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

- **Google secures 3M+ TPUs for 2028 via Intel foundry, bypassing TSMC CoWoS bottleneck.** Second-order: shifts supply chain dependency away from TSMC; signals Google's long-term bet on custom silicon over merchant GPUs for hyperscale inference. - **xAI pivots Colossus to rental compute, targeting 30%+ margins.** Implication: monetizes idle GPU capacity; pressures hyperscalers to offer flexible, short-term compute contracts to compete for variable AI workloads. - **MLCC shortage emerges as AI server ramp bottleneck; prices hiking, lead times stretching.** Second-order: delays server deployments for smaller operators; forces hyperscalers to pre-book passive components 12+ months out. - **Foxconn + Intel partner on AI data center rack systems.** Moves Foxconn upstream into system integration; Intel gains channel to hyperscale rack-level sales, bypassing server OEMs. - **Liquid cooling crosses 15kW per coldplate (CoolIT); rack densities force >50% adoption by 2027.** Implication: air-cooled infrastructure becomes stranded asset; colo providers without liquid-ready capacity lose hyperscale tenants. - **UK doubles sovereign AI cloud providers, deploys 65MW Nebius cluster.** Signals government-backed diversification from US hyperscalers; Nebius becomes European AI compute alternative, likely attracting EU-regulated workloads.