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

- **Google locks in Intel for 3M+ TPUs by 2028** as TSMC CoWoS capacity hits wall, signaling a strategic decoupling from single-node packaging and a potential shift in hyperscaler supply chains toward Intel Foundry for advanced packaging. - **xAI pivots Colossus to rental compute**, chasing 30%+ margins, while Google pays $920M/month to SpaceX for xAI capacity — indicating a secondary market for GPU clusters emerging and hyperscalers monetizing overbuilt inventory. - **MLCC shortage threatens AI server ramp**: prices hiking, lead times stretching — a second-order bottleneck beyond GPU supply, likely delaying Q3/Q4 2025 cluster deployments for mid-tier operators. - **Liquid cooling crosses 15kW per coldplate** (CoolIT) and is projected to exceed 50% adoption by 2027, driven by rack density escalation — forcing data center operators to retrofit or build new facilities with direct-to-chip cooling. - **Memory supply squeeze doubles DRAM prices**, hitting non-AI sectors — a ripple effect from HBM3e allocation to AI clusters, potentially raising server BOM costs and slowing enterprise AI adoption. - **UK doubles sovereign AI cloud providers**, deploys 65MW Nebius cluster — a policy-driven move to reduce hyperscaler dependency, with implications for European AI compute sovereignty and local data center construction demand.