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

- **Pretraining concentration**: SemiAnalysis declares pretraining dead for all but frontier labs (OpenAI, Google, Meta, xAI), accelerating a two-tier market. Second-order: GPU rental demand from mid-tier labs collapses; TensorWave’s $350M AMD cluster raise is a contrarian bet on inference/ fine-tuning workloads. - **Google’s TPU + Intel lock**: Google books 3M+ TPUs from Intel for 2028, citing TSMC CoWoS capacity wall. Implication: hyperscalers are pre-buying 3+ years out, squeezing mid-market GPU supply and forcing operators like Crusoe (5 GW pipeline, paused Wyoming site) to delay. - **xAI pivots to rental compute**: xAI shifts Colossus from internal training to rental, targeting 30%+ margins. Google pays $920M/month for that capacity. Second-order: hyperscalers become net renters, not just builders; rental GPU pricing may stabilize upward. - **Liquid cooling standard rises**: CoolIT coldplate quadruples capacity to 15kW; Supermicro shows Vera Rubin NVL72 with new coolant. Implication: 15kW+ per rack becomes baseline for Blackwell-class clusters, pressuring colo operators without liquid infrastructure. - **MLCC shortage threatens ramp**: Prices hiking, lead times stretching for multi-layer ceramic capacitors. Second-order: server assembly delays cascade into 2025 H2 GPU cluster deployments, especially for non-hyperscaler operators (e.g., Nebius UK 65MW). - **Energy pivot: battery giant into AI**: CATL invests in DeepSeek; OpenAI eyes 10GW Ohio site with Nvidia backing. Implication: AI data center siting now driven by power availability, not latency—nuclear-inspired cooling (MIT spinoff) and battery storage become critical infrastructure.