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

- **KKR launched Helix Digital Infrastructure with $10B** for AI buildout; former AWS CEO Selipsky will lead the venture. Second-order: private equity is now directly competing with hyperscalers for prime data center sites, tightening power and land availability for 2026-2028 projects. - **NVIDIA Blackwell Ultra set a new agentic AI benchmark**—20x agents per MW vs. Hopper. Implication: inference density is now the key metric; operators with Blackwell deployments will have a 20x efficiency advantage, accelerating Hopper retirement and reshaping cluster procurement. - **Google committed to 3M+ TPUs from Intel in 2028** as TSMC CoWoS hits capacity. Material change: Intel's foundry wins a major AI chip order, signaling a shift away from sole-sourced TSMC packaging for hyperscalers, with potential supply chain diversification for 2028+ builds. - **TensorWave raised $350M Series B** for AMD-powered GPU clusters. Operator move: an AMD-native cluster provider is scaling, challenging Nvidia's dominance; second-order effect is increased competition for AMD GPU allocation and potential pricing pressure on Nvidia's high-margin GPUs. - **Crusoe paused its Wyoming data center site** despite claiming a 5 GW pipeline. Implication: even well-funded operators are hitting permitting or power-delivery bottlenecks—expect more project pauses and site concentration in deregulated power markets. - **NEMA/ASHRAE/PNNL launched an AI data center framework** as power demand hits 175 TWh. Second-order: regulatory alignment is emerging, which will standardize cooling and power usage metrics, potentially raising compliance costs for smaller operators but enabling faster permitting for standardized designs.