According to a research report released by CITIC Securities, the upcoming NVIDIA GTC conference is poised to spotlight LPU (presumably a specialized chip or technology). The separation of preprocessing (P) and decoding (D) has emerged as a standard approach for large-scale model inference. Previously, NVIDIA unveiled the Rubin CPX to address the need for cost reduction in the prefill stage. Following its acquisition of Groq, there is speculation that NVIDIA may introduce LPU or similar chips at this conference, aiming to boost decoding efficiency. In light of these developments, it is advisable to closely monitor key enterprises within NVIDIA's supply chain, as well as those investing in LPU technology. Meanwhile, in the domestic computing power sector, the integration of PD separation with super-node solutions is anticipated to become the mainstream trend. Therefore, it is prudent to keep an eye on leading domestic companies specializing in super-nodes and computing power chips.