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

- **Nvidia & SK Group $500B partnership** and **Nvidia’s $1B Naver investment** signal aggressive Korea-centric AI infrastructure buildout; second-order: shifts supply chain dependency toward Korean HBM and memory partners, pressures Samsung and Micron to match SK’s capacity commitments. - **AMD commits 2GW MI450 GPUs to Anthropic** plus up to $5B investment; notable because Anthropic bypasses Nvidia for a single-vendor AMD cluster at hyperscale – second-order: validates AMD’s MI-series for frontier training, may trigger similar defections from other labs. - **OpenAI’s $20B Georgia gigawatt data center** and **Crusoe/ON.energy 5GW UPS deal** confirm hyperscalers are moving beyond colo into direct utility-scale construction; implication: power procurement now the critical path, with land seizures and grid constraints becoming regulatory flashpoints. - **70% of markets overstretched** (data center construction capacity wall) + **$500B debt subprime comparison**: physical buildout is hitting labor, materials, and permitting limits; second-order: delays will push operators toward modular/containerized designs and longer lead times for new clusters. - **KV cache offload making storage the bottleneck** and **Theta-TaN metal (1,100 W/mK)** and **NUS CIMERA compute-in-interconnect** point to memory and thermal limits shifting from GPU compute to memory hierarchy; second-order: storage vendors (Pure, NetApp, VAST) and advanced cooling startups gain strategic leverage in cluster design. - **China’s Huawei Ascend SuperPOD 1.3-1.7x behind GB300** and **Zhipu AI 1GW China-only data center** + **export control rules**: China is scaling domestic compute despite performance gap, standardizing on MoE architectures; implication: global AI supply chains bifurcate further, with China’s 1GW builds absorbing local fab capacity and competing for HBM-equivalent memory.