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

- **Google posts first negative FCF since 2004 IPO** as AI capex surges; signals hyperscale willingness to tolerate negative cash flow for strategic AI positioning, potentially resetting investor expectations across the sector. - **Nvidia weighs $250B guarantee for OpenAI's Ohio campus** and partners with SK Group on $500B infrastructure deal; indicates shift from GPU vendor to anchor financier of frontier clusters, concentrating risk on single operator. - **Safe Superintelligence partners Nvidia for 10x compute scale-up** and **Moonshot AI releases 1.56T-parameter Kimi K3 requiring 2x B200 nodes**; confirms frontier model scaling continues despite efficiency narratives, demanding denser GPU configurations. - **AMD to supply Anthropic with 2GW of MI450 GPUs, up to $5B investment**; marks first major hyperscale commitment to AMD at this scale, breaking Nvidia's near-monopoly on frontier training clusters. - **KV Cache Offload makes storage the new AI bottleneck** and **AMD-Cerebras disaggregated inference splits prompt/decode**; architectural shift toward memory-bandwidth-limited inference drives new disaggregation patterns, benefiting storage and interconnect vendors. - **Huawei Ascend SuperPOD decode throughput estimated 1.3-1.7x behind GB300**; China's domestic alternative remains materially behind on inference performance, reinforcing US export controls' effectiveness while CXMT's 471% STAR debut signals capital flight into domestic chip plays.