[DC] What Changed in AI Infra — Week 2026-W31

- **Nvidia guarantees $250B for OpenAI's Ohio campus**, signaling hyperscale capex escalation; SK Group $500B AI infra partnership and $1B Naver investment deepen Korea AI hub buildout, shifting supply chain重心 to Asia. - **AMD commits 2GW MI450 GPUs to Anthropic** with up to $5B investment, breaking Nvidia's lock on frontier-model training; AMD-Cerebras disaggregated inference (5× T/s/W) splits prompt/decode, threatening Nvidia's inference dominance. - **Epoch AI logs Google's Colossus 1 at 1e26 FLOP** (training compute milestone); Gemini 4 pretraining begins—most ambitious run yet—while Gemini 3.6 Flash beats GPT-5.6 on computer use, intensifying frontier model race and compute demand. - **Data center construction hits global capacity wall** (70% markets overstretched); $500B debt mirrors subprime risk, per analysts—Crusoe/ON.energy deploy 5GW UPS, but supply-chain bottlenecks persist. - **KV cache offload makes storage new AI bottleneck**; Supermicro's 7,000-lb double-wide rack targets B300 memory squeeze—hardware redesigns accelerate as memory bandwidth fails to keep pace with model scale. - **China blocks AI/chip exports to West**, builds 1GW Zhipu AI data center, and debuts phase-change memristor neural chip—decoupling accelerates, pressuring Nvidia's China revenue and boosting domestic alternatives like Huawei Ascend (1.3-1.7× behind GB300 decode).