NVIDIA’s Groq 3 LPX has entered full production. Last December, NVIDIA shelled out roughly $20 billion for a Groq technology license and poached founder Jonathan Ross along with key members of his team. Eight months later, the first piece of hardware from that deal is officially here.
Groq 3 LPX is a system purpose-built to accelerate large-model inference, packing 256 Groq 3 LPU chips. In Artificial Analysis testing, with Gemma 4 31B processing 100,000 input tokens, Groq 3 LPX hit 3,431 output tokens per second. In the same test, the fastest publicly callable API at the time managed only about 870 — making Groq 3 LPX nearly 4x faster.
In real deployments, the Rubin GPU handles the heavier compute while Groq 3 LPX focuses on speeding up token generation. For coding agents that hammer models dozens or hundreds of times in a row, that can seriously cut down on waiting for the model to 'spit out' its response.
Nebius will be the first AI cloud provider to adopt Groq 3 LPX. But here’s the twist: Groq itself is also among the first adopters, and it’s deploying the system alongside Dell. Last year Groq licensed its tech to NVIDIA; this year, it’s using the 'Groq' NVIDIA built.
NVIDIA has more details on its developer blog.