[DC] What Changed in AI Infra — Week 2026-W24

- **Google locks in Intel for 3M+ TPUs by 2028**, signaling TSMC CoWoS capacity is structurally constrained. Second-order: Google is vertically integrating substrate supply for custom silicon, bypassing foundry bottlenecks—expect similar moves from AWS/AMD. - **xAI pivots Colossus to rental compute**, chasing 30%+ margins. Operator move: shifts from single-tenant training to multi-tenant inference leasing. Implication: hyperscaler overcapacity in H100 clusters may accelerate spot-market commoditization. - **MLCC shortage threatens AI server ramp** as prices hike and lead times stretch. Second-order: passive component bottlenecks become the new gating factor for GPU cluster deployment, potentially delaying 2024H2 capacity adds. - **Liquid cooling crosses 15kW coldplate threshold** (CoolIT quadruples capacity). Operator move: CoolIT displaces air-cooled SKUs in high-density racks. Implication: 50% liquid cooling adoption by 2027 now plausible, forcing data center retrofits and power infrastructure redesign. - **DRAM prices double, squeezing non-AI sectors**. Second-order: memory supply is being reallocated to HBM and high-density DDR5 for AI servers, starving enterprise/consumer markets—may drive server build delays for non-hyperscaler operators. - **Google tests 'power-first' model at 1 GW Texas campus**, securing grid interconnection before compute buildout. Operator move: hyperscalers are now prioritizing power procurement over silicon procurement. Implication: data center site selection shifts to grid-constrained regions, with 3-5 year lead times for new substations.