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

- **Cerebras tokenomics play**: AWS/OpenAI deals signal shift to wafer-scale edge inference, but SemiAnalysis reveals on-chip SRAM understated by 8x—credibility risk for token-based pricing models. - **AMD doubles down on software**: $3.6M MI355X cluster access for OSS maintainers plus 75x ROCm performance jump in 14 days post-DeepSeek v4—materially closing the CUDA moat for inference workloads. - **Grid constraints bite**: PJM data shows 4-year post-approval delays; developers flee city zoning to unincorporated counties, shifting $2B grid costs to Maryland ratepayers. Georgia site used 29M gallons water unpaid—regulatory backlash imminent. - **Nvidia infrastructure moves**: Corning fiber prepay, DOE 100K-GPU supercomputer, and AI factory blueprints (4-node to 128-cluster) lock in supply chain and reference architecture—competitors face higher integration friction. - **Disaggregation economics crystallize**: B200 PD disaggregation yields 7x token throughput and cost reduction; CoreWeave tops Kimi K2.6 inference—operators with disaggregated stacks gain immediate margin advantage. - **Second-order implication**: 37 new global DC sites under construction + SoftBank water batteries by 2028 signal hyperscalers pre-buying power and cooling capacity now, compressing timeline for AMD/Cerebras to prove rack-scale reliability at Nvidia volumes.