GF Securities analysts attended the OCP Global Summit held in San Jose last week, where the event reflected a shift by hyperscale cloud providers toward full-stack co-design.

Analysts stated that the current pullback in semiconductor stocks has created buying opportunities for quality names such as NVIDIA (NVDA) and Marvell Technology (MRVL).

Analyst Jeff Pu said in an investor report: "Nvidia's networking roadmap now covers both co-packaged optics (CPO) and a new USFP pluggable standard, the latter developed jointly with Cisco (CSCO), aimed at improving density and maintainability within AI factories. USFP modules deliver up to 12.8 Tb/s per module and 400 Gb/s per lane, achieving efficiency close to CPO while retaining the field-maintenance flexibility of pluggable solutions."

USFP stands for Ultra Small Form-factor Pluggable, a new optical transceiver standard used to connect large-scale AI data center clusters.

Pu noted: "Our research on US neoclouds shows that power availability and permitting remain decisive factors for expansion. However, there is no power shortage currently or in the foreseeable future, because most sites are far from urban centers and many use stable gas power generation. On power costs, regional price spreads remain large, driving new projects toward the southern corridor and select overseas regions such as Italy. On shells, modular data center shells can be completed in as little as three months, but liquid cooling infrastructure remains the bottleneck."

The research also showed that CPUs are becoming more important in inference, especially for agentic AI, which is expected to double the total addressable market for CPUs.

Pu said: "GPU utilization in decode workloads averages only 20% to 30%, and SRAM-based accelerators perform poorly in long-context efficiency. Token costs are falling 5 to 10 times per year, but semiconductor and memory costs are rising. The bottleneck lies in NAND, DRAM and thermal management. Under the agentic AI factory model, orchestration and retrieval run on CPUs, reshaping unit economics: the CPU-to-accelerator ratio shifts from 1:4 to 1:8 in the chatbot era to 1:1 in agentic deployments."

He added: "Despite overall market weakness, we recommend investors buy quality stocks on dips. We remain bullish on Nvidia, the PTFE sector, Marvell (DSP, Google companion, Maia 300) and the CPU supply chain (AMD, Intel)."

However, risks include slowing AI demand, infrastructure deployment, supply constraints and competition.