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

- **NVIDIA Blackwell Ultra** dominates first agentic AI benchmark, delivering 20x agents/MW vs Hopper, signaling a shift toward power-constrained inference clusters for autonomous workflows. - **KKR launches Helix Digital Infrastructure** with $10B for AI buildout; **Adam Selipsky** (ex-AWS CEO) leads a separate $10B AI data center venture—two major capital vehicles targeting hyperscale supply constraints. - **Google commits to 3M+ TPUs by 2028** via Intel foundry, as TSMC CoWoS hits capacity wall; this diversifies supply chain but raises execution risk for Intel’s advanced packaging ramp. - **Wiwynn unveils SCADA server** with 2.9PB and no CPU for I/O—eliminates traditional storage controller bottleneck, enabling denser GPU clusters for checkpoint-heavy training. - **Crusoe claims 5 GW pipeline but pauses Wyoming site**; underscores tension between aggressive buildout targets and permitting/energy delays, likely slowing capacity additions in constrained regions. - **OpenAI eyes 10GW Ohio data center** with Nvidia backing—second-order implication: utility-scale AI clusters may force grid upgrades and new renewable/battery partnerships, impacting regional power markets.