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

- **Google books 3M+ TPUs for 2028, signals TSMC CoWoS bottleneck.** Google locked Intel for TPU production as TSMC CoWoS hits capacity wall. Implication: second-source foundry strategy becomes a hyperscaler necessity; Intel's advanced packaging ramp becomes a gating factor for Google's 2028 TPU deployment. - **xAI pivots Colossus to rental compute, targets 30%+ margins.** Shift from internal training to external compute rental. Implication: xAI monetizes spare GPU capacity, pressuring cloud GPU rental pricing; signals overbuild of training clusters relative to internal demand. - **MLCC shortage threatens AI server ramp; prices hiking, lead times stretching.** Passive component bottleneck emerges. Implication: server delivery timelines slip 4-6 weeks; hyperscalers will pre-buy or vertically integrate into passives; non-AI server builds get deprioritized. - **Liquid cooling crosses 50% by 2027; CoolIT quadruples coldplate capacity to 15kW.** Rack densities force thermal shift. Implication: air-cooled data center retrofits accelerate; CoolIT's capacity jump signals OEM design wins for next-gen GPU racks; 15kW coldplate enables denser H100/B200 clusters. - **Google's 1 GW Texas campus tests 'power-first' model.** Site selection driven by grid capacity, not latency. Implication: hyperscalers prioritize power availability over proximity to users; shifts data center site economics toward substation access and PPA contracts; 1 GW campus becomes new baseline for hyperscale builds. - **Memory supply squeeze hits non-AI sectors; DRAM prices double.** HBM allocation crowds out DDR5/LPDDR. Implication: non-AI server and PC OEMs face margin compression; hyperscalers allocate DRAM to AI clusters first; second-order effect: enterprise refresh cycles slow, extending legacy hardware lifespan.