AI infrastructure faces four hurdles: Power, construction, technology, funding

The AI compute bottleneck is undergoing a paradigm shift: from "can't fabricate enough chips" to "can't feed the grid." When single-cluster power demand crosses 100 MW, interconnection queues at PJM and ERCOT (2-4 year waits) and transformer delivery cycles (18-24 months) become harder constraints than EUV lithography capacity. The power wall is forcing a fundamental architectural rethinking. GB200 NVL72 racks now approach 120 kW per unit, making liquid cooling non-negotiable and shifting chip design priorities from peak FLOPS to watts-per-inference-token. This creates a structural window for AMD's MI350 and Intel's Gaudi 3, whose per-watt efficiency advantages become commercially decisive in power-constrained deployments. The more dangerous signal sits on the capital side: utility bargaining power is inverting. Echoing the 2022 ERCOT price spike that forced data center operators into 10-15 year PPAs at locked-in rates, hyperscaler capex returns are being systematically compressed. Microsoft's nuclear agreement with Talen Energy is, in essence, a compute company being forced to become an energy company. Forecast: within 18 months, "tokens per kilowatt-hour" will replace "FLOPS per dollar" as the primary procurement KPI. By 2027, partially built facilities lacking committed grid capacity will face idle-asset risk—the industry's first structural compute-power mismatch.