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

- **Nvidia doubles down on frontier compute**: Safe Superintelligence partners with Nvidia for 10x scale-up; Nvidia weighs $250B guarantee for OpenAI’s Ohio campus; $500B SK Group pact and $1B Naver investment signal Korea AI hub push. Jensen confirms open models (DeepSeek, Kimi) directly boost Nvidia sales. - **Hyperscale capex strain intensifies**: Google posts first negative FCF since 2004 IPO as AI capex hits; $500B data center debt wave draws subprime crisis parallels; 70% of global markets overstretched on construction capacity. Operators face liquidity vs. scale trade-off. - **Model scale breaks new thresholds**: Moonshot AI’s 1.56T-param Kimi K3 requires 2x B200 nodes; Epoch AI clocks Google’s Colossus 1 training at 1e26 FLOP. KV cache offload shifts bottleneck to storage, altering cluster design priorities. - **AMD challenges Nvidia on inference economics**: AMD-Cerebras disaggregated inference claims 5× T/s/W improvement via prompt/decode split; AMD to supply Anthropic with 2GW of MI450 GPUs, up to $5B investment. Huawei Ascend still 1.3-1.7× behind GB300 decode throughput. - **Infrastructure resource conflicts escalate**: AI data center sues for 287M gallons of Colorado River water; environmental pressure mounts alongside power constraints. The `pytorch_no_powerplant_blowup` env var highlights real operational fragility at frontier GPU clusters. - **Second-order implication**: Compute scale is decoupling from capital efficiency—hyperscalers face negative cash flow while model sizes double; disaggregated inference and storage-tier optimization become critical to avoid a power/water/capital triple crunch.