Commentary: US governors target AI data centers as Taiwan firms await Nvidia outlook

Core thesis: US state-level regulatory fragmentation is fundamentally rewriting the geographic economics of AI compute infrastructure. This is structural, not cyclical. Technical cascade: Stricter approvals in four states hit CoWoS advanced packaging utilization directly. A single paused mid-scale AI cluster eliminates 2-3 weeks of TSMC's monthly CoWoS-L capacity consumption. HBM supply chains in Taiwan, China (SK Hynix, Samsung) face forced shipment rescheduling. Broadcom and Marvell's networking visibility compresses from 12 months to 6, pushing the entire AI infrastructure BOM into pulse-like volatility. Compliance risk: Federal export controls governed "who can buy"; state-level permitting governs "where it can be built" — an entirely new axis. Suppliers in Taiwan, China now face dual compression: export restrictions AND destination-market closure. Compliance overhead multiplies across 50 state jurisdictions. Market dynamics: With data center revenue exceeding 70% of Nvidia's total, a one-quarter delay across four states implies $4-6B in revenue deferral. AMD accelerates MI400 ecosystem lock-in; Broadcom captures share via Trainium and Maia custom ASICs. Hyperscalers will accelerate sovereign compute diversification toward the Middle East and Southeast Asia to hedge single-grid-market risk. Trend: Within 18 months, US AI infrastructure will exhibit "interstate arbitrage" — centralized hyperscale irreversibly shifting toward distributed, sovereignty-driven deployment. Nvidia's earnings call is the tell: if management references "geographic reallocation of demand," structural slowdown is confirmed. If they cite "unchanged total demand," the market prices it as transient. The true long-tail effect: compute infrastructure power is migrating from federal to state, from centralized to sovereign.