[DC] What Changed in AI Infra — Week 2026-W29

- **Crusoe & Lancium 1GW Texas campus** breaks ground as first major AI facility to bypass grid interconnection entirely, using on-site gas generation and modular prefab construction. Sets precedent for hyperscalers seeking speed over utility dependency. - **Google adopts Intel EMIB** for 9th-gen TPU packaging, breaking TSMC’s CoWoS monopoly. Second-order: Intel foundry gains credibility for high-volume AI interposers; TSMC capacity constraints may ease for other customers. - **New York bans AI data centers >50 MW** — first U.S. state moratorium. Immediate chilling effect on Northeast buildout; operators will route 1GW+ campuses to Texas, Virginia, and international sites with faster permitting. - **Nvidia halves Asia partner list** to curb chip smuggling. Implication: gray-market H100/H200 supply tightens; Chinese firms (Alibaba, Tencent, ZTE) now rely on official H200 allocations via U.S. export licenses, not unauthorized channels. - **Meta Hyperion expansion to 5GW** in Louisiana ($50B+) signals hyperscale frontier shifting to deep-South grid-captive campuses. Second-order: local utility and transmission upgrades become gating factor for follow-on projects. - **Morgan Stanley forecasts top-5 cloud capex hitting $1.2T by 2028**, with compute capacity quadrupling. Implication: GPU supply remains the binding constraint; any packaging (CoWoS/EMIB) or cooling bottleneck will cascade into delayed cluster deployments.