GF Securities: Post-OCP Summit Semiconductor Pullback Is a Buying Opportunity; Bullish on NVIDIA, Marvell, and CPU Chain
nashnova research
GF Securities analyst Jeff Pu attended the OCP Global Summit and argues that hyperscalers are shifting to full-stack co-design; the current semiconductor pullback creates a buying window for Nvidia, Marvell, and the CPU supply chain, with agentic AI set to double CPU's addressable market.
Why does Pu call this pullback a buy?
Hyperscale cloud operators are moving from buying discrete chips to full-stack co-design — rethinking silicon, networking, and thermal together.
This means → the demand structure hasn't weakened; sentiment dropped ahead of fundamentals, opening a price gap on quality names.
Pu explicitly recommends buying the dip in Nvidia (NVDA) and Marvell (MRVL), and stays bullish on the CPU chain (AMD, Intel) and the PTFE segment.
Where does Nvidia's networking roadmap stand?
Nvidia is advancing two tracks simultaneously: co-packaged optics (CPO) — mounting optical modules right next to the chip to shorten signal paths — and a new USFP pluggable standard.
USFP, co-developed with Cisco, reaches up to 12.8 Tb/s per module and 400 Gb/s per lane.
In plain terms = USFP already approaches CPO-level efficiency, but modules can still be swapped in the field — far more maintenance-friendly than CPO.
Power and liquid cooling — where is expansion actually stuck?
Field research shows power permitting remains the decisive constraint on data-center expansion, though no actual shortage exists — most sites sit far from cities and run on stable gas-fired generation.
Regional power-price gaps are wide, pushing new builds toward the U.S. southern corridor and overseas sites like Italy.
Liquid-cooling infrastructure is still a deployment bottleneck, while modular data-center shells can be completed in as little as three months. This reflects a simple reality: building the shell is easy; plumbing the cooling is hard.
GPU utilization sits at 20–30% — is the money well spent?
GPUs average only 20%–30% utilization on decode workloads; SRAM-based accelerators perform poorly on long-context tasks.
Token costs are falling 5–10× per year, yet semiconductor and memory costs are rising — bottlenecks concentrate in NAND, DRAM, and thermal management.
This means → more GPUs alone won't solve the problem — the cost structure is shifting from compute itself to storage and cooling.
Why could agentic AI double the CPU market?
Agentic AI — a new paradigm where AI autonomously decomposes tasks, calls tools, and executes multi-step workflows — is sharply raising the CPU's importance in inference.
In agentic AI factory deployments, orchestration and retrieval run on CPUs; the CPU-to-accelerator ratio shifts from 1:4–1:8 in the chatbot era to 1:1.
In plain terms = one CPU used to support four to eight GPUs; now every GPU needs its own CPU — CPU demand multiplies several times over.
What are the risks, and what is the key validation point?
Pu flags four downside risks: slowing AI demand, infrastructure deployment pace, supply constraints, and intensifying competition.
Whether the CPU-to-GPU ratio actually shifts to 1:1 as expected is the critical node for validating the entire investment thesis.
This means → if agentic AI adoption lags expectations, the CPU chain's "doubling story" takes a haircut — this ratio is the single most important number to track from the entire report.
市场有风险，内容仅供研究参考，不构成投资建议。