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

- **Google locked in 3M+ TPUs for 2028 via Intel, signaling TSMC CoWoS capacity is constrained for next-gen AI silicon.** This is a multi-year capacity hedge; second-order effect: pressures Intel’s foundry execution and may tighten supply for other hyperscalers relying on TSMC. - **xAI pivoted Colossus to rental compute, targeting 30%+ margins**, while Google pays SpaceX $920M/month for xAI capacity. Implication: hyperscalers are now both buyers and sellers of GPU compute, blurring the line between cloud provider and AI lab—margin pressure on pure-play cloud rental. - **MLCC shortage threatens AI server ramp** as prices hike and lead times stretch. Second-order: server assembly timelines slip, potentially pushing GPU shipments into Q2, and favoring hyperscalers with long-term component contracts over smaller operators. - **Liquid cooling crossed a threshold: CoolIT’s 15kW coldplate quadruples capacity**, and rack densities force adoption past 50% by 2027. Implication: legacy air-cooled data centers face accelerated obsolescence; operators without liquid infrastructure will struggle to host next-gen Blackwell/TPU racks. - **UK doubled sovereign AI cloud providers, deploying a 65MW Nebius cluster**, while Google’s 1 GW Texas campus tests power-first siting. Implication: sovereign AI infrastructure is scaling regionally, but power availability—not GPU supply—is becoming the binding constraint for new builds. - **Memory supply squeeze doubled DRAM prices, hitting non-AI sectors.** Second-order: AI’s memory demand (HBM, DDR5) is cannibalizing general server and PC DRAM, raising costs for enterprise workloads and potentially slowing non-AI cloud expansion.