Nvidia has announced the start of volume production for its Grok 3 LPX AI inference accelerator, manufactured by Samsung Electronics on its 4nm process with yields exceeding 80%. The platform packages 256 LPUs into a rack for up to four times faster inference in agentic AI workloads. Nebius is the first customer for its Nebius Token Factory deployment at year-end. The collaboration now includes HBM4 supply for Nvidia's Rubin GPU and other components like SOCAMM2. OpenAI's Habanero chip is also using Samsung's HBM4. Samsung has raised prices by 10-15% and the 4nm line is at near 100% utilization, with analysts expecting the foundry to return to profitability in the third quarter.