Amazon’s $5 billion Nvidia AI servers sit idle in warehouses amid data center power shortages - The American Bazaar

The idle $5B GPU fleet in Amazon's warehouses is not a logistics footnote—it is a structural indictment of America's energy infrastructure. The binding constraint has migrated from wafer capacity to megawatt availability, and the mismatch between a 5-to-10-year grid permitting cycle and a 12-to-18-month AI capex cycle now throttles the entire hyperscaler cohort simultaneously. Technically, when a single B200 rack demands kilowatts rivaling a small town, chip design's north star shifts from raw FLOPS to watts-per-operation. Liquid cooling and small modular reactors transition from optional add-ons to deployment prerequisites. The upstream supply chain—transformers, switchgear, HV cables with 18-month-plus lead times—becomes the new critical path, displacing lithography as the schedule bottleneck. Strategically, sovereign AI programs in the Gulf are gaining a structural deployment-speed advantage simply because their power costs and grid headroom are superior. Microsoft and Meta face identical constraints; whoever locks in nuclear PPAs first will own the 2026 compute pricing curve. Over the next 12-to-24 months, expect GPU spot premiums to compress under "cards without current" dynamics. Cloud providers will treat megawatt reserves as strategic assets on par with silicon. The new oil of AI infrastructure isn't chips—it's power.