NVIDIA's Post-GTC 2026 Inference Stack: Why LPX and CMX Moved Ahead of CPX

Post-GTC 2026, NVIDIA's inference strategy has shifted away from GPU-only scaling toward a heterogeneous AI factory that combines Vera Rubin GPU/CPU, Groq 3 LPX/LPU, BlueField-4 STX·CMX KV-cache storage, and Spectrum-X/SPX networking. LPX does not replace HBM — it complements the low-latency decode weakness of Rubin GPU/HBM. Samsung Electronics can be reclassified as an inference memory hierarchy supplier spanning HBM4, SOCAMM2, Groq LPU foundry, and eSSD/KV-cache.

Recommendations 0

Marvell Q1 FY2027 and Korean Semiconductors: The Bottleneck Is Interconnect, Substrate, and Power — Not Just HBM

Marvell Q1 FY2027 was not a simple EPS beat. It confirmed that AI data center bottlenecks are spreading across custom XPUs, optical interconnects, scale-up networking, FC-BGA substrates, MLCCs, silicon capacitors, and test sockets. This post maps the Korea semiconductor read-through in order: Samsung Electro-Mechanics, Samsung Electronics, SK Hynix, and then ISC, Leeno Industrial, and TSE.

Korea ADR at 67: Why the Index Can Hold While Most Stocks Are Weak

Research OS local data shows Korea's 20-day advance-decline ratio falling from 113.1 in mid-April to 67.3 on May 26. KOSPI and KOSDAQ are not in a broad risk-on phase; money is compressed into AI infrastructure, semiconductor bottlenecks, back-end names, substrates, shipbuilding and defense.

AI Server Passive-Component Bottlenecks: Why Tiny Power-Stability Parts Now Matter

AI server passive-component bottlenecks are not about a shortage of GPUs. They are about the higher-spec MLCCs, silicon capacitors, inductors, and filters needed to stabilize the huge transient power demands of GPU/HBM systems. This note explains the bottleneck and why it matters for Samsung Electro-Mechanics.