSMC GPU production, nor the claimed competitive responses from AMD/Cerebras or OpenAI’s 750-token-per-second mode. Insufficient evidence.

• Likewise, the evidence supports that SpaceXAI plans to deploy Vera CPUs for agentic-AI applications, but does not establish the detailed claims about tool execution, simulation across orbital satellites, or the exact scope and timing of that deployment. 7