Why AI Infrastructure Investment Is Shifting From Chip Supply to Power Capacity

According to a recent compilation published by eciks.org, hyperscaler capital allocation to AI infrastructure has cleared $700 billion in 2026 — nearly double the $400 billion deployed in 2025 — and the binding constraint has shifted from accelerator supply to grid-side electrical capacity. Goldman Sachs modeling, dated May 2026, places annual AI capex at $765 billion this year and projects a climb to $1.6 trillion annually by 2031, yielding $7.6 trillion in cumulative spend across compute, data center facilities, and supporting power infrastructure. The capital cycle is now governed less by transistor density than by megawatts deliverable to the rack.

Where the dollars concentrate

The Goldman figures trace directly to per-generation escalation in accelerator power draw, which forces facility designs toward liquid-cooling distribution loops, higher-voltage DC delivery, and redundancy margins that legacy cloud specifications from the prior decade were never provisioned for. Construction costs reflect the shift: next-generation AI data centers run $15 million to $20 million per megawatt, against roughly $10 million per megawatt for traditional cloud facilities built in the 2010s. Capex acceleration has outrun prior analyst baselines — the 2025 increase from 2024 alone topped 50%, against early-cycle forecasts of 20% growth — and the overshoot is now absorbed in grid interconnection queues and behind-the-meter generation projects rather than in chip procurement.

Energy as the critical path

The International Energy Agency's 2026 outlook frames global electricity supply and grid investment at $1.6 trillion for the year, with grid spend alone approaching $550 billion, a roughly 20% year-on-year lift. IEA modeling indicates annual grid investment will need to expand approximately 50% above the current $400 billion baseline to clear demand through 2030. On the U.S.-specific axis, Deloitte's June 2025 estimate projected AI data center power demand growing more than thirtyfold to 123 GW, a figure that exposes the delta between compute deployment roadmaps and available transmission headroom. Hyperscalers are responding with long-term utility offtake contracts, co-investment in renewable generation, and direct construction of behind-the-meter power assets — interventions that compress individual project timelines but duplicate capital across the system and erode aggregate efficiency.

Strategic capital tracks the constraint

The pattern is encoded in recent deal flow. As reported by Pulse 2.0, SoftBank's $2 billion commitment to Intel is positioned around U.S. semiconductor manufacturing capacity tied to AI infrastructure demand. Separately, mezha.net reports Nvidia is planning up to $3 billion in Lancium to secure data center capacity backed by dedicated power generation. Both transactions route capital toward the constraint that currently determines deployment velocity. For operators and developers, the practical implication is straightforward: site selection, interconnection queue position, and behind-the-meter generation contracts now govern cluster bring-up timelines as decisively as HBM allocation or parameter-count budgeting for frontier-scale training runs.