Guotai Junan: Deepseek R1 small-scale distillation model accelerates landing of AI on the edge, with potential benefits for edge hardware.

Guotai Junan learned from GMTEight APP that Guotai Junan's research report stated that DeepSeek R1 reduces the performance requirements of computing chips based on algorithm optimization, speeds up the adaptation of domestic computing power, and with overseas computing power restrictions increasing, independent and controllable computing power is the inevitable choice. Recommendations include Semiconductor Manufacturing International Corporation (00981, 688981.SH), Sichuan Huafeng Technology (688629.SH), Beneficiaries include Bichamp Cutting Technology (002843.SZ), and Guangdong Shenling Environmental Systems (301018.SZ); the tilt of AI infrastructure towards inference computing, and ASIC chips have higher cost-effectiveness in inference scenarios. Deepseek R1 has strong performance in distilling small models, accelerating the landing of AI on the edge, and hardware on the edge is expected to benefit. Guotai Junan's main points are as follows: Deepseek R1 is open-source and low-cost, with advanced processes still being the core bottleneck, and domestic computing power accelerating growth. DeepSeek R1 has outstanding cost-effectiveness, performance comparable to o1 models, accelerating adaptation for domestic computing chips represented by Ascend, it is expected to lead the domestic AI ecosystem cycle and the domestic computing power industry chain will benefit fully. High-performance computing chips require the use of advanced processes at 7nm and below, creating development opportunities for domestic advanced processes. Inference demand drives the expansion of computing power, ASICs undertake more inference work. Deepseek R1's high inference performance drives the prosperity of inference chip clusters and AI applications, the long-term investment in AI infrastructure is tilting towards inference computing, and efficient and low-cost ASIC chips are expected to benefit. R1 has strong performance in distilling small models, accelerating the landing of AI on the edge. Deepseek R1 has excellent performance in distilling small models, driving a qualitative change in hardware experience with the upgrade of AI intelligence on the edge, and opening up growth space for hardware on the edge. Risk warning: AI technology iteration is slower than expected, and the progress of the landing of edge models may be slower than expected.