[DC] What Changed in AI Infra — Week 2026-W20
- **Cerebras disclosed wafer-scale edge deals (AWS, OpenAI) while SemiAnalysis flagged 8x on-chip SRAM understatement**, eroding architectural credibility despite tokenomics narrative. - **Nebius broke ground on 1GW Missouri AI campus** amid local opposition; PJM data confirms post-approval delays now average 4 years, shifting risk to unincorporated county sites. - **AMD launched MI350P PCIe card (40% FP8 lead vs H200 NVL) and gave OSS maintainers $3.6M MI355X cluster access**, targeting existing server install base; ROCm performance jumped 75x post-DeepSeek v4. - **B200 PD disaggregation boosted token throughput 7x, slashed cost**; Perplexity claimed 3x Blackwell inference throughput for 70B models, tightening Nvidia’s inference moat. - **SoftBank committed to water-based batteries for AI data centers by 2028**, signaling grid-scale storage pivot; Georgia data center used 29M gallons water unpaid, intensifying resource scrutiny. - **Coherent Q3 FY2026 AI optical revenue hit $1.2B**; CNAS report warns chip supply trails $700B CapEx, creating silicon wall risk for hyperscaler buildouts.
Evidence (raw JSON)
{
"kind": "dc_weekly_synthesis",
"week": "2026-W20"
}