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

- **Moonshot AI's 1.56T Kimi K3** requires 2x B200 nodes, signaling Chinese labs bypass export controls via dense model parallelism; second-order: Nvidia's B200 backlog tightens further. - **Nvidia weighs $250B guarantee for OpenAI's Ohio campus** and **SK Group $500B partnership**; operators (Nvidia, SK) pivot to capital-backed infrastructure ownership, not just chip supply. - **AMD commits up to $5B and 2GW MI450 GPUs to Anthropic**; marks first hyperscale AMD GPU win against Nvidia, with implications for disaggregated inference architectures (see AMD-Cerebras 5× T/s/W). - **Data center construction hits global capacity wall** (70% markets overstretched) while **Crusoe/ON.energy deploy 5GW AI UPS**; bottleneck shifts from chips to grid interconnection and power delivery. - **Epoch AI: Google's Colossus 1 training compute hits 1e26 FLOP** and **Gemini 4 pretraining begins**; Google asserts compute leadership, but **KV cache offload** reveals storage as new latency bottleneck. - **Huawei Ascend SuperPOD decode throughput 1.3-1.7x behind GB300**; despite China's export-block rules and phase-change memristor chip, Nvidia's inference lead persists, pressuring domestic substitution timelines.