AI Inference is an Electricity and Heat Eating Hippo... 'Samsung and SK Hynix' Emerge as Saviors

Summary

Park Se-ik reads and comments on several news articles covering China semiconductor export strategy, US 30-year bond yield crossing 5%, AI inference paradigm shift toward logic and accuracy, SK Hynix's LPCAMM2 memory as a key enabler for inference, Apple's potential Samsung foundry orders, and the history of HBM development. He warns the current semiconductor rally shows bubble characteristics similar to the dot-com era and advises caution, but suggests buying on a 20% correction around June.

• China semiconductor article: Jensen Huang argues selling high-end US chips to China can slow its chip development.
• US 30-year yield at 5% raises concerns about investment and consumption, and may signal a policy tightening shift.
• AI competition shifts from speed to logical reasoning and agentic capabilities, increasing inference compute costs.
• SK Hynix's LPCAMM2 memory offers 75% better inference power efficiency, positioning it for NVIDIA's Vera Rubin platform.
• Apple is exploring Samsung Foundry and Intel as alternatives to TSMC for processor production.
• HBM history details SK Hynix's pioneering role and eventual dominance after early struggles and competition with Samsung.
• Park Se-ik notes the semiconductor rally resembles the dot-com bubble and warns of a potential severe correction, but sees a buying opportunity on a 20% drop.