[DC] What Changed in AI Infra — Week 2026-W20

- **AMD opens MI355X cluster to OSS maintainers** — $3.6M compute grant signals shift from corporate-only access to open-source ecosystem seeding. Likely accelerates ROCm adoption ahead of Blackwell ramp. - **CoreWeave tops Kimi K2.6 inference speed** — inference performance becoming the new operator differentiator; CoreWeave’s lead pressures hyperscalers to optimize serving stacks, not just training clusters. - **Nvidia Blackwell CLC improves GEMM tile scheduling 15%** — compiler-level gains cut reliance on static persistence; second-order: extends Blackwell’s advantage over AMD in dense matrix ops for LLM training. - **SoftBank to build water-based batteries for AI DCs by 2028** — long-duration storage play targets backup power cost; if viable, lowers hyperscaler dependency on diesel generators and grid interconnection timelines. - **Rural data centers bypass city bans, shift $2B grid cost to Maryland ratepayers** — zoning arbitrage accelerates; regulators face mounting cost-pass-through disputes as municipalities lose control over siting decisions. - **OpenAI open-sources datacenter networking tech** — releases internal fabric tooling; lowers barrier for smaller operators to build custom topologies, but also exposes hyperscale networking IP to competitors.