[DC] What Changed in AI Infra — Week 2026-W30

- **Nvidia & SK Group announced $500B AI infrastructure partnership**; Nvidia also invested $1B in Naver and expanded SK Group pact for a Korea AI hub. Implication: Nvidia is vertically integrating across memory, networking, and regional hubs to lock in supply chains ahead of demand. - **AMD to supply Anthropic with 2GW of MI450 GPUs, invest up to $5B**; Meta custom AMD MI400 half-size chip targets RecSys with 144GB HBM. Implication: AMD is gaining hyperscale design wins, forcing Nvidia to defend its training/inference share with custom SKUs. - **OpenAI's $20B Georgia data center tests gigawatt buildout**; Crusoe and ON.energy deploy 5GW AI UPS at hyperscale campuses. Implication: Power delivery is now the binding constraint — infrastructure debt ($500B) and land seizures signal a capital-intensive, subprime-like risk build. - **Data center construction hits global capacity wall — 70% of markets overstretched**; zAI completes 1GW data center without Nvidia chips. Implication: Non-Nvidia clusters (AMD, Huawei, custom silicon) are filling supply gaps, accelerating ecosystem fragmentation. - **Huawei Ascend SuperPOD decode throughput estimated 1.3-1.7x behind GB300**; China's ecosystem standardizes on MoE with wide expert parallelism. Implication: China's AI hardware gap is narrowing to ~30% in inference, enabling domestic scaling despite export controls. - **Theta-TaN metal hits 1,100 W/mK thermal conductivity (3× copper)**; Supermicro 7,000-lb double-wide rack targets B300 memory squeeze. Implication: Thermal and memory bottlenecks are driving radical packaging changes — new materials and rack form factors will reshape cluster design in 2026.