Nvidia used its GTC 2026 conference in San Jose to debut the Groq 3 Language Processing Unit, the first silicon to emerge from its roughly $20 billion acquisition of inference startup Groq completed in late 2025. The SRAM-based chip, manufactured on Samsung's 4nm process, is designed specifically for low-latency token generation and slots into Nvidia's Vera Rubin platform as a dedicated decode-phase co-processor.
The Groq 3 LPX rack houses 256 LPUs across 32 compute trays connected via direct chip-to-chip copper spine. Paired with the NVL72 GPU rack, Nvidia claims the system delivers up to 35 times more tokens and 10 times more revenue opportunity for trillion-parameter models compared to Blackwell. Around 90% of Groq's staff have transitioned to Nvidia, and shipping is planned for the second half of 2026.
CEO Jensen Huang compared the deal to Nvidia's earlier Mellanox acquisition, which transformed its networking capabilities. The move signals a shift from a GPU-only company to a full-stack AI platform optimized for both training and inference, directly challenging specialized inference competitors like Cerebras.
