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

- **Meta & BlackRock commit $14B to Texas AI data center**; hyperscale capital rotation continues, with private infrastructure funds absorbing balance-sheet risk. - **Google posts first negative free cash flow since 2004 IPO** as AI capex hits; signals hyperscalers may face investor pushback on unconstrained spend. - **Nvidia weighs $250B guarantee for OpenAI’s Ohio campus**; suggests vendor financing becoming a competitive lever to lock in GPU demand. - **KV cache offload named new AI bottleneck**; storage subsystem upgrades (e.g., high-bandwidth SSDs, CXL) become critical for inference at scale. - **AMD-Cerebras disaggregated inference shows 5× T/s/W**; prompt/decode split architecture gains traction, challenging monolithic GPU designs. - **Epoch AI: Google’s Colossus 1 training compute hits 1e26 FLOP**; confirms frontier training runs now require dedicated power plant co-location, raising site-selection constraints.