广发证券Jeff Pu在OCP峰会后发布AI硬件策略纪要：建议逢低吸纳英伟达、Marvell、AMD、英特尔及PTFE板块，无个股新增订单或业绩事实

The current downturn among semiconductor stocks has created a buying opportunity for quality names, such as Nvidia ( NVDA ) and Marvell ( MRVL ), according to GF Securities. GF analysts attended the OCP Global Summit in San Jose last week, which reflected a shift toward full-stack co-design across hyperscalers. "Nvidia's networking roadmap now spans both co ‑ packaged optics and the new USFP pluggable standard, co ‑ developed with Cisco ( CSCO ) to raise density and serviceability inside AI factories," said GF analyst Jeff Pu in an investor report. "USFP modules deliver up to 12.8 Tb/s per module and 400 Gb/s per lane, achieving near ‑ CPO efficiency while retaining the flexibility of pluggables for field maintenance." The USFP stands for ultra-dense small form factor pluggable and is a new optical transceiver standard used to connect massive AI data center clusters. "Our U.S. tour to neoclouds suggests power availability and licensing remain decisive factors for expansion," Pu noted. "However, there is no power shortage issue by now and in the foreseeable future, as most sites located far from urban centers, and many use stable gas ‑ based generation. For electricity cost, regional cost spreads remain wide, driving migration of new builds toward the southern corridor and select overseas zones such as Italy. For shell, the modular datacenter shells can be completed in as little as three months, but liquid ‑ cooling infrastructure continues to be the bottleneck." The tour also revealed the rising importance of CPU for inference, especially for agentic AI. This is expected to double the total addressable market for CPUs. "GPU utilization for decode workloads averages only 20–30%, and SRAM ‑ based accelerators struggle with long ‑ context efficiency," Pu said. "Token costs are falling 5–10× every year, but semi and memory costs are rising. Bottlenecks are NAND, DRAM, and cooling. The agentic AI factory model—where orchestration and retrieval run on CPUs—reshapes unit economics: CPU ‑ to ‑ accelerator ratios move from 1:4–1:8 in chat ‑ era systems to 1:1 in agentic deployments." "Despite the broad market softness, we suggest investors accumulate quality stocks," he added. "We continue to like Nvidia, PTFE sector, MRVL (DSP, Google attach, Maia 300), and the CPU supply chain ( AMD )( INTC )." However, risks include AI demand slowdown, infrastructure deployment, supply constraints and competition.