Nvidia Says Groq 3 LPX Racks Enter Full Production, First Deployment Planned for Nebius

Nvidia senior director Dion Harris said Monday local time that Groq 3 LPX racks have entered full production and will be deployed in the data center of Nebius, a new cloud provider, with the system expected to go live later this year. The report says the rollout follows Nvidia’s roughly $20 billion technology licensing deal with Groq in December, and that the technology obtained in that transaction is now moving into commercial use. It also says several core Groq employees have joined Nvidia, while Groq founder Jonathan Ross is now Nvidia’s chief software architect. Groq 3 LPX was released in March as an inference accelerator for the Vera Rubin platform. According to the report, the architecture integrates 500 megabytes of high-speed SRAM directly on the chip die to ease memory bottlenecks. Each LPX rack can combine 256 Groq 3 chips, and Nvidia cited benchmarks showing throughput of 3,400 tokens per second. The chip is manufactured by Samsung. Harris said low-latency chips are not intended to replace GPUs. Instead, they are meant to match different processors to different parts of a workload. Cloud providers, he added, can charge more for users who are sensitive to latency. Nvidia CEO Jensen Huang plans to devote about one quarter of data-center space for programming applications to Groq chips, with the rest using Vera Rubin systems. Nvidia also expects combined sales of Blackwell and Vera Rubin to reach $1 trillion by 2027.