[DC] What Changed in AI Infra — Week 2026-W30

- **Nvidia & SK Group announce $500B AI infrastructure partnership**, with a separate $1B Naver investment to build a Korea AI hub. This signals hyperscale capital deployment shifting toward government-linked consortia, not just US hyperscalers. - **AMD secures Anthropic as anchor customer for 2GW MI450 supply**, with up to $5B investment. This is AMD’s largest single GPU commitment, directly challenging Nvidia’s dominance in frontier model training. - **Epoch AI clocks Google’s Colossus 1 training at 1e26 FLOP**, marking the first public compute benchmark for a single training run above 1e26. Expect competitive pressure on OpenAI and Anthropic to disclose or match. - **KV cache offload identified as the new storage bottleneck**, with inference serving now I/O-bound. This shifts procurement priorities toward high-bandwidth NVMe fabrics and disaggregated memory systems. - **Crusoe & ON.energy plan 5GW of dedicated AI UPS at hyperscale campuses**, reflecting a structural shift: backup power is being pre-deployed at grid scale, not retrofitted. - **China’s domestic AI ecosystem standardizes on MoE with wide expert parallelism**, as Zhipu builds a 1GW China-only data center and acquires a compiler startup. Implication: Chinese clusters are optimizing for sparse compute, not dense GPU scaling.