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

- **Nvidia Blackwell Ultra dominates agentic AI benchmark:** 20x agents/MW vs Hopper; NVFP4 precision cuts JAX training 1.8x in MaxText. Second-order: shifts inference efficiency metric to agents-per-watt, pressuring AMD and custom ASIC roadmaps. - **KKR launches Helix Digital Infrastructure with $10B** for AI buildout; ex-AWS CEO Adam Selipsky leads a separate $10B AI data center venture. Second-order: PE-backed hyperscale capacity now exceeds $20B in new commitments, signaling institutional capital rotating from cloud into dedicated AI infra. - **TensorWave raises $350M Series B for AMD-powered GPU clusters.** Second-order: AMD MI300X ecosystem gains a dedicated operator; challenges Nvidia’s dominance in non-H100 GPU-as-a-service, but supply chain and software maturity remain constraints. - **Google books Intel for 3M+ TPUs in 2028** as TSMC CoWoS hits capacity wall. Second-order: Intel foundry wins a high-volume AI chip order, decoupling from TSMC for Google’s internal silicon; CoWoS bottleneck shifts to non-TSMC advanced packaging. - **OpenAI eyes 10GW Ohio data center with Nvidia backing.** Second-order: single-site power demand exceeds current U.S. colocation total; grid interconnection and regulatory approval become the critical path, not chip supply. - **Crusoe pauses Wyoming site despite 5 GW pipeline**; NEMA/ASHRAE/PNNL launch AI data center framework as power demand hits 175 TWh. Second-order: permitting and cooling constraints force operators to pause greenfield sites; industry-wide standards emerge to manage power density and water use.