After acquiring Groq's assets for $20 billion, NVIDIA announced that the Groq 3 LPX rack has entered full-scale mass production and official commercialization. Each rack integrates 256 Groq 3 chips, capable of outputting 3,400 Tokens per second. It will be deployed alongside Vera CPUs and Rubin GPUs in Nebius data centers, with an expected launch later this year. The Groq chip features 500MB of high-speed SRAM, effectively alleviating memory bottlenecks. Manufactured by Samsung, it focuses on low-latency inference, specifically designed for the decoding phase of model serving, enabling cloud service providers to launch high-end value-added services. While low-latency chips will not replace GPUs as the mainstay in AI, NVIDIA is still expanding shipments of the Vera Rubin system. NVIDIA CEO Jensen Huang predicts that cumulative orders for Blackwell chips and Vera Rubin systems will reach $1 trillion by 2027. The company plans to allocate a quarter of its data center rack space for code development businesses to Groq chips. Additionally, NVIDIA will release its earnings report this Wednesday.