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

- **Google locked in 3M+ TPU orders with Intel for 2028** as TSMC CoWoS capacity tightens, signaling a multi-year supply chain bottleneck for advanced packaging. Second-order: hyperscalers are pre-buying compute years out, compressing lead times for smaller operators. - **xAI pivoted Colossus to rental compute** targeting 30%+ margins, mirroring a broader hyperscaler shift toward leasing excess GPU capacity. Implication: spot pricing for H100/B200 clusters may soften as supply re-enters market. - **MLCC shortage emerged as a critical AI server bottleneck** with price hikes and extended lead times, threatening Q3-Q4 rack deployment schedules. Second-order: server OEMs may prioritize high-margin AI SKUs, squeezing non-AI enterprise shipments. - **Google's 1 GW Texas campus operationalizes 'power-first' hyperscaler model**—site selection now driven by grid capacity, not latency. Implication: data center buildouts will cluster near substations, straining regional utility planning. - **Liquid cooling crossed 15kW per coldplate** (CoolIT) and is projected to exceed 50% adoption by 2027 as rack densities force thermal redesign. Second-order: legacy air-cooled colos face stranded asset risk; liquid cooling supply chain (pumps, fittings) will see 3x demand growth. - **JPMorgan, OQC, AMD co-built first quantum-AI data center for finance**, combining HPC, quantum, and classical AI in a single facility. Implication: financial services will drive early hybrid quantum-classical workloads, pressuring hyperscalers to offer integrated quantum-as-a-service.